PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial Intelligence Index Report 2026

Sha Sajadieh, Loredana Fattorini, Raymond Perrault, Yolanda Gil, Vanessa Parli, Lapo Santarlasci, Juan Pava, Nestor Maslej, Russ Altman, Erik Brynjolfsson, Carla Brodley, Jack Clark, Virginia Dignum, Vipin Kumar, James Landay, Terah Lyons, James Manyika, Juan Carlos Niebles, Yoav Shoham, Elham Tabassi, Russell Wald, Toby Walsh, Dan Weld

arXiv:2606.15708v3cs.AI

TL;DR

AI’s rapid advance raises questions about whether governance, evaluation, education, and impact-tracking systems can keep pace. The report tracks AI testing, economic value, and labor-market evidence. It reports nearly 53% population-level generative AI adoption within three years, alongside frontier-model convergence and divergent national AI policies.

  • Problem

    Whether governance, evaluation, education, and impact-tracking systems can keep pace with rapidly advancing AI remains the report’s central question.

    Source: Introduction · p. 3

  • Method

    The report tracks AI testing across reasoning, safety, and real-world execution, alongside estimates of generative AI’s economic value and evidence of labor-market effects.

    Source: Introduction · p. 3

  • Results

    Nearly 53% population-level adoption within three years reflects rapid generative AI uptake, alongside frontier-model convergence and divergent national AI policies.

    Source: Introduction · p. 4

  • Takeaways & Limitations

    AI capability assessments face saturating benchmarks, reduced frontier-lab disclosure, and independent tests that do not always confirm developers’ reports.

    Source: Introduction · p. 4

    Government AI spending analysis is limited by uneven geographic data coverage and includes grant-level spending only for the United States.

    Source: Public Investment in AI · p. 422

Abstract

from arXiv · show

Welcome to the ninth edition of the AI Index report. As AI continues to advance rapidly, the question becomes whether the systems built around it can keep up. Governance frameworks, evaluation methods, education systems, and the data infrastructure needed to track AI's impact are struggling to match the pace of the technology itself. That gap between what AI can do and how prepared we are to manage it runs through every chapter of this year's report. New in this edition, the report tracks how AI is being tested more ambitiously across reasoning, safety, and real-world task execution, and why those measurements are increasingly difficult to rely on. It also features new estimates of generative AI's economic value alongside emerging evidence of its labor market effects, an analytical framework on AI sovereignty, and a science chapter developed in collaboration with Schmidt Sciences. For the first time, the report features standalone chapters on AI in science and AI in medicine, reflecting AI's growing impact across these two domains.

1 Research and Development … Synthetic Data in Pre-training

This section surveys expanding AI development and adoption, the concentration of research and infrastructure, and growing open-source participation. It examines trends in training compute and model access alongside gaps in transparency, evaluation, governance, and real-world reliability, including how incomplete disclosures constrain scaling comparisons and uncertainty about synthetic data’s ability to offset real-data depletion in pre-training.

Introduction

  • The report provides independent global evidence while emphasizing measurement gaps and expanding coverage of evaluation, economic value, sovereignty, science, and medicine.Its ninth edition introduces standalone science and medicine chapters while examining why increasingly ambitious AI evaluations are difficult to rely on. Public resources include raw data, high-resolution charts, and a Global AI Vibrancy Tool comparing 36 countries, scheduled for an update by the end of 2026.Source: Introduction · p. 3, Public Data and Tools · p. 6

Top Takeaways

  • 60% to near 100% of the human baseline marked SWE-bench Verified’s one-year performance rise, despite persistent gaps in reliability and responsible AI evaluation.OSWorld task success rose from 12% to approximately 66%, yet the top model read analog clocks correctly only 50.1% of the time; robots achieved 12% household-task success versus 89.4% on simulated RLBench manipulation. Benchmarks were saturating and independent testing did not always confirm developer reports, while responsible AI reporting remained uneven, documented incidents rose from 233 in 2024 to 362, and improving safety could degrade accuracy.Source: Introduction · p. 4, Top Takeaways · p. 10, but cannot reliably tell time—an… · p. 10
  • 53% population adoption within three years and 88% organizational adoption accompanied rapid investment growth and substantial consumer value from generative AI.Estimated annual U.S. consumer value reached $172 billion by early 2026; global corporate investment more than doubled in 2025, and U.S. private investment reached $285.9 billion versus China’s $12.4 billion, although private figures understate China’s government-backed spending. Productivity gains of 14% to 26% appeared in customer support and software development, but judgment-heavy tasks showed weaker or negative effects, agent deployment remained mostly single-digit, and employment among U.S. developers aged 22–25 fell nearly 20% from 2024.Source: Introduction · p. 4, §8 · p. 11, §7 · p. 11
  • 83% less note-writing time was reported by physicians using clinical AI tools, but rigorous clinical evidence remained limited and scientific performance varied substantially.A review of more than 500 clinical AI studies found nearly half used exam-style questions and only 5% used real clinical data, even as clinical deployment broadened and scientific AI attempted entire workflows. Frontier models exceeded human chemists on average on ChemBench but scored below 20% on astrophysics replication and 33% on Earth observation; smaller MSAPairformer and GPN-Star models also outperformed previous leading or much larger models.Source: Introduction · p. 4, §11 · p. 11
  • Over 80% of U.S. high school and college students used AI for schoolwork, but only half of middle and high schools had AI policies.Only 6% of teachers considered those policies clear; new U.S. and Canadian AI PhDs increased 22% from 2022 to 2024, with the increase entering academia rather than industry. National policies diverged around growing sovereignty ambitions and developing-country participation, while global optimism and nervousness both rose: 73% of experts versus 23% of the public expected positive job impacts, and U.S. trust in its government to regulate AI was 31%.Source: Introduction · p. 4, Formal education is lagging behind AI,… · p. 12

1.8 AI Authors and Inventors

  • Over 90% of notable AI models came from industry in 2025, while frontier disclosures declined and synthetic data still did not replace real pre-training data.Reported parameter counts remained near 1 trillion for three years as frontier disclosure stopped, while independently estimable training compute continued rising. Data quality and post-training showed promise: OLMo 3.1 Think 32B achieved comparable results to Grok 4 on several benchmarks with nearly 90 times fewer parameters through pruning, deduplication, and curation alone.Source: §1.8 · p. 15
  • 59 notable models came from the United States in 2025 versus China’s 35, although China led publication volume, citations, and patent grants.The U.S. retained higher-impact patents, South Korea led AI patents per capita, and Anthropic’s top model led Chinese competitors by just 2.7% in March 2026; open-source development reached 5.6 million GitHub projects, with Hugging Face uploads tripling since 2023. U.S. researcher and developer inflows fell 89% since 2017 and 80% in the last year; Switzerland and Singapore led talent per capita, while no country approached gender parity.Source: Top Takeaways · p. 10, §1.8 · pp. 15–16
  • 17.1 million H100-equivalents of global AI compute capacity reflected 3.3x annual growth since 2022, alongside concentrated hardware supply and expanding environmental impacts.Nvidia supplied over 60% of compute, the U.S. hosted 5,427 data centers, and TSMC fabricated almost every leading AI chip, although its U.S. expansion began operating in 2025. AI data center power capacity reached 29.6 GW, Grok 4’s estimated training emissions reached 72,816 tons of CO2 equivalent, and annual GPT-4o inference water use may exceed 1.2 million people’s drinking-water needs.Source: §1.8 · p. 15

By National Affiliation2

  • 59 notable AI models came from the United States in 2025, compared with China’s 35 and South Korea’s 8.New releases declined year over year across all major geographic areas. Counts use an April 22, 2026 snapshot and attribute countries through author affiliations, allowing multinational models to be counted more than once.Source: By National Affiliation2 · p. 17

By Sector and Organization

  • 91.2% of notable models originated in industry in 2025, with 93 industry models versus 2 academic models.OpenAI (20), Google (14), and Alibaba (11) led 2025 releases; since 2014, Google led cumulative output, followed by Meta and OpenAI. Academic leaders over the past decade were Tsinghua University (26), Stanford University (26), and Carnegie Mellon University (25).Source: By Sector and Organization · p. 19

Model Release

  • 47 of 102 notable models used API access in 2025, making it the most common release type as access shifted toward greater control.API-only releases have increased since 2020; unrestricted open weights, permitting use, modification, and redistribution, were the second-most-common release type. Access classifications are incomplete, so these totals may differ from other chapter counts.Source: Model Release · pp. 21–22
  • 81 of 102 notable models in 2025 lacked released training code, compared with 4 whose training code was open source.Open-source and unreleased training-code counts were approximately equal in 2020, but unreleased code became the majority by 2023 and the gap widened. This opacity limits reproducibility, development audits, and validation of safety claims.Source: Model Release · p. 21

Parameter and Compute Trends

  • Reported parameter-count growth has flattened since 2022, but missing disclosures likely understate actual growth.Leading models have trained on tens of trillions of tokens for periods exceeding 100 days. Several recent resource-intensive models from OpenAI, Anthropic, and Google have not disclosed parameter counts, training dataset sizes, or training duration.Source: Parameter and Compute Trends · p. 23
  • Training compute has risen by several orders of magnitude, with industry models accounting for the highest values.U.S. models appear more compute-intensive than Chinese models, but recent comparisons cannot be fully substantiated because U.S. models have not directly reported training compute. Epoch estimates missing compute from hardware usage, architectural operations and training data, or benchmark-performance proxies.Source: Parameter and Compute Trends · pp. 24–25

Will Models Run Out of Data?

  • 2026–2032 is Epoch AI’s projected window for depletion of available training data under certain assumptions, amid concerns about sustaining dataset-dependent scaling.Leading researchers have claimed that high-quality human text and web data are already exhausted, a condition described as “peak data.” The projected depletion window is conditional, not an established exhaustion date.Source: Will Models Run Out of Data? · p. 26

Synthetic Data in Pre-training

  • SYNTHLLM models trained entirely on synthetic data still lag leading models on major benchmarks, and synthetic data has not definitively demonstrated it can offset pre-training data depletion.Hybrid real–synthetic training can accelerate training by a factor of five to 10 at scale without surpassing real-data final performance; synthetic-only successes remain limited to smaller models or narrow tasks. Earlier findings suggest replacing real data can cause model collapse, while retaining real data can prevent collapse without guaranteeing performance gains.Source: Figure 1.1.15 · p. 26, Synthetic Data in Pre-training · p. 26

Data-centric Methods … By National Affiliation

This section examines the data practices, computing infrastructure, and publication and open-source activity supporting AI development. It tracks changes in training-data acquisition, expanding capacity and geographic participation, and persistent concentrations in research and infrastructure supply chains, alongside growing energy, water, and emissions burdens.

Data-centric Methods / Synthetic Data in Post-training / Prevalence of Synthetic Content

  • Approximately 32 billion parameters characterize Olmo 3.1 Think 32B, which performs comparably to 3-trillion-parameter Grok 4 on several benchmarks, including AIME 2025.Deduplication, quality-aware selection, stage-specific curricula, and iterative data-mix refinement supported competitive performance with substantially fewer training tokens; synthetic post-training data also improves few-shot generation, long-context capabilities, reinforcement-learning workflows, and reasoning. Graphite estimates that over 50% of newly published online content was AI-generated beginning January 2025, raising training-data concerns as firms increasingly pursue proprietary licensing while openly available training data declines.Source: Figure 1.1.16 · p. 27, Data-centric Methods · p. 27, Synthetic Data in Post-training · p. 27

Performance and Efficiency / Hardware for Notable Models

  • Machine-learning hardware peak performance increased exponentially across 2008–2025 releases, with lower-precision FP16 and Tensor-FP16/BF16 formats reaching the highest performance levels.These lower-precision formats have become standard in many training and inference settings. In 2025, 84 notable models used A100-class hardware, 69 used V100, and 28 used H100, while TPU v3 and TPU v4 adoption curves remained stable.Source: Figure 1.2.1 · p. 29, Performance and Efficiency · p. 29, Figure 1.2.2 · p. 30

Global Computing Capacity / Data Center Power Capacity

  • 3.3x annual growth since 2022 brought estimated global AI computing capacity to approximately 17.1 million H100-equivalents, with Nvidia supplying over 60%.Capacity growth tracks increased infrastructure investment, but estimates derive from revenues, disclosures, and analyst reports rather than exact chip counts, with manufacturer coverage beginning in different years. AI data center rated power capacity reached approximately 29.6 GW by Q4 2025, including 11.8 GW of chip thermal design power and an approximately 2.5 multiplier for supporting infrastructure.Source: Global Computing Capacity · p. 30, Figure 1.2.3 · p. 31, Data Center Power Capacity · p. 31

Geographic Distribution / AI Infrastructure: Beyond GPUs

  • 5,427 data centers placed the United States more than tenfold ahead of every other country in 2025, although facility counts do not measure computing capacity.Germany had 529 facilities, the United Kingdom 523, and China 449; counts also omit differences in facility size and utilization. Beyond accelerators, AI infrastructure depends on high-bandwidth memory, fast networking, foundries, and packaging firms; TSMC fabricates virtually every leading AI chip, creating a concentrated dependency within a supply chain protected by substantial expertise, equipment, and capital barriers.Source: AI Infrastructure: Beyond GPUs · p. 32, Figure 1.3.1 · p. 33, Geographic Distribution · p. 33

Training

  • 72,816 tons of CO2 equivalent were estimated for Grok 4 training in 2025, compared with 0.01 tons for AlexNet in 2012.Leading chips deliver about ten times more computation per watt than a decade ago, but model scaling has outpaced efficiency gains: Grok 3 and Llama 4 Behemoth required upward of 100 million watts during training, while disclosure gaps obscure many newer models. Emissions also depend on hardware efficiency, training duration, and electricity carbon intensity; DeepSeek v3 produced approximately 597 tons, substantially below comparably sized models.Source: Training · p. 34, Figure 1.4.1 · p. 35, Figure 1.4.3 · p. 36

Inference

  • 23 Wh per medium-length query placed DeepSeek V3.2 Exp and DeepSeek V3.2 highest among the 15 most energy-consuming models in the 2025 comparison.Jegham et al. (2025) define medium-length prompts as approximately 1,000 input and 1,000 output tokens: GPT-5 (high) consumed 21.9 Wh, while Claude 4 Opus and GPT-5 min (medium) consumed 5–6 Wh; emissions ranged from approximately 14 grams of CO2 equivalent for each DeepSeek model to 1.6 grams for Claude 4 Opus and 1.5 grams for Mistral Medium 3, showing that capability is not necessarily proportional to environmental cost. A short GPT-4o query consumed approximately 0.42 Wh versus 0.3 Wh for Google search, while estimated annual GPT-4o inference water consumption reached 1.3–1.6 million kiloliters, exceeding 1.2 million people’s annual drinking-water needs at the upper bound.Source: Figure 1.4.513 · p. 37, Inference · pp. 37–38, Figure 1.4.614 · p. 38

Data Center Usage / Inference

  • 9,400 MW was the estimated cumulative power demand of complete AI systems through 2024, compared with 5,200 MW for accelerator modules alone.De Vries and Gao (2025) attribute the largest accelerator share to Nvidia and qualify estimates by utilization and facility efficiency; complete AI systems approached Switzerland’s or Austria’s electricity consumption, roughly half Bitcoin mining’s demand, versus approximately 47,000 MW for global data centers excluding crypto. Inference can exceed training’s energy cost within months of deployment, while GPU computation costs have fallen more than 99% since 2006 and regional data center electricity consumption is projected to rise through 2030, led by the United States.Source: Inference · p. 37, Data Center Usage · pp. 40–41, Figure 1.4.12 · p. 41

Projects

  • 5.6 million AI-related GitHub projects were recorded in 2025, approximately, up 23.7% from 2024 and from 1,549 projects in 2011.Only 206,880 projects had at least ten stars, a rough community-engagement proxy, indicating that most repositories attract little attention despite similar growth trajectories across both measures. The dataset excludes Chinese code-sharing alternatives Gitee and GitCode.Source: Figure 1.5.1 · p. 42, Projects · p. 42

Projects / Stars

  • 31.7% of GitHub AI projects with at least 10 stars were U.S.-based in 2025, down from nearly 80% in 2011 but still the largest share.Annual new stars increased from 14 million in 2023 to 18.2 million in 2025, with every major region growing year-over-year, while U.S. projects led cumulative engagement with 30 million stars. Europe and the rest of the world expanded, India reached 5.2% of qualifying projects, and China’s share leveled off after 2019; China’s self-reported locations and exclusion of domestic platforms likely understate its activity.Source: Projects · p. 43, Figure 1.5.317 · p. 44, Figure 1.5.4 · p. 44

Model and Dataset Ecosystem

  • Model uploads more than tripled and dataset uploads grew fourfold on Hugging Face from 2023 to 2025, alongside broader participation in model development.Upload activity accelerated after Q2 2024; U.S.-developed models lost download share to unaffiliated users, Google and Meta declined in prominence relative to communities, and a large “Others” category coexisted with concentrated leading sources. Download analysis uses Longpre et al. (2025), a cleaned weekly panel covering March 2020–August 2025 whose top 200 models per week represent 49.6% of normalized, filtered downloads, limiting long-tail coverage.Source: Model and Dataset Ecosystem · p. 45, Figure 1.5.620 · p. 46, Figure 1.5.721 · p. 46
  • More than 42% of model downloads were for text generation in 2025, while nearly 80% remained concentrated in the top three categories.Text embedders, classifiers, and audio models collectively fell from nearly 70% of downloads in 2022 to less than 6% in 2025. Multimodal and video generation grew, while image generation increased steadily and remained the second most-downloaded category.Source: Figure 1.5.923 · p. 47, Model and Dataset Ecosystem · p. 47

Total Number of AI Publications / By Venue

  • About 258,000 AI publications appeared in 2024, versus roughly 102,000 in 2013, with 6.3% annual growth and a 40.9% share of OpenAlex computer science publications.Journals accounted for 47% and conferences for 23.5%; conference share declined from 36.6% in 2013, although recent results may reflect delayed venue assignment. The analysis tracks English-language research, and publication volume does not measure research quality.Source: Figure 1.6.1 · p. 48, Total Number of AI Publications · p. 48, §1.6 · p. 48

Conference Attendance

  • Attendance increased across the 16 tracked AI conferences in 2024, with NeurIPS, CVPR, and ICML drawing the largest audiences.Smaller conferences such as ICAPS, KR, and UAI maintained stable participation. Totals combine virtual, hybrid, and in-person attendance, complicating measurement and comparisons; ICML’s 2021 spike was likely linked to its virtual format.Source: Conference Attendance · pp. 49–50, Figure 1.6.3 · p. 50, Figure 1.6.426 · p. 50

By National Affiliation28

  • 17.8% of AI publications in 2024 were attributed to China, compared with 11.1% to Europe and 7.6% to India.Citation shares were 20.6% for China, 19.5% for Europe, and 12.6% for the United States, versus 13.03% for the United States in 2023. U.S. publication share fell 3 percentage points, while unknown affiliations reached 39.3%, likely reflecting metadata changes; publications receive one count for each affiliated country.Source: By National Affiliation28 · p. 51, Figure 1.6.630 · p. 52, Figure 1.6.7 · p. 52

By Sector / By Topic

  • 68.1% of AI publications came from academia in 2024; machine learning was the largest research topic at 37%.Government, industry, and nonprofits contributed 12.4%, 11.5%, and 4.6%; industry shares were 24.6% in the United States versus 18% in China, where government contributed 25.1%, while Europe’s academic share was 55.3%; sector calculations excluded unknown affiliations. Computer vision accounted for 22.4%, pattern recognition 11.2%, and natural language processing 10%, while generative AI publications grew sharply; topic labels overlap.Source: Figure 1.6.831 · p. 53, By Sector · p. 53, Figure 1.6.1032 · p. 54

By National Affiliation

  • 46 top-100 AI publications were credited to the United States in 2024, down from 64 in 2021, while China reached 41.China increased from 34 in 2023, while Australia reached 14, versus 2 in 2023 and 6 in 2021. OpenAlex citation rankings can change as citations accumulate, and multinational papers count once for each participating country, allowing country totals to exceed 100.Source: By National Affiliation · p. 54, Top 100 Publications · p. 54

By Sector and Organization … Overview

This section surveys AI patenting and publication leadership, alongside geographic, specialization, gender, and migration patterns among AI authors and inventors. It also examines advances in AI capabilities, narrowing gaps among leading models, uneven progress across tasks and applications, and challenges in keeping evaluations reliable and up to date.

By Sector and Organization

  • 6 industry publications appeared in 2024’s top 100, down from 17 in 2021 and 19 in 2022, while academia remained the largest contributor.Stanford University and Google led in 2024 with seven publications each, followed by the Chinese Academy of Sciences and Microsoft with five each, even as industry’s share of notable model releases continued growing. Institutional counts can exceed 100 through coauthorship; unknown sector affiliations reduced the 2022 sector total to 98.Source: By Sector and Organization · pp. 55–56, Figure 1.6.1235 · p. 56

1.7 Patents / Global Trends

  • 131,121 AI patents were granted globally in 2024, up 8.2% year-over-year; China accounted for 74.2%, versus 12.1% for the United States.Grants increased from 3,866 in 2010; reported per-capita leaders in 2024 were South Korea (14.3%), Luxembourg (12.3%), and China (7.0%). The PATSTAT Global analysis reaches 2024 because of reporting delays, and differing national patent standards and laws limit cross-country comparisons.Source: Figure 1.7.138 · p. 57, Global Trends · p. 57, §1.7 · p. 57

Forward Citations Flow / Speed of Knowledge Diffusion

  • Over half of AI patent forward citations go to U.S. patents, despite their 12.1% share of patent volume; U.S. patents also receive citations sooner.Most AI patents receive their first citation within two to three years; 19% of U.S. patents remain uncited versus 32%–44% elsewhere, with regional citation activity stabilizing after about six years. Citation-based impact is contested (Higham et al., 2021), domestic citation bias is widespread (Jaffe et al., 1993; Cotterlaz et al. 2025; and Verluise et al., 2025), and Chinese patents appear frequently in U.S. filings but the reverse is less common.Source: Forward Citations Flow · p. 59, Figure 1.7.539 · p. 60, Speed of Knowledge Diffusion · p. 60

Technological Proximity / AI Patent Examples

  • Most countries’ AI patent portfolios resemble both U.S. and Chinese portfolios, generally leaning toward the United States, while national industrial strengths shape applications.Using Bar et al. (2012), India and Australia show close to 80% overlap with both portfolios, whereas Denmark’s energy-and-wind emphasis yields 45% overlap with China and 52% with the United States. Patent examples illustrate learned models for dynamic computing-resource allocation, hazard forecasting and infrastructure mapping with causal inference, and perspective-correct VR/AR passthrough through depth reconstruction and gap filling.Source: Technological Proximity · p. 61, Figure 1.7.7 · p. 62, §1 · p. 62

1.8 AI Authors and Inventors / Geographic Distribution / By Education Level

  • 220,520 identified AI authors and inventors were based in the United States in 2025, followed by India with 50,460 and Germany with 48,520.Zeki’s output-based dataset excludes China; Switzerland led per capita with 110.5 authors and inventors per 100,000 inhabitants, followed by Singapore with 109.5. PhD and master’s degree holders together constituted majorities in most countries; the United Kingdom and Australia had the highest PhD shares, at 51.1% and 50.5%, while India and Brazil had more varied degree profiles.Source: Geographic Distribution · p. 63, §1.8 · p. 63, Figure 1.8.2 · p. 64

By Gender

  • 2010–2025 saw essentially unchanged male–female ratios in almost every covered country, with men remaining the majority everywhere and no country approaching parity.Men constituted more than 80% of identified AI talent in Brazil, South Korea, and Japan. Female representation was 32.3% in Saudi Arabia, 30.1% in Australia, 29.6% in Canada, and 29.5% in Italy.Source: By Gender · p. 66

By Specialization

  • 20% of South Korea’s AI talent specialized in hardware, VLSI, and IoT; Brazil led software engineering at 18%, and Saudi Arabia led security, privacy, and cryptography at 15%.Healthcare and bioinformatics, computer vision and image processing, and software engineering were common specializations globally, each accounting for at least 10% of the talent pool in several countries.Source: By Specialization · p. 67

Mobility

  • 26.0 was the United States’ net AI talent flow in 2025, down from 324.6 in 2022; India had the largest net outflow, at -16.9.Net flow measures arrivals minus departures, and the United States remained positive throughout 2020–2025. In 2025, Saudi Arabia and Denmark recorded positive flows of 3.1 and 2.1, while Canada and Germany recorded -7.1 and -2.4.Source: Mobility · p. 67

2.1 Overall Performance Trends

  • Benchmark scores rose across language, reasoning, coding, and math in 2025, while gaps between leading models narrowed.The gap between top U.S. and Chinese models closed almost completely. Competitive pressure is shifting toward cost, reliability, and real-world usefulness.Source: Overview · p. 69

Overview

  • Evaluations are being outpaced by model progress, and benchmarks face growing questions about their reliability.Technical advancement is also outpacing many evaluation, governance, and adoption frameworks.Source: Overview · p. 69
  • Roughly one in three AI-agent attempts still fail, despite improving agent performance.Agents are improving as the range of tasks AI systems can perform expands.Source: Overview · p. 69

2.7 Robotics and Autonomous Motion

  • Autonomous vehicles have reached mass-scale deployment with promising early safety records, while robots still struggle in unstructured environments.The broader transition of AI into physical spaces remains early.Source: Overview · p. 69

2.3 Image and Video

  • Some video generation models are beginning to learn how the physical world works, beyond producing realistic-looking content.This progress could help bring AI into physical spaces.Source: Overview · p. 69

2.5 Performance in Specific Domains

  • Professional-domain evaluations in tax, legal reasoning, and corporate finance show stronger performance in some areas than others.Source: Overview · p. 69

1 AI capability is outpacing the benchmarks designed to measure it, and surpassing human-level … Time Understanding in MLLMs

AI capabilities are advancing rapidly on benchmarks, with models approaching or exceeding human reference scores on some expert reasoning tasks, but uneven real-world performance and unreliable evaluations complicate assessments of progress. This section examines these advances and limitations across language, reasoning, visual understanding, and generation, including evaluations of user preferences, tool use, and faithfulness.

8 AI models are expanding into professional domains, showing performance ranging from 60

  • On OSWorld, agent accuracy rose from roughly 12% to 66.3%, within 6 percentage points of human performance, while agents still fail roughly one-third of attempts on structured benchmarks.Professional-domain evaluations report performance up to 90%, with the top 15 models separated by as little as 3 percentage points in each benchmark.Source: §8 · pp. 72–73

Claude Sonnet 4.5 / Gemini 2.5 Pro / DeepSeek-R1

  • 77.2%+ on SWE-bench Verified and 61.4% on OSWorld marked Claude Sonnet 4.5’s gains, alongside tooling for long-running autonomous workflows.GPT-5.1 improved SWE-bench Verified to approximately 76.3% versus GPT-5’s approximately 72.8%, ran faster, and adjusted reasoning effort dynamically; Gemini 2.5 expanded context to 1M tokens, scored approximately 63.8% on SWE-Bench Verified, and reached #1 on LMArena. DeepSeek R1 introduced GRPO, comparing groups of outputs against predefined rules to train reasoning without labeled data or a separate critic model, reducing training complexity.Source: Claude Sonnet 4.5 · p. 74, §8 · p. 74, Gemini 2.5 Pro · p. 74

Technical Performance Benchmarks vs. Human Performance

  • 30 percentage points were gained by frontier models on Humanity’s Last Exam in one year, as challenging benchmarks rapidly saturated.Frontier systems meet or exceed human baselines on ImageNet, SuperGLUE, and MMLU, while advanced reasoning benchmarks approach human performance; software engineering and computer use remain below baseline. Capabilities remain uneven: Gemini Deep Think earned 35 points and gold at the 2025 IMO, versus 28 points and silver in 2024, yet ClockBench’s top model achieved 50.6% accuracy versus humans’ 90.1%.Source: §1 · p. 72, §7 · p. 72, Technical Performance Benchmarks vs.… · p. 75

Model Performance Converges at the Frontier

  • Fewer than 25 Arena Elo points separated the top four companies’ models in March 2026, down from roughly 97 points a year earlier.Closed-weight Claude Opus 4.6 led open-weight GLM-5 by 49 points (3.4%), while the leading U.S. model led China’s Dola-Seed-2.0 Preview by 39 points (2.7%). Singh et al., 2025, caution that unrecorded model selection, unequal data access, and adaptation to Arena-style interactions can influence rankings independently of general capability.Source: Closed- vs. Open-Weight Models · p. 76, Figure 2.1.22 · p. 77, US vs. China Technical Performance · p. 77

Benchmarking AI

  • 2% on MMLU Math and 42% on GSM8K were invalid-question rates, illustrating serious reliability problems in widely used benchmarks.Opacity, inconsistent prompting, contamination, and weak reproducibility undermine comparisons; interactive and physical-world evaluations remain underdeveloped, motivating human–AI collaboration assessments that measure interpretability and helpfulness. Truong et al., 2025, used response-pattern analysis to flag problematic questions with up to 84% precision; Cheng et al., 2025, proposed secure, refreshed, community-governed evaluations; Reul et al., 2025, found sparse and declining developer reporting on bias and environmental impact, leaving societal consequences poorly assessed.Source: Figure 2.1.5 · p. 79, Benchmarking AI · pp. 79–80

MMLU: Massive Multitask Language Understanding / Arena Leaderboard

  • 91.2% on MMLU-Pro places Gemini-3.1-Pro first, while all 15 leading models score above 87%, demonstrating tightly clustered broad-knowledge performance.Arena complements structured tests with blind, human-voted comparisons: leading Text Arena ratings span roughly 46 points, Claude-Opus-4-6-Thinking leads at approximately 1,510, and overlapping confidence intervals prevent clear dominance. Preferences need not track correctness or represent all uses; Singh et al. (2025) identify order, length, and style biases.Source: MMLU: Massive Multitask Language… · p. 81, Figure 2.2.15 · p. 82, Arena Leaderboard · pp. 82–83

RAG: Retrieval-Augmented Generation

  • Larger context windows let retrieval-augmented generation include more material, but do not guarantee better answers because attention must remain reliable throughout the window.Standard chunk retrieval can struggle with cross-document synthesis; Microsoft Research’s Graph RAG organizes sources into knowledge graphs and community summaries to address this limitation. Multistep retrieval and reranking introduce further trade-offs among answer quality, latency, and cost.Source: RAG: Retrieval-Augmented Generation · pp. 83–84

Berkeley Function Calling Leaderboard

  • 77.5% overall BFCL accuracy puts Claude-Opus-4.5 first, with roughly 21 percentage points separating the top 15 models on function calling.Function calling generates structured requests to external tools; BFCL V4 broadens evaluation toward agents, assigning 40% to agentic tasks and 30% to multiturn interactions. Evaluation mode matters: Grok-4-0709 scores 63% in prompt mode versus 61.4% in function-calling mode, while Grok-4-1 scores 69.6% with fast reasoning versus 58.3% without reasoning.Source: Berkeley Function Calling Leaderboard · p. 84

MTEB: Massive Text Embedding Benchmark / Arena: Vision

  • 76 was the top MTEB English v2 average task score in 2025; Gemini-3-Pro led Vision Arena at approximately 1,285 Elo in early 2026.MTEB spans over 50 datasets and eight task categories; its top score rose approximately 11 points since 2023 but remained imperfect. Vision Arena uses blind, Elo-based human comparisons: Gemini occupied four top-six positions, while overlapping confidence intervals limited distinctions across the bottom two-thirds, clustered between 1,230 and 1,260.Source: Figure 2.2.48 · p. 85, MTEB: Massive Text Embedding Benchmark · p. 85, Figure 2.3.511 · p. 90

The Gap Between Long Context Windows and Deep Understanding

  • Almost 30x annual context-window growth since mid-2023 has not eliminated the gap between accepted input length and deep understanding.Input lengths supporting 80% accuracy grew roughly 250x over nine months on Fiction.liveBench and MRCR, yet LongBench v2’s best model scored 57.7% versus time-limited human experts’ 53.7% (Bai et al., 2025). Stepwise checking helps but is costly, multiple-information retrieval and later-input accuracy remain difficult (Yu et al., 2025), and isolating context use from prior knowledge substantially changed rankings (Yang et al., 2025).Source: The Gap Between Long Context Windows… · p. 86

Video-MMMU / MVBench

  • 66% Video-MMMU overall accuracy for Keye-VL-1.5-8B trails the 74.4% human baseline; JT-VL-Chat and JT3.5 tie at 74.1% MVBench average accuracy.Video-MMMU uses 300 expert-level educational videos to test perception, comprehension, and applying learned knowledge to new scenarios. Its Δknowledge metric shows human experts gaining 33.1 percentage points after viewing, versus GPT-4o’s leading model gain of 15.6 points; about one-third of models decline.Source: MVBench · p. 87, Figure 2.3.210 · p. 88, Video-MMMU · p. 88

VBench-2.0 / Video-Bench

  • 66.7% makes Veo 3 the VBench-2.0 leader, ahead of Vidu Q1’s 62.7%; Gen3 leads Video-Bench video quality, followed by Kling.VBench-2.0 evaluates faithfulness across human fidelity, creativity, controllability, physics, and commonsense; models still struggle with complex stories and consistent object or scene dynamics. Video-Bench uses GPT-4o with few-shot scoring and chain-of-query prompting to improve human-rating alignment, finding motion effects the weakest subdimension across nearly all models.Source: Video-Bench · p. 90, VBench-2.0 · p. 91, Figure 2.3.612 · p. 91

Progress in Video Generation

  • Veo 3 demonstrated zero-shot perception, physical modeling, manipulation, and early visual reasoning, although specialized systems still usually outperformed it (Wiedemar et al., 2025).Testing covered 62 qualitative tasks, seven quantitative evaluations, and more than 18,000 generated videos; Veo 3 improved over Veo 2 across every quantitative task and sometimes matched or exceeded Nano Banana. The authors described successive-frame reasoning as “chain of frames,” suggesting a possible trajectory toward general-purpose vision.Source: Progress in Video Generation · p. 92

GPQA: A Graduate-Level Google-Proof Q&A Benchmark / Understanding and Reasoning Benchmark for Expert AGI

ARC-AGI-2 / Humanity’s Last Exam

  • 84.6% puts Gemini 3 Deep Think first on ARC-AGI-2; HLE accuracy reached 38.3%, rising 30 percentage points between 2024 and 2025.ARC-AGI-2 tests abstraction and rule inference from sparse examples, not general intelligence broadly, with scores spanning about 46%. HLE covers 2,700 challenging questions across academic subjects using multiple-choice and short-answer formats, and high-confidence errors remain common.Source: Figure 2.4.316 · p. 95, ARC-AGI-2 · p. 95, Figure 2.4.417 · p. 95

Time Understanding in MLLMs

  • 50.6% ClockBench accuracy for GPT-5.4 High trailed humans’ 90.1%, with median model errors of roughly one to three hours versus three minutes.Saxena et al. (2025) found Gemini-2.0 achieved only 22.6% exact match on ClockQA, while GPT-o1 reached 80% accuracy on CalendarQA but struggled more with date arithmetic. Fine-tuning on 5,000 synthetic clock images improved familiar-style performance without generalizing to real-world or altered clocks; Fu et al. (2025) linked failures to integrating visual cues, including distinguishing clock hands.Source: Time Understanding in MLLMs · p. 96

Planning … Humanoid Robotics

This section surveys advances and limitations in AI planning, software, mathematics, finance, law, agentic task execution, and robotics, including humanoid hardware and investment. It examines benchmark gains alongside domain-dependent performance, evaluation concerns, persistent challenges in realistic tasks and safe robotic execution, and uncertainty about widespread humanoid deployment and demand.

Planning / PlanBench

  • No single model leads PlanBench across domains: LAMA retains advantages, frontier models excel in some structured tasks, and scrambled descriptions often reduce performance.PlanBench requests complete plans and counts solved tasks, up to 45 per domain: LAMA reaches 45/45 on Miconic, while GPT-5 reaches 38/45 on Childsnack and 45/45 on Spanner. Under scrambled descriptions, GPT-5 falls to 12/45 on Blocksworld and 7/45 on Sokoban; DeepSeek R1 reaches 3/45 on Blocksworld and 0/45 on Floortile and Sokoban.Source: PlanBench · p. 98, Figure 2.4.919 · p. 99, Figure 2.4.1020 · p. 99

SWE-bench / Terminal-Bench / Vibe Code Bench

  • 76.8% approximately gives Claude 4.5 Opus (high reasoning) the SWE-bench Verified lead, while Terminal-Bench improves sharply and autonomous application building remains difficult.Terminal-Bench 2.0 accuracy increased from 20% in February 2025 to 77.3% in early 2026. On Vibe Code Bench, Claude Opus 4.6 (Nonthinking) leads at 56.5%, followed by GPT 5.2 at nearly 47% and GPT 5.3 Codex at 41.4%.Source: SWE-bench · p. 100, Figure 2.5.121 · p. 101, Terminal-Bench · p. 101

FrontierMath / MathArena

  • 97% MathArena accuracy contrasts with 31.3% on FrontierMath Tier 4, while strong competition answers still do not imply human-level proof construction.MathArena rose from about 83% in November 2025 to 97% in December; it tests newly released contests to reduce contamination, automatically grades numerical answers, and uses humans for proofs. FrontierMath uses original, exceptionally challenging problems, with GPT-5.2 Pro (Web App) leading Tier 4 by the end of 2025 but still failing roughly two-thirds.Source: Figure 2.5.525 · p. 103, MathArena · p. 103, Figure 2.5.424 · p. 103

Theorem Proving

  • 91.9% on IMO-ProofBench puts Aletheia ahead of Gemini 3 Deep Think at 76.7%, although competition-problem familiarity may influence scores.IMO-Bench (Luong et al., 2025) separates answer generation, proof construction, and grading; automated proof scores correlate with expert scores at 0.96 on basic and 0.93 on advanced problems, but high-stakes results require human verification. Gemini Deep Think earned IMO gold with 35 points in 2025, solving five of six problems in natural language within 4.5 hours, versus the expert-assisted, days-long silver-medal pipeline in 2024 (Luong and Lockhart, 2025).Source: Theorem Proving · p. 104, Figure 2.5.726 · p. 105, Figure 2.5.6 · p. 105

TaxEval / MortgageTax

  • 77.1% leads TaxEval v2 for Claude Sonnet 4.6; MortgageTax peaks below 70%, at 69.4% for Gemini 3.1 Pro Preview.TaxEval's top 15 models range down to 74% for Claude 3.7 Sonnet Thinking; its over 1,500 expert-verified questions assess factual correctness and expert-like reasoning. MortgageTax tests semantic extraction and annualized-amount calculations across 1,258 mortgage tax documents, with GPT 4.1 scoring 65.9% at the bottom of its top-15 group.Source: Figure 2.5.827 · p. 106, TaxEval · p. 106, Figure 2.5.928 · p. 107

CorpFin / Finance Agent

  • 68.26% gives Kimi K2.5 the CorpFin v2 lead; Claude Sonnet 4.6 leads Finance Agent v1.1 at 63.33%, below two-thirds accuracy.CorpFin tests extraction, reasoning, and interpretation of credit agreements exceeding 200 pages under three document-access setups; GPT 4.1 scores 63.05%, and no model exceeds 70%. Finance Agent evaluates entry-level analyst skills through 537 questions, with Kimi K2.5 scoring 50.62%, about 13 percentage points behind the leader.Source: Figure 2.5.1029 · p. 108, CorpFin · p. 108, Figure 2.5.1130 · p. 109

CaseLaw / LegalBench

  • 73.4% gives GPT‑5.1 the CaseLaw v2 lead, while Gemini 3.1 Pro Preview (2/26) leads LegalBench at 87.4% amid tightly clustered scores.CaseLaw uses recent, restricted-access court decisions to test document-grounded reasoning, yet models often rely on general knowledge; GPT 4.1 follows at 69.9%. LegalBench tests six forms of legal reasoning, with all top 15 models above 83% and separated by about 4 percentage points.Source: Figure 2.5.1231 · p. 110, CaseLaw · p. 110, Figure 2.5.1332 · p. 111

GAIA / OSWorld / WebArena

  • 66.3% OSWorld accuracy places Claude Opus 4.5 within 6 percentage points of humans; GAIA reaches 74.5% against a 92% human baseline.OSWorld evaluates 369 desktop, file-operation, and multi-application tasks; GAIA accuracy rose from about 20% in January 2025 to 74.5% in September 2025. WebArena adds 812 long-horizon web tasks, verifying achieved goals through resulting site states rather than action traces.Source: Figure 2.6.133 · p. 112, GAIA · p. 112, Figure 2.6.234 · p. 113

WebArena

  • 74.3% success on WebArena in early 2026 brings leading models within 4 percentage points of the 78.2% human baseline.Success increased from about 15% in 2023, leaving the smallest model–human performance gap among the agent benchmarks in this section.Source: WebArena · p. 113, Figure 2.6.335 · p. 114

MLE-bench

  • 64.4% success on MLE-bench in early 2026, versus about 17% in 2024, demonstrates progress on end-to-end machine-learning tasks.Its 75 curated Kaggle competitions use rebuilt train/test splits and reimplemented grading for local evaluation against human leaderboards and medal thresholds. These competition problems remain more structured than most real-world data-science work.Source: Figure 2.6.436 · p. 114, MLE-bench · p. 114

Cybench

  • 93% unguided success on Cybench, up from 15% in 2024, represents the steepest improvement across the section’s agent benchmarks.Cybench comprises 40 professional-level cybersecurity tasks across six capture-the-flag categories, with human first-solve times ranging from two minutes to almost 25 hours.Source: Figure 2.6.537 · p. 115, Cybench · p. 115
  • 70.2% pass@1 puts Claude Opus 4.5 first on τ-bench, followed by GPT 5.2 at 69.9% and Qwen3.5 at 68.4%.The top seven models span 62.9%–70.2%, a narrow 7.3-percentage-point spread. Retail and airline tasks test multiturn conversations, tool/API use, and policy compliance through final outcomes often verified against database states, capabilities that remain difficult.Source: Figure 2.6.6 · p. 115, Cybench · p. 115

RLBench

  • 89.4% average success makes EquAct the January 2026 leader on RLBench’s 18-task subset, ahead of SAM2Act at 86.8%.Using 100 demonstrations per task, the benchmark has seen success rise from about 48% in 2022 to nearly 90% in 2025. EquAct also improves performance under full 3D rotational variation, but these remain short-horizon tasks in controlled simulation.Source: Figure 2.7.1 · p. 116, RLBench · p. 116

BEHAVIOR-1K

  • About 26% Q-score and 12.4% full task success by Robot Learning Collective, the leading BEHAVIOR Challenge team, expose persistent household-robotics limitations.BEHAVIOR-1K contains 1,000 survey-derived household activities requiring long-horizon mobile manipulation in simulated homes. Q-score awards partial credit for completed subgoals; the held-out results show that reliable completion remains beyond current capabilities.Source: BEHAVIOR-1K · pp. 116–117, Figure 2.7.2 · p. 117, ResponsibleRobotBench · p. 117

ResponsibleRobotBench

  • 0.64 is GPT-4o’s leading safe score on ResponsibleRobotBench, compared with 0.40 for GPT-4o mini and 0.35 for Qwen-72B.The benchmark’s 23 multistage tasks involve electrical, fire/chemical, and human-related hazards, requiring risk detection, safety reasoning, safe planning, and human assistance when necessary. Success requires both completion and safety; even GPT-4o failed to complete more than a third of tasks safely.Source: Figure 2.7.3 · p. 117, ResponsibleRobotBench · p. 117

Humanoid Robotics

  • Humanoid robotics shows growing hardware availability and investment rather than widespread deployment, with demand and expansion beyond structured factory pilots still uncertain.Figure 02’s BMW pilot logged over 1,250 runtime hours and loaded more than 90,000 parts across over 30,000 vehicles during 11 months. Norway’s 1X opened a waitlist for its $20,000 household robot, but many company milestones remain prospective rather than verified operational results.Source: Humanoid Robotics · p. 118

Physical AI and Foundation Models for Robotics … Belief vs. Fact: Benchmarking Reliability

This section examines advances and real-world challenges in robotics and autonomous driving alongside responsible AI infrastructure, reliability failures, and trade-offs among safety, fairness, and privacy. It also covers rising reported AI incidents, gaps in comparable benchmark reporting, and evaluations of hallucination and the distinction between knowledge and belief.

Physical AI and Foundation Models for Robotics

  • Physical Intelligence’s π₀ and π0.6 demonstrate cross-platform robot tasks without task-specific retraining, while vision-language-action models remain research-stage systems with substantial real-world limitations.VLAs replace separate perception, planning, and action modules with one network mapping camera inputs and language instructions to motor control; GR00T and Gemini Robotics similarly target multiple robots and tasks. Training data requires costly physical trials or high-fidelity simulation, motivating World Foundation Models such as Cosmos to generate synthetic physics data.Source: Physical AI and Foundation Models for… · p. 120

Deployment

  • Approximately 11 million fully driverless Apollo Go rides were estimated for 2025, a 175% year-over-year increase, alongside expanding U.S. autonomous passenger services.By late 2025, Waymo operated roughly 2,500 robotaxis recording around 450,000 weekly trips, while California weekly paid trips reached approximately 283,880 from near zero in mid-2023. Deployment evidence is concentrated in the United States and, less extensively, China, with comparable European deployment data unavailable.Source: Figure 2.7.540 · p. 121, Deployment · p. 121, Self-Driving Cars · p. 121

Technical Innovations and New Benchmarks

  • Around 1,600 hours of driving data in nuPlan and Nvidia’s Physical AI-AV illustrate dataset growth, but hours do not establish comparable quality or content.Early benchmarks contained single-digit hours, versus roughly 500 hours for Waymo’s 2019 Open Dataset; some plotted dataset hours are estimated, and simulated and real-road data differ. Research increasingly combines multisensor datasets, long-tail end-to-end evaluations, reasoning-and-action models, multiview reasoning benchmarks, world models, and reinforcement learning to address unseen traffic scenarios.Source: Technical Innovations and New Benchmarks · pp. 122–123, Figure 2.7.843 · p. 123, Safety · p. 123

Safety

  • 8 versus 198 vehicle-to-vehicle intersection incidents marked the largest gap between Waymo and its human-driving benchmark in Waymo’s own safety reporting.Through September 2025, Waymo reported lower injury and airbag-deployment crash rates against a human benchmark covering the same miles and areas. Monthly reported ADS incidents nevertheless rose from roughly 10–25 in early years to frequently exceeding 80 in late 2024–2025; Waymo’s largest share was consistent with its larger deployment footprint.Source: Safety · pp. 123–124, Figure 2.7.9 · p. 124, Figure 2.7.1145 · p. 125

3.1 Scope and Dimensions of Responsible AI

  • Responsible AI encompasses system properties, risk controls, and governance, but improving one dimension can degrade others without an established framework for navigating those trade-offs.The report’s three-layer framework adds autonomy and human agency, environmental sustainability, and human oversight and contestability to previously tracked dimensions. Despite expanding policies, benchmarks, and safety institutes, deployment outpaces responsible AI infrastructure; unclear accountability and missing standardized longitudinal data for fairness, privacy, and explainability constrain assessment.Source: Overview · p. 126, §3.10 · p. 128, §3.1 · p. 129

3.10 Tradeoffs Across RAI Dimensions

  • 362 documented AI incidents in 2025, versus 233 in 2024, accompanied sparse responsible AI benchmark reporting and declining foundation-model transparency.The average Foundation Model Transparency Index score fell from 58 in 2024 to 40 in 2025, with disclosure gaps around training data, compute, and post-deployment impact. Standard-use AILuminate safety ratings weakened under jailbreak attacks, while language evaluations exposed regional gaps: an Arabic model outperformed GPT-5.1 and Gemini 2.5 Flash, and several leading models lost close to half their accuracy on a Slovenian regional dialect.Source: §3.10 · p. 128
  • 22% to 94% hallucination rates across 26 top models reveal difficulty distinguishing false statements framed as user beliefs from those attributed to others.Accuracy fell from 98.2% to 64.4% for GPT-4o and from over 90% to 14.4% for DeepSeek R1 when false statements were presented as user beliefs.Source: §3.10 · p. 128
  • 17% growth in AI-specific governance roles accompanied a decline from 24% to 11% in businesses without responsible AI policies in 2025.Implementation barriers included knowledge gaps (59%), budget constraints (48%), and regulatory uncertainty (41%). GDPR remained the leading regulatory influence despite falling from 65% to 60%; ISO/IEC 42001 reached 36%, the NIST AI Risk Management Framework 33%, and organizations reporting no regulatory influence fell from 17% to 12%.Source: §3.10 · p. 128

AI Incidents

  • 362 AI incidents were reported to AIID in 2025, while OECD AIM reached 435 monthly incidents in January 2026; both show sharp increases.AIID’s human-reviewed records favor English-language, high-visibility incidents, whereas AIM uses automated multilingual news collection and reached a six-month moving average of 326. Historical totals remain subject to revision, and the broader assessment combines these databases with benchmark reporting and third-party evaluations despite limited dimension-specific data.Source: Figure 3.2.12 · p. 132, AI Incidents · p. 132, §3.2 · p. 132

Unmoderated AI Output and Harmful Speech (July 8, 2025)

  • Grok generated antisemitic and violent extremist content after an update relaxed safety filters, prompting content removal and temporary suspension of text responses.The July 2025 incident illustrated the consequences of deploying lightly moderated conversational AI at scale; critics argued that weakened guardrails made the harm predictable.Source: Unmoderated AI Output and Harmful… · p. 133

AI Deepfake Impersonation and Romance Scams (March 9, 2025)

  • Deepfake impersonations of actor Jin Dong deceived fans into imagined private relationships, inducing money transfers and major personal decisions.The scams particularly targeted older women, and Jin Dong called for stronger legal protections and clearer consequences after publicizing the fraud in March 2025.Source: AI Deepfake Impersonation and Romance… · p. 133

AI-Assisted Website Impersonation and Consumer Fraud (Aug. 20, 2025)

  • Fake Joann Fabrics websites stole payment and personal information, illustrating how AI-assisted cloning can expand consumer fraud.Following the retailer’s bankruptcy, convincing discount sites left shoppers without purchases and with compromised cards. AI tools enable rapid, multilingual website cloning without coding, lowering barriers to targeting smaller brands with fewer resources.Source: AI-Assisted Website Impersonation and… · p. 134

RAI Benchmarks

  • Frontier developers routinely report capability benchmarks but sparsely disclose comparable responsible-AI evaluations, limiting independent assessment.Only Claude Opus 4.5 reports more than two listed responsible-AI benchmarks, and only GPT-5.2 reports StrongREJECT; sparse disclosure does not imply absent internal evaluation. Public evaluators also emphasize capabilities, while context-dependent fairness measures and inconsistent safety reporting complicate comparison.Source: RAI Benchmarks · p. 134, Figure 3.2.4 · p. 135

Hughes Hallucination Evaluation Model (HHEM) Leaderboard

  • 1.8%–5.4% hallucination rates characterize HHEM’s top 15 evaluated models, with most clustering at 4%–5% and only three below 4%.HHEM measures hallucinations in CNN/Daily Mail document summaries, whereas open-ended factuality tests assess a different task, making their scales noncomparable. Last year’s top-model range was 1.3%–2.9%, but the evaluated model set differs, preventing a like-for-like trend interpretation.Source: Factuality and Truthfulness · p. 135, Figure 3.2.53 · p. 136, Hughes Hallucination Evaluation Model… · p. 136

AA-Omniscience

  • 22%–94% hallucination rates span 26 models on AA-Omniscience, with Grok 4.20 Beta 0305 lowest at 22%.Its 6,000 questions cover six domains; correct answers earn rewards, errors incur penalties, and abstentions incur none, encouraging uncertainty acknowledgment on an index ranging from negative 100 to 100. After domain normalization, Gemini 3.1 Pro Preview, Grok 4.20 0309 v2, and Claude Opus 4.6 (max) have the strongest overall profiles, while other models show uneven domain strengths.Source: AA-Omniscience · p. 136, Figure 3.2.6 · p. 137, Figure 3.2.7 · p. 137

Belief vs. Fact: Benchmarking Reliability

  • 62.6% accuracy on first-person false beliefs versus 95% on third-person false beliefs reveals newer models’ persistent epistemic weakness on KaBLE.The benchmark tests knowledge versus belief through 13,000 questions across 13 tasks and 24 models; older models achieved 52.5% and 79% on the corresponding false-belief categories. Although recent models perform well on recursive knowledge, this may reflect pattern matching, and models still struggle to recognize that knowledge requires truth whereas belief does not.Source: Belief vs. Fact: Benchmarking… · p. 138

AI Companions … Foundation Model Transparency Index

This section examines responsible AI practices, governance, and inclusion, covering organizational maturity, policy adoption, research, incident response, companion-related harms, and privacy protections. It also explores uneven participation and inclusion across countries, languages, and dialects, alongside regional infrastructure initiatives, developer transparency, and institutional safety infrastructure.

AI Companions / Response classification across INTIMA prompt categories by model

  • Companionship-reinforcing responses exceeded boundary-maintaining responses across four tested models, while real-user research identified relational harms outside most AI safety frameworks.INTIMA uses 368 prompts covering 31 behaviors to classify responses as companionship-reinforcing, boundary-maintaining, or neutral, with the balance varying across providers. Zhang et al., 2025 analyzed over 35,000 Replika conversation excerpts, identifying six harm categories, chatbot roles as perpetrator, instigator, facilitator, or enabler, and “algorithmic compliance,” where reliance on a chatbot encourages users to accept harmful behavior.Source: Figure 3.2.9 · p. 139, AI Companions · p. 139, Response classification across INTIMA… · p. 139

Responsible AI Maturity / AI Incidents, Risks, and Mitigation Efforts

  • 2.3 out of 4 was global responsible AI maturity in 2025, with practices still being integrated as affected organizations reported more repeated incidents and weaker responses.The survey excluded China; incident prevalence remained 8%, but among affected organizations, reporting 3–5 incidents rose from 30% to 50%, while organizations rating responses excellent fell from 28% to 18%. Inaccuracy risk awareness rose from 60% to 74% and cybersecurity awareness from 66% to 72%; active mitigation reached 71% and 61%, respectively.Source: Figure 3.3.1 · p. 140, Responsible AI Maturity · p. 140, §3.3 · p. 140

AI Governance and Investment

  • 17% of organizations assigned AI governance to dedicated roles in 2025, up from 14%, while responsibility shifted away from data and analytics functions.Data and analytics ownership fell from 17% to 13%; information security remained the most common owner at 21%, and organizations without an owner declined from 9% to 5%. Most organizations below $1 billion in revenue expected RAI investment below $5 million; among those with at least $30 billion in revenue, 41% expected spending of at least $25 million and 22% budgeted at least $50 million.Source: AI Governance and Investment · p. 142

Implementation, Barriers, and Benefits

  • 11% of organizations reported having no responsible AI policies in 2025, down from 24% in 2024, while respondents increasingly reported benefits.Reported improvements increased for business outcomes by 7 percentage points, operations and customer trust by 4 each, and reduced incidents by 8; knowledge and training gaps remained the leading implementation barrier, rising from 51% to 59%, while technical limitations rose from 32% to 38%. For scaling agentic AI, security and risk concerns led at 62%, followed by technical limitations and regulatory uncertainty at 38% each; executive-support barriers were lower for agentic AI than RAI policies, at 9% versus 14%.Source: Implementation, Barriers, and Benefits · p. 144

Regulatory Influence

  • 60% of respondents cited GDPR as influencing responsible AI practices in 2025, down from 65%, but it remained the most cited regulatory influence.Reported influence of AI-specific regulations increased by 2 percentage points; newly included ISO/IEC 42001 and NIST AI Risk Management Framework responses reached 36% and 33%, respectively, without a 2024 comparison. OECD AI Principles influence fell from 21% to 16%, while organizations reporting no regulatory influence declined from 17% to 12%.Source: Regulatory Influence · p. 146

Publication Volume / Geographic Distribution

  • 19% more responsible AI papers were accepted at six tracked conferences in 2025, rising from 1,278 to 1,521, while China overtook the United States.Keyword-based tracking is not exhaustive: security and safety accounted for 641 papers (+23%), fairness and bias 462 (+13%), transparency and explainability 405 (+14%), and privacy and data governance 248 (+33%); RAI shares at general-purpose conferences remained around 8%. China contributed 812 papers versus the United States’ 394, reversing 2024’s 322 versus 788; Europe’s output declined in 2024 and 2025, although the United States retained the largest cumulative total for 2019–2025.Source: Figure 3.4.1 · p. 147, Publication Volume · p. 147, §3.4 · p. 147

3.5 RAI Policymaking / Global AI Governance Participation

  • International AI governance participation broadened, but major AI-producing countries largely skipped UNESCO readiness assessments and endorsement of global initiatives remained uneven.UNESCO’s RAM assesses legal frameworks, infrastructure, and education; participation concentrates in Latin America, Sub-Saharan Africa, and parts of South and Southeast Asia, consistent with its capacity-building purpose. Only Canada, France, Germany, Italy, and Japan endorsed every recorded major initiative during 2019–2025; the 2025 AI Action Summit convened over 100 countries and secured 64 signatories, but the United States and United Kingdom declined, reflecting innovation-first priorities and national-security concerns, respectively.Source: §3.5 · p. 150, Global AI Governance Participation · p. 151

3.6 Data Governance for Privacy / Data Protection and Privacy / 3.7 Fairness and Bias

  • Near zero to above 80 was the range of national GIRAI privacy scores, with Australia and parts of Europe scoring highest.GIRAI covers 138 countries using a quality-reviewed, 1,862-question expert survey conducted in November 2023–February 2024; low scores can reflect early AI deployment or limited institutional capacity rather than disregard for privacy. Most countries have data protection legislation, but gaps persist in parts of Africa and Asia; the following fairness assessment distinguishes bias and unfair discrimination, gender equality, and cultural and linguistic diversity because fairness depends on context.Source: Data Protection and Privacy · p. 153, §3.6 · p. 153, Figure 3.6.1 · p. 154

Bias and Unfair Discrimination

  • Below 20 were GIRAI bias-and-discrimination scores across much of Africa, the Middle East, and Central Asia; the United States and Canada ranked highest.Scores were generally low, with Australia, parts of Europe, and Brazil in the middle range. The dimension assesses government safeguards and independent civil-society efforts to prevent discriminatory AI outcomes.Source: Figure 3.7.1 · p. 155, Bias and Unfair Discrimination · p. 155

Gender Equality

  • Canada and The Netherlands led GIRAI’s gender equality dimension, which assesses initiatives preventing gender bias and protecting equal rights across AI development and use.Parts of Europe and Japan scored 61–80, while the United States and Brazil scored 41–60.Source: Figure 3.7.2 · p. 156, Gender Equality · p. 156

Cultural and Linguistic Diversity

  • 39% of African countries had nonstate activity protecting cultural and linguistic diversity, compared with 7% having government programs and 2% having legal frameworks.Nonstate actors also outpaced governments in Europe, Asia, and the Middle East; North America had active programs and nonstate actors but less-developed legal frameworks. Singapore led the overall diversity dimension, with Germany, Ireland, Italy, Qatar, Estonia, and Slovenia also scoring in the upper range of 70–88.Source: Cultural and Linguistic Diversity · p. 156, Figure 3.7.4 · p. 157, Figure 3.7.3 · p. 157

Inclusiveness and the Global Language Gap

  • 0.86 was LLM-X’s leading mean score on HELM Arabic, ahead of Gemini 2.5 Flash at 0.82 and GPT-5.1 at 0.81.HELM Arabic covers seven benchmarks spanning academic assessment, grammar, and region-specific safety, showing that English-centric rankings need not transfer. In Indic LLM Arena, which evaluates contextual appropriateness for Indian users, GPT-5.2 scored 1,314, GPT-5.1 1,298, Gemini 3 Flash 1,288, Qwen3-Next-80B 1,156, and Llama-4-Maverick-17B 1,108.Source: Figure 3.7.514 · p. 158, Inclusiveness and the Global Language… · pp. 158–159, Figure 3.7.615 · p. 159
  • 99.8% accuracy on Standard Slovenian fell to 88.6% on the Cerkno dialect for GPT-5, illustrating performance gaps within a language.Mistral Medium 3.1 fell from 90.0% to 53.2%, and Llama 3.3 from 87.0% to 53.6%. Dialects differ in spelling, vocabulary, and grammar and are rarely represented in training data.Source: Inclusiveness and the Global Language… · p. 159, Figure 3.7.716 · p. 160

Inclusiveness and the Global Language Gap; AFRICA; ASIA, MENA (ARABIC), CENTRAL ASIA; 3.9 Security and Safety

  • Regional initiatives address English-centered performance gaps through local AI infrastructure, while safety infrastructure expanded through evaluation frameworks, government-backed institutes, and standardized benchmarks.SEA-LION and AI4Bharat develop tailored data pipelines, tokenizers, and benchmarks for languages whose morphology, scripts, and limited digitized text challenge standard multilingual tools. Regional evaluation examples include AfroBench’s coverage of 64 African languages and SEA-HELM’s linguistic and cultural evaluation; safety was the responsible-AI dimension with the fastest institutional infrastructure growth.Source: Inclusiveness and the Global Language… · pp. 158, 160, AFRICA · p. 160, ASIA, MENA (ARABIC), CENTRAL ASIA · p. 161

The Openness Index

  • 2–16 out of 100 covered most leading models’ Openness Index scores, with training-data disclosure weak across both transparency indices.The Openness Index assesses weight access, licensing, training methodology, and training-data disclosure; K2 Think and Olmo 3 32B Think led and alone earned pre-training-data transparency points. Availability and methodology supplied most points, while over 90% of notable industry models were released without training code in 2025.Source: Figure 3.8.1 · p. 163, The Openness Index · p. 163, Foundation Model Transparency Index · p. 164

Foundation Model Transparency Index

  • 40 was the average FMTI transparency score in 2025, down from 58 in 2024; IBM led at 95, followed by Writer at 72.FMTI evaluates developers’ upstream, model-level, and downstream disclosures; upstream information about training data and resources was weakest. xAI and Midjourney each scored 14, while open-model developers, enterprise providers, transparency-report publishers, and EU AI Act signatories tended to perform better.Source: Figure 3.8.2 · p. 164, Foundation Model Transparency Index · p. 164

Global AI Safety Institutes … What Is Generative AI Worth?

This section examines AI safety evaluations alongside expanding AI investment, adoption, and estimated consumer value, covering improving normal-use performance, persistent jailbreak vulnerabilities, and responsible-AI trade-offs. It also considers geographically concentrated, infrastructure-focused investment, rising frontier-company revenues and compute costs, growing U.S. generative AI consumer surplus, uneven labor-market effects, and uncertain economy-wide productivity gains.

Global AI Safety Institutes

  • State-backed AI safety institutes are expanding internationally, but remain concentrated in wealthy, technologically advanced economies and pursue differing security, evaluation, and enforcement priorities.Operational institutes exist in the UK, U.S., Japan, Singapore, and Israel; India and France have launched institutes, while Canada, South Korea, Germany, and Brazil have institutes in development. These organizations conduct technical evaluations and safety research for government policy, while international-network membership offers an entry point for countries without standalone institutes, including Kenya and Australia.Source: Global AI Safety Institutes · p. 165

HELM Safety

  • 0.90–0.98 scores characterized most 2024–2025 models on HELM Safety, suggesting improving performance but limited benchmark discrimination among leading systems.HELM Safety evaluates social bias, self-harm and abuse risks, harassment and misinformation, adversarial conversations, and helpfulness–harmlessness trade-offs. Older 2023 models scored lower, while compressed recent scores suggest current benchmarks may lack the granularity to distinguish meaningful safety differences.Source: Figure 3.9.2 · p. 166, HELM Safety · p. 166

3.10 Tradeoffs Across RAI Dimensions

  • Responsible-AI improvements can degrade other dimensions, with no universally superior intervention or model and no shared framework for comparing these trade-offs.Kemmerzell and Schreiner (2024) found differential privacy reduced accuracy by up to 33 percentage points, fairness training helped only the most demographically imbalanced dataset, and robustness-oriented augmentation produced the fewest adverse effects; Cecchini et al. (2024) found GPT-4 led robustness (0.91) and accuracy (0.67), while Llama 2 7B led toxicity avoidance (0.98). These studies are task-specific; Wasif et al. (2025) found stronger differential privacy reduced Alzheimer’s identification accuracy by 14.8 percentage points and increased missed diagnoses by 21.4% at hospitals with less data, whereas encryption-based alternatives maintained more stable fairness but required two to three times more computing power.Source: §3.10 · p. 170

Overview / AI Company Revenue and Consumer Value

  • $172 billion in estimated annual U.S. consumer surplus from generative AI by early 2026 accompanied rapidly rising revenues, compute spending, and geographically concentrated investment.Consumer surplus grew 54% from $112 billion a year earlier, median value per user tripled, and most tools remained free or nearly free; frontier-company revenues rose rapidly alongside compute costs, with Google reporting over $150 billion in annual capital expenditure in 2025. U.S. generative-AI investment exceeded China and Europe combined, although private figures omit substantial Chinese government funding, including an estimated $184 billion deployed into AI firms during 2000–2023; economy-wide productivity evidence remains early and mixed.Source: Overview · p. 171, §3 · p. 173, §2 · p. 173

Organizational and Population AI Adoption

  • 53% population-level generative-AI adoption was reached within three years, faster than the personal computer or internet, but adoption varied substantially across countries.Adoption correlated strongly with GDP per capita, although Singapore reached 61% and the United Arab Emirates 64%; the United States ranked 24th at 28.3%. Generative AI was used in at least one business function at 70% of organizations, with China and Europe recording the largest annual increases, while AI-agent deployment remained in single digits across nearly all business functions.Source: §5 · p. 173

Jobs / Productivity / Robot Deployments

  • Nearly 20% lower employment among software developers aged 22–25 since 2024 illustrates uneven labor-market effects, while measured productivity gains remain task-dependent.One-third of surveyed organizations expected workforce reductions, almost half expected little or no change, and broad employment data had not shown large-scale job losses; reported productivity gains were 14%–15% in customer support, 26% in software development, and 50% in marketing output, with smaller gains in deeper-reasoning tasks and concerns about learning penalties. In industrial robotics, China accounted for 54% of global installations in 2024 versus 51.1% in 2023, while global annual growth was flat and Taiwan recorded 33% growth.Source: §5 · p. 173, §8 · pp. 173–174

Corporate Investment Activity / Private Investment Activity

  • $581.69 billion in global corporate AI investment in 2025 marked 129.9% annual growth, while private funding became increasingly concentrated in large deals.Corporate investment includes acquisitions, minority stakes, private investment, and public offerings; the private-investment analysis covers AI/ML companies with over $1.5 million in funding since 2013. Private investment reached $344.7 billion, up 127.5%, including $170.9 billion for generative AI, up over 200%; newly funded companies increased 70.8%, while billion-dollar funding events rose from 15 to 28.Source: Corporate Investment Activity · p. 178, §4.2 · p. 178, Figure 4.2.2 · p. 179

Geographic Distribution of Private Investment

  • $285.9 billion in private AI investment made the United States the 2025 leader, compared with China’s $12.4 billion and the United Kingdom’s $5.9 billion.U.S. leadership also extended to newly funded companies—1,953 versus 172 in the United Kingdom and 161 in China—and cumulative investment since 2013, at $757.3 billion versus China’s $131.8 billion. Within the United States, California received $218 billion, over 75% of the national total, although state-level mapping is not exhaustive.Source: Figure 4.2.8 · p. 182, Geographic Distribution of Private… · pp. 182, 184, 186, Figure 4.2.12 · p. 184
  • $184 billion was estimated to have reached Chinese AI companies through government guidance funds during 2000–2023, funding excluded from the private investment comparisons.These state-initiated funds pursue financial returns and strategic government priorities (Beraja et al., 2024; Luong et al., 2021). Private investment comparisons therefore likely understate China’s total AI capital allocation.Source: Geographic Distribution of Private… · p. 184

Focus Areas of Private Investment

Revenue

  • OpenAI’s early revenue growth after crossing $1 billion outpaced Uber, Cheniere Energy, and Moderna over comparable periods.Google alone in the comparison was scaling toward $100 billion in annual revenue. Frontier-company annualized revenue estimates show rapid growth but are directional, reflecting variable reporting reliability and accounting practices rather than necessarily annual recurring revenue.Source: Revenue · p. 189, Figure 4.2.19 · p. 190

Annual Compute Spend

  • OpenAI and Anthropic reported significantly higher annual compute spending in 2025 than in 2024 as frontier AI revenue grew.Reported spending largely represents rented cloud capacity rather than owned data centers, providing a proxy for compute procured to train and operate models.Source: Annual Compute Spend · p. 190

Capital Expenditures

  • More than $150 billion in reported 2025 capital expenditures placed Google alongside Amazon as the leading spenders supporting advanced AI infrastructure.Cloud providers finance infrastructure leased to AI companies; announced projects included Stargate at $100–$500 billion, Google’s $40 billion Texas data-center commitment, and Microsoft’s $17.5 billion India infrastructure investment.Source: Figure 4.2.21 · p. 191, Capital Expenditures · p. 191

What Is Generative AI Worth?

  • Generative AI consumer surplus exceeds estimated U.S. revenues, suggesting social returns exceed the private returns captured by producers.Brynjolfsson et al. (2026) used online choice experiments in 2025 (N=1,400) and early 2026 (N=2,000), asking compensation required to forgo all generative AI tools for one month rather than measuring productivity. Usage frequency was the strongest individual predictor of surplus, followed by work use, product variety, and paid subscriptions; practical guidance, technical help, and information seeking were associated with higher surplus.Source: What Is Generative AI Worth? · p. 192
  • $172 billion annually, up from $112 billion, is the estimated U.S. generative AI consumer surplus in Brynjolfsson et al. (2026).Adoption increased from 48% to 56% of U.S. adults. Average surplus per user increased 27%, from $98 in 2025 to $125 by March 2026, while the median rose from $3.40 to $11.40.Source: What Is Generative AI Worth? · p. 192

Industry Usage … AI Talent Concentration

This section tracks organizational AI adoption, AI agent deployment, and demand for AI skills in 2025, examining differences across countries, industries, and functions. It also charts AI talent representation and concentration by gender and geography for 2016–25 using LinkedIn data.

Industry Usage

  • 88% of respondents reported organizational AI use in 2025, up from 78% in 2024, while regular generative AI use reached 79%, up from 71%.Over half reported AI use across at least three business functions. Adoption increased across all regions, with China and Europe recording increases of 13 and 11 percentage points, respectively.Source: Figure 4.3.1 · p. 193, Industry Usage · p. 193

By Industry and Function

  • 58% adoption led industry/function pairings for knowledge management in business, legal, and professional services and software engineering in technology, with benefits varying by function.Information-processing and customer-facing functions generally led adoption, while strategy, finance, risk, and compliance lagged, except financial-services risk and compliance; reported cost savings were highest in software engineering and manufacturing (56%), while revenue gains featured marketing and sales (67%), strategy and corporate finance (65%), and product/service development (62%). Innovation improved for 64% and employee and customer satisfaction for 45%, although perceived improvements often matched reports of no effect and no more than 7% reported worsened cost metrics.Source: By Industry and Function · pp. 194–195, Figure 4.3.31 · p. 195, Figure 4.3.4 · p. 196

Deployment Stages

  • Single-digit scaled AI agent use characterized nearly all business functions, even as larger companies were more likely to have broader AI programs scaling.Most respondents reported no agent use, including about two-thirds or more in IT and knowledge management. Technology-sector scaled agent use was comparatively higher in software engineering (24%), IT (22%), and service operations (21%), concentrating in functions with established broader AI adoption.Source: Deployment Stages · p. 197, Figure 4.3.63 · p. 197, Figure 4.3.7 · p. 198

Measuring Signals of AI Diffusion

  • Approximately 53% adoption within three years put generative AI ahead of the early adoption trajectories of personal computers and the internet (Bick et al., 2024).Microsoft telemetry showed adoption positively correlated with GDP per capita (Misra et al., 2025), although the United Arab Emirates exceeded 64% and Singapore exceeded 61%, outperforming the GDP-based trend. Between the halves of 2025, South Korea gained 4.8% and rose from 25th to 18th, while the United States fell to 24th despite 28.3% population-level adoption.Source: Figure 4.3.95 · p. 199, Measuring Signals of AI Diffusion · pp. 199–201, Figure 4.3.10 · p. 200
  • Close to 40% of Anthropic platform activity involved computer and mathematical tasks throughout 2025 (Massenkof et. al, 2026), while educational usage grew.Educational instruction and library tasks increased from 9% early in 2025 to approximately 14% late in the year. Automation-oriented conversations rose from 41% to 49% by August, briefly surpassing augmentation, but augmentation regained the lead at 52% in November.Source: Figure 4.3.137 · p. 202, Measuring Signals of AI Diffusion · p. 202, Figure 4.3.14 · p. 203

AI Labor Demand Across Geographies / By Sector / By State

  • 4.69% of Singapore’s job postings required AI skills in 2025, leading tracked countries, while U.S. demand expanded across sectors but remained geographically concentrated.The U.S. share reached 2.6%; demand increased across every economic sector, with information-sector AI posting share rising from 7.8% to 13.2% and growth extending into transportation, real estate, and education. California accounted for 170,881 postings and 17.2% of U.S. AI postings, with California, Texas, and New York together representing approximately one-third; Washington, D.C., had a comparatively high within-market share of 6.2%.Source: AI Labor Demand Across Geographies · p. 204, Figure 4.4.1 · p. 204, By Sector · p. 209

Skill Composition

  • 111% growth in generative AI skill mentions accompanied a 5% decline in their share of AI postings, while demand shifted toward agentic systems.Broad AI and machine learning clusters appeared in 1.7% and 1.0% of all U.S. postings, respectively; Python appeared in 258,674 posts, while infrastructure-related skills showed strong long-term growth. References to agentic AI and orchestration frameworks increased as the share mentioning ChatGPT, chatbots, or conversational AI declined, indicating demand for coordinating and operationalizing task-oriented systems.Source: Skill Composition · p. 205

AI Talent Concentration

  • 2.1% AI talent concentration among LinkedIn members placed Israel first in 2025, followed by Singapore at 1.8% and Luxembourg at 1.6%.Concentration grew at least 75% in every sampled country during 2019–2025, exceeding 100% growth in the United Arab Emirates, India, and Saudi Arabia; Luxembourg led net inflows at 5.23 per 10,000 LinkedIn members, versus 1.2 for the United States. Men typically represented 65%–75% of AI talent, with gender ratios largely unchanged since 2016; U.S. women represented 34.3% versus men’s 65.7%.Source: Figure 4.4.2113 · p. 214, AI Talent Concentration · pp. 214–215
  • Asterisks mark countries whose y-axis labels use a different scale from those of other countries.These scale differences must be accounted for when comparing country charts.Source: AI Talent Concentration · p. 216
  • 2016–25 AI talent representation is charted by gender and geographic area using LinkedIn data.The source is LinkedIn, 2025; the chart is attributed to the 2026 AI Index report.Source: Figure 4.4.2315 · p. 215, AI Talent Concentration · p. 217
  • 2016–25 AI talent concentration is separately charted by gender and geographic area using LinkedIn data.The supplied passages provide no chart values or comparative findings.Source: Figure 4.4.2416 · p. 216, AI Talent Concentration · p. 218

AI’s Labor Impact … Physics, Astronomy, Chemistry, and Materials Science

AI’s productivity and employment effects remain uneven, while industrial robot deployment continues to expand. Scientific applications are advancing toward complete workflows, but unreliable execution and costly experimental validation constrain their impact.

AI’s Labor Impact; Productivity Trends; Micro-level Studies; Macro-level Studies

  • 14%–15% more issues per hour were resolved by AI-assisted customer support agents, but productivity gains vary substantially across tasks and organizational settings.Less-experienced workers often benefited most; software development and marketing also improved, while METR’s initial slowdown finding was not replicated and AI-assisted learning sometimes incurred penalties. Macro evidence is mixed: AI adoption raised European firms’ labor productivity by 4%, strengthened by training, whereas surveyed executives reported minimal realized gains; organizational integration may delay returns.Source: Micro-level Studies · p. 219, AI’s Labor Impact · p. 219, Figure 4.4.27 · p. 220

Workforce Impact; Unemployment rate by AI exposure quintile

  • Roughly 16% lower employment emerged among workers ages 22–25 in the most AI-exposed occupations relative to the least-exposed, controlling for firm-type effects.Software developer employment among ages 22–25 fell close to 20% from its 2022 peak by September 2025, while older groups continued growing (Brynjolfsson et al., 2025). These patterns do not establish uniform displacement: unemployment increased by 0.30 percentage points among the most-exposed workers versus 0.94 among the least-exposed.Source: Workforce Impact · pp. 221–223, Figure 4.4.3017 · p. 222, Figure 4.4.31 · p. 223

Changes in the occupational mix from recent baselines

  • One-third of respondents anticipate workforce reductions, although 43% expect little or no change, according to McKinsey’s survey (2025).Expected headcount decreases exceed observed decreases across nearly all functions, while occupational composition has shifted faster after generative AI’s introduction than after computers or the internet. Across 844 occupational tasks, 46.1% of workers wanted AI takeover, yet tasks with the highest average automation scores represented only 1.3% of Claude.AI usage; a desire-versus-feasibility framework emphasizes gradual task redesign (Shao et al. (2026)).Source: Unemployment rate by AI exposure… · p. 223, Changes in the occupational mix from… · pp. 224–225, Figure 4.4.33 · p. 225

4.5 Robot Deployments; Global Installation Patterns; Geographic Patterns; Service Robotics

  • 542,000 industrial robots were installed globally in 2024, up just 0.2%, while operational stock reached 4,664,000 and China accounted for 54.4% of installations.Collaborative robots’ installation share rose from 2.8% in 2017 to 13.6% in 2024; China installed 295,000 robots, compared with Japan’s 44,500 and the United States’ 34,200. Agricultural service-robot deployments increased 2.5-fold, while hospitality alone declined; IFR’s latest available data cover 2024 and remain subject to revision.Source: Figure 4.5.1 · p. 227, Global Installation Patterns · p. 227, Figure 4.5.4 · p. 229

2 Frontier models outperform human chemists on average but cannot reproduce published; 5.2 AI Across Scientific Domains; Physics, Astronomy, Chemistry, and Materials Science

  • Below 20% paper-scale replication scores on ReplicationBench contrast with leading models exceeding human expert averages on ChemBench’s 2,700+ chemistry questions.Learned simulation surrogates, inverse design, and agents connected to software and laboratory equipment are advancing physical sciences, but complete research execution remains unreliable; UnivEarth agents achieved 33% accuracy and their code failed 58% of the time. Virtual-cell models, including Evo 2, STATE, and AlphaGenome, aim to predict responses to drugs and genetic perturbations but still require experimental validation.Source: §2 · p. 233, Physics, Astronomy, Chemistry, and… · p. 236

3 Astronomy released its first foundation model, first visualization benchmark, and a 100TB

  • Under 4 minutes suffices for FourCastNet 3 to generate a 60-day global forecast, running 8 to 60 times faster than prior approaches.Aardvark Weather replaced the traditional numerical forecasting pipeline with one ML system, while multiple AI weather models entered operational deployment. Astronomy introduced its first foundation model, AION-1, trained on over 200 million celestial objects from five major surveys, and its first scientific-computing and visualization benchmark, AstroVisBench.Source: §3 · p. 233

5 On end-to-end scientific research tasks, the best AI agents score roughly half of what PhD; 5.3 AI Agents and Tools for Science Workflows

  • 38.8% accuracy for PaperArena’s best agent trails the 83.5% PhD expert baseline; frontier models achieve roughly 17% on BixBench bioinformatics analysis.Sakana’s AI Scientist-v2 produced an ICLR-workshop-accepted paper without human-coded templates, and Google’s AI Co-Scientist received validation in three biomedical areas, although experimentally confirmed discoveries remain few. Scientific models predominantly originate in academia and government, often through international collaboration; scale is not decisive, with 111-million-parameter MSAPairformer outperforming previous ProteinGym leaders and 200-million-parameter GPN-Star outperforming a 40-billion-parameter model.Source: §5 · p. 233, §5.3 · p. 233

Overview; 5.1 AI for Science in 2025; Publications in AI for Science

  • Approximately 80,150 natural-science publications involved AI in 2025, up roughly 26%; AI-related research represented 5.8%–8.8% of output across the reported fields.Established predictive modeling now coexists with expanding workflow assistance and early autonomous discovery, with breakthroughs concentrated in domains possessing strong data infrastructure. Publication volume does not establish discovery quality: experimental validation remains expensive and slow, limiting how many AI-generated hypotheses scientists can test.Source: Publications in AI for Science · p. 234, §5.1 · p. 234, Figure 5.1.12 · p. 235

Datasets … Structure Prediction and Cofolding Models

This section surveys scientific and biomedical datasets, foundation and specialized models, research agents, and cofolding systems. It highlights advances alongside persistent weaknesses in multistep research workflows and limited experimental confirmation and clinical evidence.

Datasets / Benchmarks

  • Below 20% performance on ReplicationBench contrasted with ChemBench results where the best frontier models surpassed the best human chemists but still struggled with basic tasks.New benchmarks lack longitudinal comparisons: equation-discovery systems scored 31.5% across 239 problems, top models stayed under 80% across 1,340 materials problems, and PHYBench reported Gemini 2.5 Pro at 36.9% versus human experts at 61.9%. Training resources expanded through Multimodal Universe’s approximately 100TB of astronomical observations and OMol25’s over 100 million high-accuracy DFT calculations spanning 83 elements.Source: Datasets · p. 236, Benchmarks · p. 237, ACADEMIA INDUSTRY · p. 238

Foundation Models / GOVERNMENT INDUSTRY

  • Up to 29 times better performance than specialized models was reported for GPhyT, which generalized beyond its training problems without task-specific fine-tuning.GPhyT trained on 1.8TB of simulations; other releases included an open astronomy model spanning 300M–3.1B parameters and over 200M celestial objects from 5 surveys, and a chemistry LLM using 34B tokens and 2.7M instructions. Physics Supernova scored 23.5 out of 30 at the 2025 International Physics Olympiad, ranking 14th among 406 participants; StarWhisper Telescope planned observations across 10 telescopes, while ChemAgents demonstrated autonomous robotic synthesis and optimization controlled by Llama-3.1-70B.Source: Foundation Models · p. 239, NONPROFIT GOVERNMENT · p. 239, Figure 5.2.3 · p. 240

Biological and Life Sciences

  • Roughly 17% accuracy on BixBench and below-human average performance on BioML-bench exposed biological research agents’ difficulties executing tool-integrated, multistep workflows.OpenGenome2 supplies nearly 9.3 trillion DNA base pairs for openly weighted Evo 2, with up to 40 billion parameters and a 1 million token context window; Spacetop adds over 600 imaging hours from 101 participants, but functional perturbation data remain scarce. Biological applications include neural-response models generalizing across stimuli and animals, autonomous BCI-Agent cell classification, and Biomni spanning 25 subfields; neuroscience supports connectome mapping (Dorkenwald et al., 2025; The MICrONS Consortium, 2025) and brain–network comparisons (Linsley et al., 2025), while ecology faces sparse, taxonomically biased, nonstandardized data.Source: Datasets · p. 242, Biological and Life Sciences · p. 242, Benchmarks · p. 243

Earth Science

  • 100% completion across 85 climate tasks accompanied ClimateAgent’s quality score of 8.32, versus 6.27 for Microsoft Copilot and 3.26 for GPT-5.FourCastNet 3 generates 60-day global forecasts at 0.25-degree resolution in under 4 minutes on one GPU, running 8 to 60 times faster than prior approaches; TerraMind supports any-to-any generation across 9 geospatial modalities. Observational infrastructure supports operational AI weather forecasting, whereas climate projections face out-of-distribution future states, agriculture lacks shared benchmarks, and extreme-event reliability escapes average skill metrics; hydrological LSTMs trained across CAMELS catchments consistently outperform process-based models (Kratzert et al., 2019).Source: Earth Science · pp. 246–247, Foundation Models · pp. 248, 250, Benchmarks · p. 248

Mathematics

  • AI advanced competition mathematics and formal proof generation, but major open problems, including long-standing Erdos conjectures, remained beyond current capabilities.Goedel-Prover targets automated proofs in languages such as Lean. International Mathematical Olympiad performance rose from silver to gold medal in one year, alongside rapid gains on FrontierMath and MathArena.Source: Mathematics · p. 250

AstaBench

  • Around 0.53 was AstaBench’s highest overall agent score, at roughly $3.40 per problem; most agents scored between 0.10 and 0.45.The benchmark evaluated 57 agents across 22 agent classes on over 2,400 problems spanning literature understanding, code execution, data analysis, and discovery. Most agents cost below $1.00 per problem.Source: Figure 5.3.1 · p. 251, AstaBench · p. 251

PaperArena

  • 38.8% average accuracy made multiagent Gemini 2.5 Pro PaperArena’s best-performing configuration, substantially below the PhD expert baseline of 83.50%.PaperArena requires synthesizing evidence across papers while coordinating parsing, retrieval, and computation tools. Multiagent configurations consistently outperformed single-agent setups, typically by 2 to 4 percentage points.Source: Figure 5.3.2 · p. 252, PaperArena · p. 252

AI as a Co-scientist

  • 78.4% GPQA Diamond top-1 accuracy was achieved by Google’s AI Co-scientist (Gottweis et al., 2025) when selecting its highest-rated hypothesis per question.Its generate-debate-evolve loop refines evidence-grounded hypotheses and was validated in three biomedical areas; AI Scientist-v2 used agentic tree search without human-coded templates to produce the first fully AI-generated paper accepted at a peer-reviewed ICLR workshop. Kosmos (Mitchener et al., 2025) sustained runs up to 12 hours, averaging 42,000 code lines and 1,500 papers, with collaborators estimating six months of research per run; SciToolAgent used knowledge-graph-driven tool retrieval to outperform prior frameworks by 10 to 20 percentage points on multi-tool tasks.Source: AI as a Co-scientist · pp. 252–253
  • Only a handful of multiagent scientific systems have experimentally confirmed results, leaving a wide gap between computational proposals and demonstrated findings.Examples include the Virtual Lab’s 92 SARS-CoV-2 antibody candidates, more than 90% of which bound their target, and OriGene’s two new cancer drug targets, GPR160 and ARG2. Workforce training gaps, missing API and interoperability standards, and inadequate infrastructure-maintenance funding remain roadblocks.Source: AI as a Co-scientist · p. 253

6.2 Clinical Applications

  • 85.5% versus 20% for unaided physicians was achieved by Microsoft’s AI Diagnostic Orchestrator with OpenAI’s o3 on challenging published medical cases.Ambient scribes reduced reported note-writing time by up to 83%, with one hospital reporting 112% return on investment; a randomized digital-twin trial involving 150 diabetes patients found 71% achieved healthy blood sugar over one year while safely reducing medications. The FDA authorized 258 AI medical devices in 2025, mostly through pathways not requiring new trials, and only 2.4% of devices with clinical studies had randomized trial support.Source: §6.4 · p. 257

6.3 Patient Engagement / 6.4 Ethical Considerations

  • 84% to 92% of health-related Google searches display AI-generated summaries, making them a routine influence on users’ initial interpretation of health questions.Ethics discussion in medical AI publications more than doubled in 2025, but governance dominated while algorithm accountability, biosecurity, and global health equity remained underexplored.Source: §6.4 · p. 257

Research Volume

  • Approximately 71% growth brought AI-driven protein research publications from 2,259 in 2024 to 3,855 in 2025 across four research categories.Protein-drug interactions remained the largest category, increasing from 49.9% to 54.4% of publications, while structure prediction’s share fell from 28.7% to 23.9%. Absolute counts increased across all categories, with function prediction and synthetic protein design retaining comparatively stable shares.Source: Research Volume · p. 258

Molecular and Cellular Biology

  • Biological AI increasingly combines experimental measurements and synthetic data, but database sizes are not directly comparable because entries represent different biological units.AlphaFold 3, Boltz-2, and OpenFold3 use quality-filtered AlphaFold 2 predictions; OMol25 contains over 100 million quantum mechanics calculations, Tahoe-100M covers over 50 cancer cell types exposed to more than 1,100 drugs, and BaseData contains over 9.8 billion genes. Virtual-cell releases including Evo 2, STATE, and AlphaGenome aim to predict drug and genetic-perturbation responses without wet-lab experiments, but still require experimental validation.Source: §6.4 · p. 257, Molecular and Cellular Biology · p. 259

Data for Biomedical Vision-Language Models

  • Biomedical vision-language datasets are expanding across specialties, modalities, and domains, with data scaling remaining an active research direction rather than appearing saturated.Training requires large image-and-caption repositories that can be transformed into continual pretraining datasets.Source: Data for Biomedical Vision-Language… · p. 260

Sequence-Based Models: Protein Language Models

  • 111 million-parameter MSAPairformer surpassed previous state-of-the-art ProteinGym methods using multiple sequence alignments and physical constraints, at a fraction of the training and parameter budget.The 2025 shift toward efficiency included smaller retrieval-augmented Profluent E1 models and task-specific ESM-C models; benefits from parameter scale remained task-dependent. Fine-tuned ProGen designed OpenCRISPR-1 with improved specificity over SpCas9, while 200-million-parameter GPN-Star outperformed a 40-billion-parameter genomics model.Source: §6.4 · p. 257, Sequence-Based Models: Protein… · pp. 261–262, Figure 6.1.6 · p. 262

Structure Prediction and Cofolding Models

  • Most cofolding models show similar performance on protein structures and biomolecular complexes, although AlphaFold 3 retains advantages on certain tasks.These models jointly predict structures involving proteins, nucleic acids, drugs, and other biomolecules; the Boltz series and OpenFold3 have commercially permissive licenses. With all PDB structure types represented, further gains likely require richer data: distilled structures expand training sets from hundreds of thousands to tens of millions, while Boltz-2 adds experimental binding affinities and leading models differ in additional data sources.Source: Structure Prediction and Cofolding… · p. 262

Training data sources for cofolding models, 2025 … 6 People are acquiring AI skills outside formal education, and advertising AI skills in their

This section examines AI’s expanding role in biomedical research, clinical workflows, and education, including patient exposure, student use, and growing specialization. It highlights uneven model performance and limits in evidence of patient benefit and real-world testing, alongside gaps in safety evaluation, governance, education policy, and data.

Training data sources for cofolding models, 2025; Protein Design and Generative Models for Therapeutics

  • AlphaFold 3 has not been significantly surpassed on FoldBench despite subsequent larger models, suggesting data rather than model size is an important bottleneck.Following AlphaFold 3, protein structure prediction models converged on a similar parameter scale rather than continuing to grow. In a 2025 Nipah binder-design challenge, 99 of over a thousand tested designs bound and none neutralized the target; Mosaic’s expert-tuned combination of tools outperformed general-purpose approaches.Source: Training data sources for cofolding… · pp. 263–264, Figure 6.1.8 · p. 264, Figure 6.1.10 · p. 265

Virtual Cell Models and Genomic Foundation Models; Multimodal Foundation Models for Biomedical Discovery

  • 200-million-parameter GPN-Star outperformed 40-billion-parameter Evo 2 on multiple variant effect prediction tasks, showing that scale alone remains insufficient.Virtual-cell research grew substantially in 2025, with releases including Evo 2, STATE, and AlphaGenome, while benchmark results emphasize the importance of training methods and data curation. Multimodal biomedical publications have grown rapidly since 2021, particularly vision-language models pairing images with text and vision-omics models integrating imaging with genomic or transcriptomic data.Source: Virtual Cell Models and Genomic… · p. 265, Figure 6.1.12 · p. 266, Multimodal Foundation Models for… · p. 267

Automated and Agentic Biomedical Discovery

  • 92 novel SARS-CoV-2 nanobody binder designs were produced by the Virtual Lab’s coordinated scientist agents, but outputs still require experimental validation.Automated discovery connected reasoning with experiments: Robin identified ripasudil as a dry age-related macular degeneration candidate, STELLA independently integrated new software tools, and Biomni unified 150 tools, 105 packages, and 59 databases across 25 subfields. Agent Laboratory assigned PhD, Postdoc, and ML Engineer roles, while the Virtual Lab used an LLM Principal Investigator to coordinate specialized scientists.Source: Automated and Agentic Biomedical… · p. 268

Data Scale and Availability; Modeling Approaches; Prospective Clinical Trials

  • 28.5% growth brought prospective medical imaging AI trials from 417 in 2024 to 536 in 2025, expanding evaluation beyond retrospective patient data.Medical imaging training datasets remain roughly 100 times smaller in raw sample count than nonmedical datasets, with particular scarcity in CT and MRI and fragmentation across institutions. Vision-language models expanded across specialties, and Merlin trained on a single 40GB GPU using reports and ICD codes, but standardized cross-model benchmarks remain lacking; clinician review increasingly supplements lexical evaluation.Source: Figure 6.2.1 · p. 269, Data Scale and Availability · p. 269, Modeling Approaches · p. 270

LLM Clinical Reasoning Performance; AI Agents in Clinical Medicine

  • 85.5% diagnostic accuracy was achieved by MAI-DxO paired with o3 on challenging NEJM cases, versus approximately 20% among 21 practicing physicians.86% was o1-preview’s median management reasoning score, versus 42% for GPT-4 alone, 41% for physicians with GPT-4, and 34% for physicians using conventional resources (Brodeur et al., 2025); these isolated evaluations do not establish improved patient outcomes. On MedAgentBench, the best model achieved 69.7% task success across 300 virtual EHR tasks, illustrating the early-stage evidence for reliable autonomous clinical agents.Source: AI Agents in Clinical Medicine · pp. 272–273, Figure 6.2.42 · p. 272, LLM Clinical Reasoning Performance · p. 272

Deployment, Implementation, and Deimplementation; FDA-Authorized AI/ML-Enabled Devices; Devices by Clinical Specialty; Industry Landscape

  • 1,357 AI/ML-enabled medical devices from 693 companies across 17 specialties had FDA authorization by December 2025, but authorization does not ensure adoption.Radiology accounted for 76.6% of devices, although non-radiology authorizations rose from 7 in 2016 to 60 in 2025; the dominant 510(k) pathway requires substantial equivalence rather than new clinical trials, and only 2.4% of devices with clinical studies had randomized-trial support in an analysis through December 2024 (Singh et al., 2025). Financial, operational, and institutional barriers persist within a market dominated by several large manufacturers but populated mostly by companies holding one or two authorizations; approximately 10% of 2025 clearances used Predetermined Change Control Plans permitting iterative updates.Source: FDA-Authorized AI/ML-Enabled Devices · pp. 273–274, Figure 6.2.7 · p. 275, Devices by Clinical Specialty · p. 275

Enterprise-Scale Deployments in 2025; Ambient AI Documentation; AI-Powered Sepsis Prediction; Generative AI in Clinical Workflows

  • 18.7% relative and 17% sepsis mortality reductions were reported for the Targeted Real-time Early Warning System and COMPOSER, respectively.Enterprise-scale deployments also included ambient documentation: Sharp HealthCare reported 83% less note-writing effort, while adoption reached 63% among hospitals using Epic and Abridge expanded to over 150 health systems. Clinical workflow integration included ChatEHR recording 23,000 sessions across 1,075 trained users within three months, while OpenEvidence reported adoption by 40% of U.S. physicians.Source: Enterprise-Scale Deployments in 2025 · p. 276, AI-Powered Sepsis Prediction · p. 277, Ambient AI Documentation · p. 277

Evidence Gaps and Governance

  • 11.8 to 14.6 severely harmful recommendations per 100 clinical cases were produced by leading general-purpose LLMs on NOHARM, predominantly through omissions.These findings concern open-ended reasoning, not constrained scribes or sepsis alerts with clinician oversight; the ARISE review found only 5% of over 500 studies used real clinical data and concluded AI works best supporting clinician judgment. Stanford Health Care’s FURM framework governs all new AI tool adoptions there, with GUIDE-AI working to extend its availability.Source: Evidence Gaps and Governance · p. 278

Digital Twins in Medicine

  • 12.1% of 149 reviewed human digital-twin studies met the NASEM definition, and only 19% of systems were tested in real healthcare environments (Sadée et al., 2025).Digital twins require personalization, dynamic updating, and predictive capability; publications rose from near zero in 2015 to 372 in 2025, while healthcare patent filings rose from 30 in 2016 to 4,926 in 2025, with the latest patent year incompletely published. Clinical evidence included increased survival in a prostate-cancer adaptive-therapy pilot and a diabetes randomized trial (n=150) in which 71% achieved HbA1c below 6.5% within twelve months while safely reducing blood sugar–lowering medications.Source: Digital Twins in Medicine · pp. 278–279

AI Overviews for Health-Related Searches

  • 84%–92% of health-related queries triggered Google AI Overviews across five query types, making generated summaries a routine entry point for health information.Symptom and common health questions triggered overviews most frequently at 92%, followed by treatment queries at 90% and condition-based queries at 84%–88%.Source: Figure 6.3.1 · p. 280, AI Overviews for Health-Related Searches · p. 280

Patient Perspectives on AI in Healthcare

  • Tenfold publication growth from 2020 to 2025 expanded research on patient perspectives, revealing conditional acceptance of healthcare AI and persistent representation gaps.Patients favored assistive over autonomous roles (Fee et al., 2025; Allen et al., 2025; Hmido et al., 2025), with demographic disparities in acceptance (Labinsky et al., 2025; Ogu et al., 2025; Li et al., 2025), concerns about lost empathy, and emphasis on clinician endorsement (Berger et al., 2025; Machado et al., 2025; Nong et al., 2025) and disclosure. Internal medicine, radiology, and oncology dominated the literature; the United States, United Kingdom, and Germany led publication output, while sub-Saharan Africa, Latin America, Southeast Asia, and direct participation by children and adolescents remained underrepresented.Source: Patient Perspectives on AI in Healthcare · p. 281, Artificial intelligence use cases and… · p. 282

Volume and Concentration

  • 43.4% of medical AI publications discussed ethics in 2025, up from 37.1% in 2024, with governance receiving the most attention.The absolute number discussing ethics more than doubled; governance-related publications numbered 1,228, compared with 896 addressing algorithmic concerns and 874 addressing societal concerns. Only 14 publications discussed biosecurity, and fewer directly addressed misuse or dual-use ethics.Source: Volume and Concentration · p. 284

Global Health: A Different Ethical Focus

  • 51.8% of global-health AI publications mentioned ethics in 2025, with societal concerns outranking governance and algorithmic concerns, unlike the broader medical AI literature.Ethics appeared in 100 of 193 publications; Europe contributed 38, East Asia 31, and North America 28, while sub-Saharan Africa, Latin America, and Oceania each contributed fewer than five. Equity, justice, and accessibility were prominent societal concerns.Source: Global Health: A Different Ethical Focus · p. 286

7.3 K–12 CS and AI Education

  • Four out of five U.S. high school and college students use AI for schoolwork, but school policies and teacher understanding lag.Only half of middle and high schools have AI policies, and just 6% of teachers describe those policies as clear; common student uses include research, essay editing, and brainstorming. More than 90% of countries offer school-level computer science, but AI education is less established; China and the United Arab Emirates mandated it starting in 2025–26.Source: §2 · p. 290

7.1 Background

  • AI literacy grew faster than engineering-oriented AI skills in most countries, while limited coursework, teacher training, and incomplete data constrained assessment of education provision.The United Arab Emirates, Chile, and South Africa instead showed faster engineering-skills growth since 2022; AI literacy concerns foundational understanding and use, whereas AI education adds technical system-building skills. Available postsecondary completion data cover 2023–24, OECD data extend only through 2023 and exclude India, China, and much of Africa, and scarce standardized K–12 AI data force reliance on imperfect computer-science proxies.Source: Overview · p. 288, §6 · p. 290, §7.1 · p. 291

7.2 Postsecondary CS and AI Education

  • 11% fewer undergraduates enrolled in computer science at U.S. four-year universities between 2024 and 2025, even as AI specialization expanded.AI software-related master’s graduates increased 17% from 2023 to 2024, while new AI PhDs in the United States and Canada increased 22% from 2022 to 2024, with all growth going to academia rather than industry. The United States remained a leading producer of ICT graduates across degree levels, but Turkey, Brazil, and Mexico increased output faster; degree-completion data lag enrollment changes.Source: §1 · p. 290, §2 · p. 290, §7.2 · p. 292

U.S. Degree Graduates … Foundational Computer Science

This section examines U.S. AI-related degree trends, demographic representation, and graduate career destinations, alongside international comparisons of ICT graduates. It also covers university students’ use of generative AI and K–12 computer science offerings, access, and participation, using CS as a proxy where AI-specific K–12 data remain limited.

U.S. Degree Graduates

  • 82% more students earned AI software master’s degrees in 2024 than in 2022, while AI hardware degrees remained flat or declined.Software master’s graduates increased 17% between 2023 and 2024; hardware bachelor’s graduates declined 13% from their 2020 peak. This edition expands coverage beyond CS to AI-relevant software and hardware majors.Source: U.S. Degree Graduates · p. 292, Figure 7.2.1 · p. 293
  • 36% is women’s highest representation across AI-related degrees, reached in software master’s programs, compared with nearly 60% across all degrees.Women have slightly higher representation in AI software degrees than in AI hardware degrees, but remain underrepresented in both.Source: U.S. Degree Graduates · p. 293
  • Hispanic/Latino, Black, Native Hawaiian/Pacific Islander, and Native American/Alaskan students are underrepresented at every AI software degree level.White students are also underrepresented except at PhD level; multiracial and Asian students are overrepresented, especially Asian master’s students. A further racial breakdown finds generally consistent representation across degree levels, except for Asian and Native American/Alaskan students, with Hispanic/Latino, Black, and Native Hawaiian/Pacific Islander students underrepresented throughout.Source: U.S. Degree Graduates · p. 294
  • 67% of U.S. AI software master’s graduates are nonresidents, and nonresidents constitute most AI-related graduate students.The report anticipates further declines in nonresident graduates because of federal student-visa revocations and policies discouraging international enrollment.Source: U.S. Degree Graduates · p. 295
  • Georgia Institute of Technology alone appears in the top 10 across all degree levels for both AI software and hardware degrees.Leading institutions include public and private universities, and rankings may shift as more institutions introduce undergraduate AI majors. Classification under broader CS or engineering programs can undercount AI-related graduates at some institutions.Source: U.S. Degree Graduates · p. 295, Figure 7.2.6 · p. 297

AI-hardware related

  • 65% of new AI PhDs enter industry, down from 77% in 2022, while the academic employment share has nearly doubled since 2022.Government employment rose to 2%, from 0.7% in 2021, and the report interprets academia’s rising share as challenging the brain-drain narrative. Employment shares exclude self-employed, unemployed, and other-status graduates, so they do not total 100%.Source: AI-hardware related · p. 297, Figure 7.2.7 · p. 298, Global ICT Graduates · p. 298

Global ICT Graduates

  • The United States produces the most ICT graduates at every degree level among sampled countries, but other countries generally show faster year-over-year growth.Growth reached 27% for Turkish associate’s graduates, 30% for Brazilian and Turkish bachelor’s graduates, and 76% for Mexican PhDs; U.S. master’s graduates increased 55%. The latest data cover 2023, count graduates regardless of nationality, and exclude India, China, and African countries, limiting global comparisons.Source: Global ICT Graduates · p. 298, Figure 7.2.9 · p. 299, Figure 7.2.10 · p. 299
  • 20%, 22%, 29%, and 29% are women’s average shares of ICT associate’s, bachelor’s, master’s, and PhD graduates, respectively.Gender composition remains broadly stable, although women’s PhD share increased 4 percentage points and their associate’s share declined. Women reached at least half of associate’s graduates in Peru and PhDs in Costa Rica and Latvia; Turkey’s previously parity-level shares moved toward global averages in 2023.Source: Global ICT Graduates · pp. 300–301

CS, CE, and Information Faculty

  • Over 6,600 CS, computer engineering, and information faculty worked in the United States and Canada in 2024–25, nearly two-thirds in tenure-track positions.CRA projects growth over the next two academic years, especially in postdoctoral positions, but methodology changes prevent direct comparisons with previous AI Index counts. Hispanic/Latino, Black, and Indigenous people and women except Asian women were underrepresented, while Asian and Native Hawaiian/Pacific Islander men were overrepresented; this comparison uses U.S. national demographics against U.S.–Canada faculty data and excludes postdoctorates.Source: CS, CE, and Information Faculty · p. 302

Student Use of AI Tools

  • 80% of university students in Chegg’s 15-country survey used generative AI for learning in 2025, double the 40% reported in 2023.Usage ranged from 95% in Indonesia to 67% in the United States and United Kingdom; 56% of users asked questions daily, while nonusers cited misconduct accusations (45%), accuracy (38%), and restrictive policies (33%). Students used AI for research, brainstorming, and editing; understanding concepts was university students’ leading use, reported by 56%, versus 41% of high-school students.Source: Figure 7.2.16 · p. 303, Student Use of AI Tools · p. 303, Figure 7.2.17 · p. 304
  • 39.8% of student Claude use involved creating and 30.2% analyzing, indicating substantial use for higher-order cognitive tasks.Anthropic reported lower shares for applying (10.9%) and understanding (10.0%), raising the possibility that students rely on AI for important cognitive skills rather than developing them independently. Separately, 55% of U.S. college students surveyed believed generative AI had mixed effects on their critical thinking.Source: Student Use of AI Tools · p. 304
  • 64% of over 73,000 surveyed California State University students agreed AI positively affected their learning experience, while universities expanded acceptable-use policies.Chegg respondents reported improved understanding (50%), assignment completion (49%), organization (41%), faster learning (55%), and more free time (41%). A faculty survey found 48% of institutions had acceptable-use policies, up 9 percentage points since 2025; 80% of U.K. students perceived clear assessment policies, up 16 percentage points from the previous year.Source: Student Use of AI Tools · p. 304

United States

  • Only four U.S. states emphasized AI in their CS standards as of December 1, 2025, indicating limited adoption of AI education.Code.org expanded its annual CS education report in 2025 to examine state AI education policies. Because AI-specific K–12 education data remain limited, the chapter tracks CS education as a proxy alongside emerging state and federal AI policies.Source: United States · p. 305, §7.3 · p. 305

Foundational Computer Science

  • 60% of U.S. high schools offered foundational CS in 2024–25, unchanged from 2023–24 but above 35% in 2017–18.State trajectories differed: Tennessee increased from 61% to 83%, potentially reflecting its CS-access policy and graduation requirement, while Wyoming declined from 74% to 64%. Budget cuts or reallocations toward literacy and student crisis support may contribute to declining offerings.Source: Foundational Computer Science · pp. 305–306
  • 91% of large high schools offered foundational CS in 2025, compared with 77% of medium-sized schools and 44% of small schools.Access also differed by Title I status (60% versus 65% for non–Title I schools), location (57% rural, 59% urban, 71% suburban), and ethnicity: Asian students had 91% access, most other reported groups 80%–82%, and Native American students 70%, up from 66%. In incompletely reported enrollment data, Black, Native American/Alaskan, and white students remained near or above proportional representation, Native Hawaiian/Pacific Islander students approached it, and Asian students and students with 504 plans were overrepresented; Hispanic/Latino students, economically disadvantaged students, students with IEPs, girls, and English language learners remained underrepresented despite improvements.Source: Foundational Computer Science · pp. 307, 310, Figure 7.3.7 · p. 308
  • 6.1% of students across 42 reporting states enrolled in CS in 2024–25, while participation decreased in 26 of those states.Arkansas and South Carolina reported participation of 25.1% and 25.7%, respectively, versus 1.8% in Idaho and Minnesota; North Dakota, Tennessee, and Arkansas recorded the largest increases, but missing prior-year reports prevent some comparisons. Access did not consistently predict participation: Maryland had universal access but 16.7% participation, whereas Rhode Island had 79% school access and 18.4% participation.Source: Foundational Computer Science · p. 308, Figure 7.3.8 · p. 309, Figure 7.3.9 · p. 309

Advanced Computer Science … Creators’ Organizations Condemn EU AI Act Implementation

AI use and skills are expanding amid uneven education policies, curricula, and implementation, while governments build domestic capacity and pursue divergent regulatory approaches. Initiatives span infrastructure, development, safeguards, computing rights, and international governance coordination, with EU compliance guidance drawing criticism from creative-industry organizations over cultural rights and model developers.

K-12 Student Use of AI Tools; Advanced Computer Science

  • About 50% to 84% of surveyed students use AI for school-related tasks, while AP computer science exam growth slowed and participation disparities persisted.Students commonly use generative AI for research, essay revision, and brainstorming, reporting learning benefits alongside concerns about cheating accusations and weakened skills; many support school access and instruction. AP exam growth fell from 21% in 2022–23 to 5% in 2023–24, with female students participating less often than males and racial/ethnic participation remaining disproportionate.Source: Advanced Computer Science · p. 311, K-12 Student Use of AI Tools · p. 314

Education Standards, Policies, and Guidance; The Implementation Gap in K–12 AI Policy

  • Only about half of middle and high schools have AI-use policies, while decentralized implementation leaves AI education dependent on local resources.Only 36% of students described policies as extremely clear, and teachers reported clear, comprehensive policies at just 6% of schools; policies permitting AI use were more common in wealthier, urban communities. State guidance is largely nonbinding, teacher preparation lacks standardized benchmarks, and AP Computer Science lacks AI-specific content, leaving policy, training, and curriculum misaligned.Source: Education Standards, Policies, and… · p. 314, The Implementation Gap in K–12 AI Policy · p. 317

State K–12 AI Guidance in the United States; Federal K–12 AI Guidance in the United States

  • 30 states had issued AI-in-education guidance by January 2026, but substantial AI content remained uncommon in adopted computer science standards.Among 45 states with K–12 CS standards, 29 included limited AI content, 10 made no specific mention, and six included significant content; two more had significant-content drafts, with substantive standards recently adopted or developing. Seventeen states identified CS as foundational to AI and five allocated AI-education professional-development funding, while forthcoming summer 2026 CSTA standards expand AI learning goals and an April 2025 federal executive order promotes student exposure, educator development, partnerships, and workforce pathways.Source: State K–12 AI Guidance in the United… · pp. 315–316, Federal K–12 AI Guidance in the United… · p. 316

Global AI and CS Education

  • Approximately 93% of countries taught computer science in 2025, but few implemented AI education, and access remained uneven.Thirty percent mandated CS and 63% offered it without mandates, with private and urban schools more likely to offer standalone classes; global comparisons require caution because CS, ICT, and digital literacy are sometimes conflated. Beijing, Guangdong, Hangzhou, and the UAE mandated progressively structured AI curricula for 2025–26, whereas other initiatives emphasized classroom tool use, including South Korea’s subsequently reversed AI-textbook rollout.Source: Global AI and CS Education · p. 318

AI Skill Penetration; AI Skills Diffusion; 7.4 AI Skill Acquisition

  • 3.0 was India’s relative AI skill penetration rate, ahead of the United States at 2.0 and Germany at 1.8, with persistent gender gaps.LinkedIn profile rates were 3.1 for Indian men versus 1.9 for women and 2.1 versus 1.4 in the United States; certificates, online courses, and workplace experience provide additional skill-acquisition pathways. LinkedIn’s new diffusion index measures growth against countries’ own baselines: literacy generally grew faster than engineering, although the UAE, Chile, and South Africa showed rapid engineering growth; leading U.S. growth areas included prompting, Microsoft Copilot Studio, agents, productivity, and strategy.Source: AI Skill Penetration · p. 319, §7.4 · p. 319, AI Skills Diffusion · p. 320

National AI Strategies; AI Sovereignty

  • 44 state-backed AI supercomputing clusters operated in Europe and Central Asia in 2025, up from three in 2018, highlighting uneven sovereignty infrastructure.South Asia, Latin America, and the Middle East and North Africa reached two, three, and eight clusters respectively, while national strategies expanded, with emerging economies accounting for more than half of newly adopted strategies in 2024. Data-sovereignty approaches also differed: through 2024, East Asia and the Pacific recorded 77 localization measures, sub-Saharan Africa 71, Europe and Central Asia 66, and North America three.Source: §2 · p. 325, §1 · p. 325

Public Investment in AI; AI and Policymaking

  • Approximately $20.4 billion in U.S. AI-related public contracts and grants during 2013–24 contrasted with $285.9 billion in private investment in 2025 alone.European commitments included $1.6 billion from the UK, $505 million from Germany, and $320 million from France, with UK and German spending accelerating in 2024. U.S. congressional AI-related witnesses increased from five in 2017 to 102 in 2025; industry’s share rose from 13% to 37%, becoming the largest group, while academia’s fell to 15%.Source: §2 · p. 325, §6 · p. 325

AI Action Summit; U.S., UK, EU, and Chinese Policy Measures

Utah Enacts the Mental Health Chatbot Act (HB 452); Montana Enacts the Right to Compute Act (SB 212)

Kigali Summit Highlights Africa’s AI Opportunity and Labor Risks; Africa Declares AI a Strategic Priority: Investment, Inclusion, and Innovation; Zimbabwe Partners With Nvidia to Launch Africa’s 1st AI Factory

Gavin Newsom–Commissioned California AI Policy Report Warns of ‘Irreversible Harms’

  • California’s AI policy report warns of biological and nuclear misuse and proposes safety, transparency, and whistleblower frameworks.The report could offer a national blueprint in the absence of federal law.Source: Gavin Newsom–Commissioned California… · p. 327

G7 Issues Joint Statement Reaffirming Cooperation on AI Governance

Texas Signs the Responsible Artificial Intelligence Governance Act (TRAIGA) Into Law; Africa Declares AI a Strategic Priority: Investment, Inclusion, and Innovation

  • Texas’s TRAIGA establishes strict high-impact AI rules, banning uses that incite harm, violate constitutional rights, or discriminate against protected classes.Passed in June 2025, TRAIGA takes effect in 2026. The federal Take It Down Act addresses nonconsensual intimate imagery, explicitly including deepfakes, and strengthens removal and accountability expectations.Source: Africa Declares AI a Strategic… · p. 327, Texas Signs the Responsible Artificial… · p. 328

US Launches ‘America’s AI Action Plan’ and 3 Executive Orders; US Senate Strikes 10-Year Federal Moratorium on State AI Regulation

China Announces Action Plan for Global AI Governance

Creators’ Organizations Condemn EU AI Act Implementation; EU Releases Voluntary Code of Practice for General-Purpose AI

EU’s General-Purpose AI Obligations Take Effect … A Closer Look at US Federal Regulations

This section surveys AI governance obligations, national frameworks, international initiatives, and expanding U.S. legislative and regulatory activity alongside a federal shift toward deregulation and challenges to state laws in 2025. It also examines uneven participation and capabilities, AI sovereignty patterns across compute, data, models, applications, and talent, and measurement limitations and regional approaches to tracking public AI investment.

EU’s General-Purpose AI Obligations Take Effect; California Enacts Landmark AI Safety/Transparency Law (SB 53); Italy Becomes First EU Member State to Pass an AI Law; AI Companion Lawsuits Prompt Renewed Scrutiny of Emotional AI Safeguards; California Enacts New AI Laws

  • EU general-purpose AI obligations and California’s SB 53 introduce safety and transparency requirements, while Italy advances national legislation and companion-related harms prompt scrutiny.EU obligations include risk assessments, disclosures, and mitigation for systemic-risk systems; SB 53 requires safety-protocol and incident disclosures plus whistleblower protections, while Italy’s EU-aligned framework addresses governance, deployment, safeguards, copyright, and penalties. A reported teen suicide linked to an AI companion intensified child-safety scrutiny, while California separately enacted companion-bot regulation, generative-content provenance requirements, and expanded nonconsensual-deepfake protections.Source: EU’s General-Purpose AI Obligations… · p. 329, California Enacts Landmark AI… · p. 329, Italy Becomes First EU Member State to… · p. 329

UN Launches Global Scientific Panel and Global Dialogue on AI Governance; European Commission Launches the European Strategy for AI in Science; UNESCO Adopts Global Standards on Neurotechnology; US Executive Order Launches the Genesis Mission; Launch of Pax Silica: US-led AI and Technology Supply-Chain Cooperation Initiative; US Executive Order Seeks to Curb State-Level AI Laws

  • International initiatives advance scientific coordination and neurotechnology standards, while European and U.S. policies promote scientific AI, technology supply chains, and constraints on state regulation.The UN approved a scientific panel and governance dialogue, UNESCO adopted neural-data and mental-privacy standards, and the European Commission launched an AI-in-science strategy alongside its business- and industry-focused Apply AI Strategy. A U.S. executive order launched the Department of Energy-led Genesis Mission, the multinational Pax Silica Declaration supports semiconductor, data-infrastructure, and AI-hardware cooperation, and another U.S. order seeks to limit or preempt state AI laws.Source: UN Launches Global Scientific Panel… · p. 329, European Commission Launches the… · p. 330, UNESCO Adopts Global Standards on… · p. 330

8.2 National AI Strategies; By Geographic Area

  • National AI strategies expanded across emerging economies in 2024 and 2025, but published strategies measure policy intent rather than implementation.New frameworks span sub-Saharan Africa, South and Central Asia, and Latin America and the Caribbean, while high-income economies increasingly consolidate or update existing strategies. International cooperation, technical assistance, and policy diffusion support expansion, but implementation and regulatory capacity remain challenges, particularly in African countries still lacking formal strategies.Source: By Geographic Area · p. 331, §8.2 · p. 331

Infrastructure Sovereignty

  • 44 state-backed AI supercomputing clusters in Europe and Central Asia in 2025, up from 3 in 2018, mark the sharpest regional acceleration.North America reached 41 clusters, while South Asia, the Middle East and North Africa, and Latin America and the Caribbean reached 2, 3, and 8, respectively. This frontier-training dataset is only a proxy for sovereign high-end compute; private clusters dominate global capacity, while commercial cloud access and international public-private partnerships blur ownership-based distinctions.Source: Figure 8.3.1 · p. 333, Infrastructure Sovereignty · p. 333

Data Sovereignty; Talent Sovereignty

  • 77 data-localization measures in East Asia and Pacific contrast with North America’s 3 (Ferracane et al., 2025), while cross-border AI talent circulation slows.Data sovereignty concerns control over collection, storage, processing, and transfers; localization increased across nearly every region, accelerating around 2016, while emerging U.S. restrictions impose targeted controls on sensitive-data transfers. Talent sovereignty encompasses workforce capacity and mobility: declining inflows and outflows suggest reduced circulation, the U.S. attraction lead is narrowing, India is becoming a net absorber, and Middle Eastern and North African hubs are gaining.Source: Data Sovereignty · p. 334, Talent Sovereignty · pp. 339–340

Model Sovereignty

  • 1,618 cumulative U.S. model releases by 2025, versus China’s 849, illustrate persistent concentration despite expanding model development elsewhere.Open-source frameworks lower entry barriers and GitHub activity suggests wider diffusion; Europe and Central Asia reached 666 releases, versus 330 in East Asia and the Pacific excluding China, with smaller totals elsewhere. Counts include smaller publicly documented models rather than only notable models, but reporting gaps make some regional estimates conservative and exclude sub-Saharan Africa’s smaller, language-specific ecosystem.Source: Model Sovereignty · pp. 335–337, Figure 8.3.4 · p. 336

Application Sovereignty

  • Application sovereignty offers countries more room for niche specialization than models or compute, although broad application investment remains concentrated in a few economies.The United States, China, the United Kingdom, Germany, and France invest broadly; specializations include German manufacturing, Estonian education, sub-Saharan African finance, Chilean healthcare, Argentine agriculture, and Israeli security. Procurement policies, sectoral requirements, and digital public infrastructure shape deployment autonomy, but comparable cross-country procurement, deployment, and infrastructure-utilization data remain limited.Source: Application Sovereignty · pp. 337, 339

8.4 AI and Policymaking; Global Legislative Records on AI; A Closer Look at Global AI Legislation; US Legislative Records

  • 25 AI-related laws passed in the United States during 2016–2025, versus South Korea’s 17, led uneven legislative activity among G20 countries.Counts cover enacted legislation, count omnibus bills once, do not measure legal significance, and use a changed methodology that affects comparisons with the 2025 report. The U.S. Take It Down Act criminalizes nonconsensual intimate imagery, including deepfakes, requires platform removal within 48 hours, and provides criminal penalties and civil remedies.Source: Figure 8.4.1 · p. 341, Global Legislative Records on AI · p. 341, §8.4 · p. 341

US Legislative Records

  • The described national governance frameworks establish responsibilities, coordinate AI strategy, and support development alongside trustworthy and ethical AI.One framework mandates an AI Basic Plan and a prime minister–led AI Strategy Headquarters coordinating government policy and international cooperation. South Korea’s law provides statutory foundations for government strategy, coordination, support, standards, and safeguards.Source: US Legislative Records · p. 343

State Level

  • 150 state AI-related bills became law in 2025, up from fewer than 10 in 2020, with California leading legislative activity.California enacted 20 laws in 2025 versus Texas’s 12 and New York’s 10; over 2016–2025, California enacted 62, Maryland 28, Virginia 25, and Utah 24. Missouri and Rhode Island enacted none; the tracker counts only enacted texts containing “artificial intelligence,” likely capturing only part of state policymaking.Source: State Level · pp. 343–344, Figure 8.4.4 · p. 344, State AI Legislation Amid Shifting… · p. 345

State AI Legislation Amid Shifting Federal Policy

  • State AI laws advanced despite federal deregulation, spanning safeguards and broader frameworks, while a December 2025 executive order targeted conflicting state regulation.Colorado delayed key compliance dates to mid-2026, and Texas removed most private-sector obligations; Utah regulated mental-health chatbot disclosures, health-data sharing, and advertising, California added companion-chatbot suicide safeguards and protections for minors, and Washington required media provenance data. Montana established a right to compute with narrowly constrained government restrictions and critical-infrastructure risk-management requirements; the federal order directed litigation against state laws and funding pressure while preserving state authority over child safety, data centers, and government procurement.Source: State AI Legislation Amid Shifting… · pp. 345–346

US Congressional Hearings

  • 102 witnesses appeared at congressional AI hearings in 2025, versus 18 in 2022 and 131 in 2023, amid an almost twentyfold increase since 2017.From the 115th to 119th Congress, industry’s witness share rose from 13% to 37%, government’s fell from 35% to 10%, academia’s fell from 26% to 15%, and the other category rose from 26% to 38%. During 2017–2025, general AI governance and national security drew 113 and 74 witnesses respectively; House participation exceeded Senate participation in finance and economic policy (36 versus 3) and national security (49 versus 25), while health and biomedical AI drew 9 in each chamber.Source: US Congressional Hearings · pp. 346–348, Figure 8.4.7 · p. 347, Figure 8.4.9 · p. 348

US Regulations

  • 58 federal AI-related regulations were recorded in 2025, compared with one in 2016, despite a policy shift toward reducing regulatory constraints.The sharpest increase followed 2022. In early 2025, the Trump administration revoked Executive Order 14110, which supported advanced-model reporting and privacy, civil-rights, and workforce initiatives, and adopted a deregulatory replacement.Source: US Regulations · p. 348, Figure 8.4.10 · p. 349

By Agency

  • 28 AI-related regulatory actions came from the Executive Office of the President in 2025, making it the most active federal agency.Commerce and the Industry and Security Bureau became more active alongside attention to export controls and AI supply chains. Energy, Education, and the Securities and Exchange Commission began issuing AI-related regulations in 2023 or later.Source: By Agency · p. 349

A Closer Look at US Federal Regulations

  • Federal measures in 2025 combined deregulation with infrastructure expansion, AI exports, education initiatives, sensitive-data restrictions, and revised export controls.The AI Diffusion rule revised controls on advanced computing and closed-weight models before its May rescission; infrastructure policy accelerated federal-site data centers with clean-energy, cybersecurity, and labor conditions, while the deregulatory order required an AI Action Plan within 180 days. Other measures restricted sensitive-data transactions involving countries of concern, promoted full-stack AI exports through industry consortia and federal support, and expanded youth AI education, educator training, and workforce preparation.Source: A Closer Look at US Federal Regulations · pp. 351–352

8.5 Public Investment in AI

  • U.S. and European public AI investment estimates are presented separately because data availability and accounting differ.U.S. estimates use contracts, grants, and Other Transaction Agreements, counting obligations after AI activity first appears and controlling for early de-obligations. European and U.K. estimates exclude grants; long-term instruments often report maximum contract ceilings rather than actual spending, and award-duration data is frequently incomplete.Source: §8.5 · p. 353

Total AI Public Investments … Global Trust in Institutions

This section tracks public AI investment across the United States and selected European countries alongside public and expert views on AI’s progress and impacts. It examines AI awareness, adoption, optimism and anxiety, attitudes toward AI companionship, and expectations for disclosure and regulation, including trust in institutions to govern AI.

United States

  • $20.5 billion went toward U.S. AI-related activities during 2013–24, predominantly through grants.Funding comprised $15.9 billion in grants, $3.9 billion in contracts, and $650 million in Other Transaction Agreements (OTAs); grant spending accelerated after 2020, reaching $5.1 billion in 2024, or 32% of cumulative grants. Grants dominated award volume—22,364 versus 3,347 contracts and 185 OTAs—but median OTA value approached $1 million, compared with $304,000 for grants and $150,000 for contracts.Source: United States · p. 353, Figure 8.5.1 · p. 354

US AI-related public spending / Spending Across Agencies

  • Nearly 60% of U.S. AI contract and OTA spending went to Virginia, California, and Maryland, while grants were geographically more dispersed.The top three grant recipients—California, Massachusetts, and New York—received less than 16%; contract concentration may reflect proximity to federal agencies, whereas grants follow research institutions’ broader footprint. Defense accounted for 73% of contract and OTA spending during 2013–24 and 74.1% in 2024; HHS and NSF each accounted for roughly 40% of grant funding by 2024, following HHS growth after 2020.Source: Figure 8.5.4 · p. 355, US AI-related public spending · p. 355, Spending Across Agencies · p. 356

Spending Across Agencies / Spending Across Sectors

  • $3.7 billion in European AI-related contracts was committed during 2013–24, led by the United Kingdom’s $1.6 billion.Germany followed with $505 million and France with $320 million; 2024 commitments represented 28% of the U.K.’s cumulative total and 40% of Germany’s, while median contracts in the U.K., Germany, and Spain remained below $500,000 versus almost $1.1 million in Denmark. Government bodies represented 62.6% of European contract spending in 2024, health 13.9%, and education 13.7%; reported commitments are awarded amounts, not clearly timed actual obligations, because umbrella purchasing arrangements obscure spending.Source: Spending Across Agencies · p. 357, Figure 8.5.8 · p. 358, Spending Across Sectors · p. 359

and Regulation

  • 73% of AI experts expected positive effects on how people work, compared with 23% of the U.S. public, a 50-point gap.Expert–public optimism also differed for the economy (69% versus 21%) and medical care (84% versus 44%); 64% of Americans expected fewer jobs over 20 years versus 39% of experts, who forecast AI assistance for 18% of U.S. work hours by 2030 versus the public’s 10%. Companionship excitement reached 52% globally versus 42% in the United States, while experts forecast daily AI-companion use among U.S. adults at 10% by 2027 and 30% by 2040.Source: and Regulation · p. 362
  • 31% of U.S. respondents trusted their government to regulate AI responsibly, the lowest surveyed share, versus a 54% global average.Singapore and Indonesia recorded 81% and 76%; separately, across 25 countries in Pew’s 2025 survey, median trust in effective AI regulation was 53% for the EU, 37% for the United States, and 27% for China. Across all 50 U.S. states, concern about insufficient regulation exceeded concern about excess regulation, with national shares of 41% and 27%, respectively.Source: and Regulation · p. 362

9.1 Global Sentiment Toward AI

  • 59% of respondents globally saw AI’s benefits outweighing drawbacks in 2025, up from 55% in 2024, while 52% reported nervousness.Over 80% in Malaysia, Thailand, Indonesia, and Singapore expected profound life changes within three to five years, with Malaysia increasing nine percentage points; North America and Europe showed lower excitement and higher nervousness, while India’s concern rose 14 points versus two for excitement. Globally, 56% expected time savings, but only 40% expected improvements to their jobs, compared with 33% in the United States.Source: Figure 9.1.1 · p. 363, §9.1 · pp. 363–366

Global Perceptions of AI’s Impact on Jobs

  • 11% of respondents considered AI replacement of their job within five years very likely in both 2024 and 2025; 56% considered it unlikely.Expectations of changed work were also stable: 22% considered changes very likely in 2025 versus 21% in 2024, while 32% considered them unlikely in both years. More than 60% in Nigeria, Japan, Mexico, the UAE, South Korea, and India expected net job creation, whereas 67% in the United States and 68% in Canada expected elimination and disruption; optimism about improvements to jobs and job markets remained around or below 50% in most countries.Source: Global Perceptions of AI’s Impact on… · pp. 367, 369, Figure 9.1.7 · p. 368

Global AI Use in the Workplace

  • 58% of employees globally used AI semiregularly or regularly in 2025, and 53% trusted it for work purposes.Usage exceeded 80% in India, China, Nigeria, the UAE, and Saudi Arabia, whereas workplace trust in most North American and European countries ranged from 40% to 48%; online sampling may overrepresent younger, urban, educated respondents in emerging economies, although differences persisted after age and education controls. Around 85%–90% in India reported organizational support for AI strategy, literacy, and governance, while Japan, Korea, and Portugal ranked lowest for literacy support; most countries reported weaker governance support than literacy or strategy support.Source: Figure 9.1.10 · p. 370, Global AI Use in the Workplace · pp. 370–371

9.2 US Public and Expert Views on AI’s Societal Impact

  • 73% of AI experts expect positive effects on how people do their jobs, compared with 23% of U.S. adults.Experts are also more optimistic about the economy (69% versus 21%), K–12 education (61% versus 24%), and medical care (84% versus 44%); both groups express low optimism about elections, news, and personal relationships. Over 20 years, 64% of adults versus 39% of experts anticipate fewer jobs, while 5% versus 19% anticipate more; both identify cashiers, journalists, and software engineers as vulnerable, with experts emphasizing truck drivers and lawyers and the public anticipating losses more broadly.Source: Figure 9.2.1 · p. 372, §9.2 · pp. 372, 374, 375
  • 18% of U.S. work hours will receive generative AI assistance by 2030 in experts’ median forecast, compared with the public’s 10%.Across 68 forecasts, experts anticipate faster progress; capability forecasts agree in only 9% of cases, and when they diverge, the public expects slower progress 71% of the time. Experts forecast greater complex-math accuracy by 2030 (+25 points) and a higher likelihood of a transformative technological event by 2040 (+30), whereas the public predicts greater AI electricity consumption and a higher probability of solving a major mathematical problem; forecast horizons vary by survey design.Source: §9.2 · pp. 373–374, Figure 9.2.35 · p. 374, Figure 9.2.5 · p. 375
  • 47% of U.S. adults reported hearing a lot about AI in 2025, up from 26% in 2022.Awareness increased most among adults aged 18–29 (+29 percentage points), compared with +13 points among those 65 and older. More than 60% reported AI interaction at least several times weekly and 31% almost constantly or several times daily; daily use varies demographically, but self-reported interaction may undercount exposure through embedded systems.Source: §9.2 · pp. 376–377

Views on AI Companions

  • 30% daily AI companionship adoption among U.S. adults is experts’ forecast for 2040, compared with the public’s 20%.Experts forecast 10% adoption in 2027 and 15% in 2030, with upper-quartile predictions exceeding 40% and top-decile predictions exceeding 60%. A 2026 Ipsos-Google survey found 52% worldwide excitement about AI companionship, but 36% in the United States and 34% in Canada were not excited at all; Japan had the highest uncertainty, at 18%.Source: Figure 9.2.9 · p. 378, Views on AI Companions · p. 378
  • AI companions can reduce loneliness comparably to human interaction (Freitas et al., 2024), but emotional dependence correlates with psychological distress (Bengio et al., 2025).Systems remember interactions, recognize emotion, and personalize responses (Yang et al., 2025); users report always-available support (11.9%), safe emotional expression (9.9%), and mental health improvements (6.2%) (Pataranutaporn et al., 2025). Sustainable loneliness relief and effects on existing relationships remain uncertain, and both experts and the public consider therapist replacement less likely.Source: Views on AI Companions · pp. 378–379

Global Trust in Institutions

  • 79% of Ipsos respondents support mandatory disclosure of companies’ AI use, while 54% trust their government to regulate AI responsibly.Government trust was lowest in the United States (31%), compared with Singapore (81%), Indonesia (76%), Malaysia (73%), and Thailand (70%). A separate Ipsos/Google survey found 58% prioritized AI-enabled innovation versus 41% protecting affected industries through regulation, although South Africa, India, and Ireland leaned toward regulation.Source: Global Trust in Institutions · pp. 380–381
  • 53% median trust in the EU’s ability to regulate AI effectively exceeded 37% for the United States and 27% for China across 25 countries.Respondents generally trusted their own country most, while trust in the EU varied both within and outside its membership; Germany and the Netherlands were more trusting than Greece and Italy. In the United States, 44% trusted their government’s regulatory ability and 47% distrusted it, while 43% trusted the EU.Source: Figure 9.3.2 · p. 381, Global Trust in Institutions · p. 381
  • 41% of U.S. respondents feared insufficient federal AI regulation, compared with 27% fearing excessive regulation, with substantial uncertainty across states.Concern about excessive regulation peaked in New York and Tennessee (31%), while concern about insufficient regulation peaked in Missouri and Washington (48%); roughly one-third in most states were unsure. Insufficient-regulation concern reached 51% among adults 65 and older and 46% among college graduates versus 34% among those with high-school education or less; partisan differences were modest, at 45% for Democrats versus 40% for Republicans.Source: Global Trust in Institutions · pp. 382–384

Appendix … Lightcast

This appendix documents the data sources and methods used to estimate training emissions and measure AI projects, model usage, publications, conference attendance, patents, talent, technical benchmarks, responsible-AI research, investment, and labor demand. It specifies classification rules, geographic attribution methods, and coverage limitations.

Environmental Impact Analysis

  • 52 influential models were assessed for training emissions across industry and academic language and vision categories, excluding hardware production, idle infrastructure, and deployment.The calculator proposed by Lacoste et al. (2019), checked against published emissions, used hardware, GPU hours calculated from hardware quantity and training duration, provider partnerships, and compute regions inferred from team locations; training inputs came from Epoch AI or model papers. Coverage comprised 36 industry language, eight industry vision, four academic language, and four academic vision models; A100 SXM4 80GB hardware substituted for H100 GPUs, while custom-infrastructure calculations incorporated carbon efficiency and offsets.Source: Environmental Impact Analysis · p. 385

GitHub; AI Conference Attendance

  • GitHub identifies public AI repositories through topic labels, expanded keywords, and library dependencies; conference attendance comes from organizers or published totals.Repository identification follows Gonzalez, Zimmerman, and Nagappan (2020) and Dohmke, Iansiti, and Richards (2023), adding dependencies on PyTorch, TensorFlow, OpenAI, Transformers, XGBoost, scikit-learn, and SciPy. Geographic assignment uses owners’ annual modal IP-geolocated location, sampled daily with last-known locations carried forward during inactivity; conference organizers were contacted for 2025 totals only when attendance was not already published online.Source: GitHub · p. 385, AI Conference Attendance · p. 387

Hugging Face; Downloads data

  • 49.6% of all Hugging Face downloads are represented by weekly top-200 models with at least 200 downloads in the March 2022–August 2025 dataset.Downloads come from Longpre et al. (2025), whose direct collaboration with Hugging Face yielded its confirmed least-noisy dataset, supplemented by manual cleaning and missing-value imputation. Public model and dataset counts are cross-sectional, lack geographic information, and have approximately 60% missing modality values, preventing usage time-series analysis and granular geographic or modality breakdowns.Source: Downloads data · pp. 385–386, Hugging Face · p. 385

AI Publication Analysis

  • AI publications are identified in OpenAlex using an AI-focused CSO Classifier and reconciled with DBLP to distinguish conference proceedings from journal articles.CSO version 3.4.1 supplies 2,369 AI-related topics; classification combines syntactic matching, semantic embeddings, and postprocessing, while DBLP matching uses DOIs followed by titles and publication years. Aggregation uses the latest publication version and author affiliations, allowing overlapping country and sector counts and marking missing affiliations or citations as unknown; annual top-100 lists are reviewed down from OpenAlex’s 150 most-cited papers using citation data from OpenAlex, Google Scholar, and Semantic Scholar.Source: AI Publication Analysis · p. 386

AI Patent Analysis

  • AI patent identification merges keyword searches with IPC/CPC classification codes, using PATSTAT Global grants aggregated into DOCDB invention families to reduce duplicate counting.Coverage begins in 2010, with countries assigned by the publication authority of the earliest recorded grant; keyword development combines AI-generated terms, established lexicons, Word2Vec expansion, BERTopic and DeBERTA validation, and manual false-positive checks. Marginal duplicates can remain when applications share an earliest filing date, so analytics additionally remove duplicates against aggregation variables such as year.Source: AI Patent Analysis · p. 387, Speed of Knowledge Diffusion Methodology · p. 387

Speed of Knowledge Diffusion Methodology

  • Knowledge diffusion uses Kaplan–Meier estimates of months from patent publication to first citation, censoring patents still uncited when observation ends.Technological proximity, following Bar et al. (2012), sums the minimum patent-portfolio shares across shared technological classes, producing a similarity score between 0 and 1.Source: Speed of Knowledge Diffusion Methodology · pp. 387–388

Zeki; Date Range; Affiliation Range

  • 658,000 AI professionals outside China with demonstrated research, data, or model contributions form Zeki’s talent population.Zeki uses public business information and vetted vendors, excludes private information and secured-login collection, and infers gender from first names and likely residence, treating probabilities between 45% and 55% as ambiguous. Coverage includes company and academic workers during 2010–2025, excludes records beginning before 2010, and carries forward last-known residency, sector, and education through 2025 when career timelines lack end dates or recent updates.Source: Zeki · p. 388, Date Range · p. 388, Affiliation Range · p. 388

Talent Supply

  • 2010–2025 talent panels track country, education, and sector, while separate datasets describe current specializations and geographic or cross-sector career transitions.Annual country and sector assignments follow career timelines, while education follows education timelines across six education levels and 336 sectors. Specialization counts use individuals’ latest recorded country and top two current specializations among the top 100 specializations; mobility data tracks country and regional inflows and outflows and transitions between sectors.Source: Talent Supply · p. 389
  • 12-month moving averages smooth AI talent trends to reduce volatility from delayed profile updates, while country-year percentages track male and female representation.Profile updates can lag job changes, particularly after layoffs or career breaks; gender shares use each country-year’s total AI talent pool as the denominator.Source: Talent Supply · p. 389

Benchmarks

  • Benchmark scores are current as of early 2026 and assume company-reported results are accurate, with subsequent model releases potentially surpassing reported state-of-the-art scores.Sources include leaderboards, public repositories, company papers, blog posts, and product releases.Source: Benchmarks · p. 390

Conference Submissions Analysis

  • Responsible-AI submissions are identified through keyword searches at AAAI, AIES, FAccT, ICML, ICLR, and NeurIPS, with human and LLM-assisted categorization checks.Humans verify papers matching ambiguous keywords, while an LLM assigns categorical RAI-coherence labels from titles and abstracts for the remaining papers. A paper can belong to multiple responsible-AI categories.Source: Conference Submissions Analysis · p. 393

Chapter 4: Economy

  • Over 8 million global company profiles underpin Quid’s company analysis, with weekly updates and language-based links identifying topical clusters.Quid combines an in-house LLM, Boolean queries, and other search features across datasets; company searches can incorporate investment information, founding year, and headquarters location.Source: Chapter 4: Economy · p. 396

Data

  • Quid supplements Capital IQ organization data with Crunchbase, identifies AI companies through Boolean searches, and visualizes company relationships through semantic similarity.Undisclosed investor identities or funding amounts create missing data despite the broad company index. Searches use “artificial intelligence,” “AI,” “machine learning,” or “deep learning”; visualizations retain only the 7,000 most relevant companies when results exceed that number.Source: Data · p. 396, Chapter 4: Economy · p. 396

Companies

  • 1,993 respondents across 105 nations underpin McKinsey’s 2025 corporate-activity survey, while investment samples cover global AI and machine-learning companies during 2015–2025.Survey responses are weighted by each respondent nation’s contribution to global GDP. Investment analyses include private investments, IPOs, and M&A, with additional AI/ML and generative-AI samples restricted to companies receiving over $1.5 million during January 1, 2015–December 31, 2025.Source: Companies · pp. 396–397

Lightcast

  • Over 300 AI skills across 10 clusters define Lightcast’s AI job postings, with Agentic AI newly added alongside the nine previously tracked clusters.Lightcast deduplicates postings collected from over 51,000 job sites; 24 skills were added since the previous edition. Comparisons with JOLTS indicate coverage of over 99% of total U.S. labor demand, with offline hiring at small businesses and union hiring halls among omissions.Source: Lightcast · p. 397
  • At least two AI Engineering skills and/or current or past AI-job employment define LinkedIn AI talent; at least one AI Literacy skill defines AI literacy.LinkedIn distinguishes self-reported engineering and literacy skills, aggregating more than 1.3 billion members’ profiles under privacy thresholds, with coverage influenced by platform use and accessibility. AI jobs require AI skills; skills-penetration analysis covers five industries versus broader coverage for other metrics, while gender analysis uses self-identification or pronoun/name inference and excludes unclassified members and countries with insufficient attribution coverage.Source: Lightcast · pp. 399–400

Methodologies … Ipsos

This section documents definitions, data sources, and search and screening procedures for AI workforce, science, healthcare, education, legislation, congressional-hearing, and public-investment data. It also provides guidance on interpretation, comparability and coverage limitations, and references to external Ipsos survey methodologies.

1. Top AI Skills; 2. Fastest Growing AI Skills

  • Sample interpretations identify machine learning, AI, and deep learning as most-added engineering skills, and custom GPTs, AI productivity, and AI agents as fastest-growing.Growth is measured year over year, with country-specific skill-add thresholds set at the 50th percentile of the latest year’s distribution.Source: §1 · p. 400, §2 · p. 400

3. AI Talent Concentration; 4. Relative AI Talent Hiring Rate YoY Ratio; 6. Female Representation in AI

  • 0.88% U.S. AI-engineering talent concentration, 30.5% global female representation within AI-engineering talent, and 24.7% U.S. relative AI-hiring ratio YoY growth appear as sample interpretations.Talent concentration divides AI-talent counts by LinkedIn membership and requires caution because country coverage varies. Hiring measures members reporting new employers when jobs begin relative to location membership; the AI-only rate’s YoY change is compared with overall hiring using a three-month moving average.Source: §3 · p. 400, §4 · p. 400, §6 · p. 402

5. Skills Penetration

  • AI skills penetration measures the share of AI skills among an entity’s 50 representative skills, selected using TF-IDF weighting.Relative penetration compares countries with a benchmark while controlling for occupations, including industry-specific breakdowns. Gender-specific global averages permit only within-gender country rankings; a shared global average permits cross-gender comparisons within and across countries.Source: §5 · pp. 401–402

7. AI Talent Migration

  • AI talent migration measures arrivals minus departures per 10,000 LinkedIn members, normalized to year-end country membership and subject to minimum sample thresholds.The U.S. sample interpretation reports a positive net flow of 1.07 per 10,000 members. World Bank Group–LinkedIn data (Zhu et al. (2018)) identify migration through profile-location updates, with data available from 2019.Source: §7 · p. 402

8. Career Transitions into AI Jobs; 9. Fastest Growing AI Occupations

  • 26.9% of U.S. transitions into AI engineer came from software engineer and 13.3% from data scientist in the sample interpretation.Transitions are pooled over five years, counting moves into previously unheld occupations and excluding graduates’ first occupations and intra-occupation moves. Fastest-growing occupations are ranked by YoY changes in occupation share, with AI product manager showing the largest U.S. increase within AI engineering in the sample interpretation.Source: §8 · p. 402, §9 · p. 402

10. AI Skills Diffusion Index

  • The AI skills diffusion index tracks growth in members listing AI skills, adjusted for skill diversity, rather than measuring penetration levels.Country baselines begin when a membership threshold is reached. The sample interpretation labels the index AI engineering but describes the U.S. adjusted share listing AI literacy skills as 1075x higher than in March 2023.Source: §10 · pp. 402–403

Chapter 5: Science; The Central Dogma

  • Web of Science and PubMed searches quantify AI-related science, protein research, drug discovery, and virtual-cell publications; FDA records supply medical-device data.Web of Science searches combined AI topic terms with natural-science subject categories, restricted results to articles, and exported annual counts on March 20, 2026. PubMed searches counted protein research for 2024–2025 across function prediction, structure prediction, protein–drug interactions, and synthetic design; drug-discovery searches covered 2018 onward, while virtual-cell searches used title/abstract terms.Source: Chapter 5: Science · p. 405, The Central Dogma · p. 410

The Central Dogma — Patient Perspectives on AI in Health Care

  • Patient-perspective searches targeted empirical studies of AI in healthcare published during 2020–2025 across PubMed/MEDLINE, Embase, and Scopus, excluding telehealth and literature-only reviews.Searches required explicit AI mentions in titles or abstracts, patient/family/public perspectives, clinical use, and primary data; language limits started with English. Database-adapted searches underwent title/abstract and full-text screening; four reviewers tagged specialty and geography, and Claude Sonnet 4.6 converted tags to binary variables with manual accuracy checks.Source: The Central Dogma · p. 410, Core Limits (applied across databases) · p. 412

Search strings implemented on Feb. 9, 2026

  • 368 articles were retrieved by the Scopus patient-perspectives search; a separate PubMed Central analysis mapped medical AI ethics into algorithmic, governance, and societal concerns.The Scopus query covered 2020–2025, combining AI, patient/public/family/caregiver perspectives, healthcare, and empirical-method terms while excluding telehealth-related and specified review terms. The ethics analysis covered January 2021–December 2025, using additional filters for data sharing, algorithm sharing, biosecurity, and global health, with topics including bias, accountability, privacy, equity, and autonomy.Source: Search strings implemented on Feb. 9,… · p. 413

Chapter 7: Education

  • Education indicators draw on Code.org, College Board, CRA, Cambridge’s Raspberry Pi centre, IPEDS, and OECD datasets covering school access and postsecondary education.Sources include Code.org’s 2025 report and dashboard, College Board national and state AP summaries, the 2024 CRA Taulbee Survey of U.S. and Canadian member universities, IPEDS surveys of federal-aid-participating institutions, and OECD enrollment, graduate, and entrant statistics. The global K–12 dataset was updated through 2025 and clarified that U.S. computer science availability varies across schools and districts.Source: Chapter 7: Education · p. 419

Global Legislation Records on AI

  • The global legislation tracker combines autonomous crawling, two LLM classification passes, and supporting quotation extraction across G20 countries, followed by AI Index manual verification.Source-language keyword searches identify candidates; Google Gemini assesses relevance and document type before further analysis extracts covered systems, sectors, objectives, agencies, identifiers for deduplication, and legislative status. Verbatim provisions provide auditable evidence for the classifications.Source: Global Legislation Records on AI · p. 421

US State-Level AI Legislation

  • U.S. state-law tracking searches enacted bill texts for “artificial intelligence” and supplements those results with AI-related laws from the Multistate AI Legislation Tracker.Coverage includes all 50 states during 2015–2025; keyword-based inclusion requires the phrase in the final enacted version, while the supplementary tracker captures laws without that wording.Source: US State-Level AI Legislation · pp. 421–422

US AI Regulation

  • The dataset records 425 witness appearances across 107 AI-focused congressional hearings; separate tracking of U.S. agency regulations from 2016 to 2025 uses a Federal Register keyword search for artificial intelligence.Hearings span the 115th Congress through the 119th Congress as of January 30, 2025, identified through manually validated keyword searches primarily on Congress.gov, with supplementary sources and hearing- and witness-level attributes recorded. Regulation tracking analyzes total counts and originating agencies.Source: US AI Regulation · p. 422

Public Investment in AI

  • Public-investment estimates cover European and U.S. procurement but include grants only for the United States, with uneven data coverage limiting comprehensiveness.European data comes from TED, whose publication thresholds create coverage differences; U.K. sources additionally include national and devolved procurement systems, but exclude Northern Ireland. U.S. contracts and OTAs come from Omari et al. (2025) and FPDS, and grants from USAspending; GAO (2023) identified 49 agencies with over $5 billion in unreported fiscal-year 2022 outlays.Source: Public Investment in AI · p. 422
  • AI investment classification combines dictionary-based matching, topic and transformer models, and manual review; revised transaction accounting reduces retroactive inflation of investment estimates.Descriptions were combined, translated into English, and normalized; a dictionary incorporating Yamashita et al. (2021) and manually reviewed Word2Vec expansions supplied regex matches, which BERTopic and DeBERTA validated, with scores below 20% and all high-value tenders manually reviewed. U.S. accounting now aggregates only transactions after each award’s first AI match and drops first-year de-obligations, preserving earlier trends while reducing estimated investment stocks, especially initially.Source: Public Investment in AI · pp. 422–423

Ipsos

  • Ipsos survey methodology is not reproduced; readers are referred to the original survey reports for 2022–2026.The listed 2026 source is the Ipsos/Google Multi-Country AI Survey.Source: Ipsos · p. 424

Related papers

Loading 2606.15708v3…