PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial Intelligence Index Report 2025

Nestor Maslej, Loredana Fattorini, Raymond Perrault, Yolanda Gil, Vanessa Parli, Njenga Kariuki, Emily Capstick, Anka Reuel, Erik Brynjolfsson, John Etchemendy, Katrina Ligett, Terah Lyons, James Manyika, Juan Carlos Niebles, Yoav Shoham, Russell Wald, Toby Walsh, Armin Hamrah, Lapo Santarlasci, Julia Betts Lotufo, Alexandra Rome, Andrew Shi, Sukrut Oak

arXiv:2504.07139v3cs.AI

TL;DR

The AI Index addresses the need for rigorous, globally sourced evidence on rapidly changing AI trends by synthesizing longitudinal data across technology, research, business, policy, and society. It reports that AI mentions rose 21.3%, from 1,557 in 2023 to 1,889, and have grown more than ninefold since 2016.

  • Problem

    Rapidly intensifying AI activity across society, the economy, and governance requires accurate, globally sourced, longitudinal evidence to track and interpret change.

  • Method

    The Index synthesizes longitudinal trend data, using automated classification of computer-science publications and indirect methods to estimate training compute.

  • Results

    21.3%: AI mentions increased from 1,557 in 2023 to 1,889, while total mentions grew more than ninefold since 2016.

  • Takeaways & Limitations

    The Index provides context for understanding how AI is shaping current developments and collective decision-making.

  • Takeaways & Limitations

    Publication estimates may undercount AI research outside computer science, and this year’s findings differ slightly from previous reports because the data and classification methods changed.

Abstract

from arXiv · show

Welcome to the eighth edition of the AI Index report. The 2025 Index is our most comprehensive to date and arrives at an important moment, as AI's influence across society, the economy, and global governance continues to intensify. New in this year's report are in-depth analyses of the evolving landscape of AI hardware, novel estimates of inference costs, and new analyses of AI publication and patenting trends. We also introduce fresh data on corporate adoption of responsible AI practices, along with expanded coverage of AI's growing role in science and medicine. Since its founding in 2017 as an offshoot of the One Hundred Year Study of Artificial Intelligence, the AI Index has been committed to equipping policymakers, journalists, executives, researchers, and the public with accurate, rigorously validated, and globally sourced data. Our mission has always been to help these stakeholders make better-informed decisions about the development and deployment of AI. In a world where AI is discussed everywhere - from boardrooms to kitchen tables - this mission has never been more essential. The AI Index continues to lead in tracking and interpreting the most critical trends shaping the field - from the shifting geopolitical landscape and the rapid evolution of underlying technologies, to AI's expanding role in business, policymaking, and public life. Longitudinal tracking remains at the heart of our mission. In a domain advancing at breakneck speed, the Index provides essential context - helping us understand where AI stands today, how it got here, and where it may be headed next. Recognized globally as one of the most authoritative resources on artificial intelligence, the AI Index has been cited in major media outlets such as The New York Times, Bloomberg, and The Guardian; referenced in hundreds of academic papers; and used by policymakers and government agencies around the world.

Introduction to the AI Index Report 2025 … 7. AI optimism registers sharp increase among countries that previously showed the most skepticism.

The AI Index 2025 documents rapid advances in AI performance, adoption, affordability, investment, governance, education, science, and medicine, while emphasizing persistent challenges involving safety, trust, bias, access, and complex reasoning. It provides globally sourced, longitudinal data to help stakeholders understand AI’s present development and societal impact.

  • Top Takeaways: AI performance advanced sharply, with scores rising 18.8, 48.9, and 67.3 percentage points on MMMU, GPQA, and SWE-bench, respectively.AI systems also made major strides in video generation and, in some time-limited programming settings, outperformed humans.
  • Top Takeaways: AI adoption and investment accelerated: organizational use reached 78% in 2024 from 55% in 2023, while U.S. private AI investment reached $109.1 billion.Generative AI attracted $33.9 billion globally, and research indicates AI boosts productivity and often narrows workforce skill gaps.
  • Top Takeaways: The AI landscape is becoming more globally competitive: U.S. institutions produced 40 notable models versus China’s 15 and Europe’s three, while Chinese models approached parity on major benchmarks.Performance gaps between leading U.S. and Chinese models narrowed substantially across MMLU, MMMU, MATH, and HumanEval during 2024.
  • 7. AI becomes more efficient, affordable, and accessible. Driven by increasingly capable small models, the inference: AI became cheaper and more accessible: inference costs for GPT-3.5-level performance fell over 280-fold, while open-weight models narrowed their performance gap with closed models from 8% to 1.7%.Hardware costs declined 30% annually and energy efficiency improved 40% each year, lowering barriers to advanced AI.
  • 8. Governments are stepping up on AI—with regulation and investment. In 2024, U.S. federal agencies introduced: Governments expanded AI regulation, infrastructure investment, and safety coordination, including 59 U.S. federal AI-related regulations and major national commitments such as China’s $47.5 billion semiconductor fund.Legislative mentions of AI across 75 countries rose 21.3% in 2024, while international organizations released frameworks emphasizing transparency, trustworthiness, and fairness.
  • 5. The responsible AI ecosystem evolves—unevenly. AI-related incidents are rising sharply, yet standardized RAI: Responsible AI progress remains uneven: AI-related incidents reached 233 in 2024, a 56.4% increase, while standardized evaluations remain rare and companies often lag in mitigating recognized risks.New benchmarks such as HELM Safety, AIR-Bench, and FACTS provide tools for assessing safety and factuality, while governments increased cooperation on AI governance.
  • 9. AI and computer science education is expanding—but gaps in access and readiness persist. Two-thirds: AI education and scientific use expanded, but access and readiness gaps persisted: two-thirds of countries offer or plan K–12 computer science education, while fewer than half of U.S. teachers feel equipped to teach AI.AI received major scientific recognition through Nobel Prizes in physics and chemistry and the Turing Award for reinforcement learning.

Chapter Highlights · 5. AI models get increasingly bigger, more computationally demanding, and more energy intensive. · 8. AI hardware gets faster, cheaper, and more energy efficient. New research suggests that machine learning

The AI Index 2025 reports accelerating AI development, with industry increasingly leading notable model production, China leading publication volume, and the United States leading highly cited research and notable models. AI systems are becoming more computationally demanding and carbon intensive, while hardware improvements and falling inference costs make AI increasingly affordable to deploy.

  • Chapter Highlights: Industry produced nearly 90% of notable AI models in 2024, up from 60% in 2023, while academia remained the leading producer of highly cited publications.Industry’s share of notable model production reached 90.2% in 2024, whereas academia continued to lead top-cited research.
  • Chapter Highlights: China led 2023 AI publication volume and citations at 23.2% and 22.6%, while U.S. institutions contributed the most top-100-cited publications.The United States also produced 40 notable AI models in 2024, compared with China’s 15 and Europe’s combined three.
  • Chapter Highlights: AI publications more than doubled from approximately 102,000 in 2013 to over 242,000 in 2023, rising from 21.6% to 41.8% of computer science publications.The growth reflects increasing AI interest across computer science and related disciplines.
  • 5. AI models get increasingly bigger, more computationally demanding, and more energy intensive.: Training compute for notable AI models doubles approximately every five months, while LLM training datasets double every eight months and training power requirements double annually.Industry investment continues to drive model scaling and performance gains.
  • 5. AI models get increasingly bigger, more computationally demanding, and more energy intensive.: The cost of querying a GPT-3.5-equivalent model fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024, a more than 280-fold reduction.Depending on the task, LLM inference prices fell by 9 to 900 times.
  • 8. AI hardware gets faster, cheaper, and more energy efficient. New research suggests that machine learning: AI hardware performance grew 43% annually from 2008 to 2024, doubling every 1.9 years, while fixed-performance costs fell 30% annually and energy efficiency increased 40% annually.The H100 delivers 22 billion FLOP per second per dollar, approximately 1.7 times the A100’s and 16.9 times the P100’s price-performance.
  • 8. AI hardware gets faster, cheaper, and more energy efficient. New research suggests that machine learning: Training emissions rose from 0.01 tons for AlexNet in 2012 to 588 tons for GPT-3, 5,184 tons for GPT-4, and 8,930 tons for Llama 3.1 405B.These figures contrast with 18 tons of annual carbon emissions for the average American.

Chapter 2: Technical Performance (cont’d) … 9. Complex reasoning remains a problem. Even though the addition of mechanisms such as chain-of-thought

The 2025 AI Index reports rapid technical progress across model openness, international competition, reasoning, benchmarking, and video generation, while complex reasoning remains sensitive to time budgets. Open-weight and Chinese models are catching up, frontier performance is converging, and new benchmarks expose persistent limitations despite major gains.

  • 2. Open-weight models catch up: By February 2025, the performance gap between the leading closed-weight and open-weight models narrowed from 8.04% to 1.70% on the Chatbot Arena Leaderboard.The report attributes this improvement to the emergence of high-performing open-weight models such as Llama 3.1 and DeepSeek V3.
  • 3. The gap between Chinese and US models closes: By the end of 2024, U.S.–China performance gaps narrowed to 0.3, 8.1, 1.6, and 3.7 percentage points on MMLU, MMMU, MATH, and HumanEval, respectively.The overall gap between the top U.S. and Chinese models also fell from 9.3% in January 2024 to 1.7% by February 2025.
  • Chapter 2: Technical Performance (cont’d): The gap between the top and 10th-ranked Chatbot Arena models shrank from 11.9% to 5.4%, while the gap between the top two fell from 4.9% to 0.7%.These changes indicate increasingly convergent frontier performance and a more competitive AI landscape.
  • 5. New reasoning paradigms like test-time compute improve model performance: Test-time compute improved reasoning performance, with o1 scoring 74.4% on an International Mathematical Olympiad qualifying exam.The report presents this as an example of new reasoning paradigms improving model performance.
  • 6. More challenging benchmarks are continually proposed: New benchmarks reveal substantial remaining difficulty: Humanity’s Last Exam reaches only 8.80%, FrontierMath 2%, and BigCodeBench 35.5%, versus a 97% human standard for coding.These benchmarks were introduced as traditional evaluations such as MMLU approach saturation.
  • 7. High-quality AI video generators demonstrate significant improvement: Advanced 2024 video generators, including Meta’s Movie Gen and Google DeepMind’s Veo 2, produced substantially higher-quality videos than 2023 systems.The report also notes a 142-fold reduction in the model size required to exceed 60% on MMLU, from PaLM’s 540 billion parameters to Phi-3 Mini’s 3.8 billion.
  • 9. Complex reasoning remains a problem: On complex reasoning tasks, AI systems score four times higher than human experts with a two-hour budget, but humans outperform AI two to one at 32 hours.AI agents already match human expertise on selected tasks such as writing Triton kernels while producing results faster and at lower cost.

1. Evaluating AI systems with responsible AI criteria is still uncommon, but new benchmarks are beginning

Standardized responsible AI benchmarks for LLMs remain uncommon, although HELM Safety and AIR-Bench are emerging to address the gap. AI-related incident reports also reached a record high in 2024.

  • 1. Evaluating AI systems with responsible AI criteria is still uncommon, but new benchmarks are beginning: Standardized responsible AI benchmarks for LLMs remain lacking, but HELM Safety and AIR-Bench are beginning to fill the gap.The AI Index previously highlighted the lack of standardized RAI benchmarks, and the issue persists.
  • 2. The number of AI incident reports continues to increase: 233 AI-related incidents were reported in 2024, a record high and a 56.4% increase over 2023.The AI Incidents Database recorded the increase.

3. Organizations acknowledge RAI risks, but mitigation efforts lag. A McKinsey survey on organizations’ RAI

Organizations recognize several responsible AI risks, but their concern is not matched by uniformly active mitigation efforts. In a McKinsey survey, inaccuracy, regulatory compliance, and cybersecurity were cited as concerns by 64%, 63%, and 60% of respondents, respectively.

  • Organizations acknowledge RAI risks, but mitigation efforts lag.: 64% of respondents cited inaccuracy as a responsible AI concern, compared with 63% for regulatory compliance and 60% for cybersecurity.These risks were among the issues most salient to organizational leaders, although not all organizations were taking active steps to address identified risks.

6. Foundation model research transparency improves, yet more work remains. The updated Foundation

Foundation model transparency improved substantially, with average scores among major developers rising from 37% to 58%, but considerable room for improvement remains. Newer, more comprehensive factuality and truthfulness evaluations have emerged after earlier benchmarks failed to gain widespread adoption.

  • Foundation model research transparency improves: 58%: Average transparency scores among major foundation model developers increased from 37% in October 2023 to 58% in May 2024.The Model Transparency Index tracks transparency in the foundation model ecosystem.
  • Foundation model research transparency improves: Considerable room for improvement remains despite the promising transparency gains.
  • Better benchmarks for factuality and truthfulness: Newer evaluations, including the updated Hughes Hallucination Evaluation Model leaderboard, FACTS, and SimpleQA, address factuality and truthfulness after HaluEval and TruthfulQA failed to gain widespread adoption.

8. AI-related election misinformation spread globally, but its impact remains unclear. In 2024, numerous · 9. LLMs trained to be explicitly unbiased continue to demonstrate implicit bias. Many advanced LLMs—

AI-generated election misinformation spread across countries in 2024, although its real-world impact remains uncertain. LLMs designed to be unbiased continued to exhibit implicit biases, including racial, gender, and occupational stereotypes that increased with model scale.

  • Racial Classification in Multimodal Models: Larger vision models also amplified racial bias: in ViT-L models, classification probabilities for Black and Latino men as criminals increased by up to 69% as dataset size grew.Larger datasets improved human classification by reducing misidentification of nonhuman entities, but they also increased racial bias.
  • Measuring Implicit Bias in Explicitly Unbiased LLMs: LLMs trained to be explicitly unbiased showed systemic implicit biases, associating negative terms with Black individuals, women with humanities over STEM, and men with leadership.The study introduced two methods for detecting implicit bias and evaluated biases across domains including race, gender, religion, and health.
  • Measuring Implicit Bias in Explicitly Unbiased LLMs: As models scale, implicit biases increase even as decision bias and rejection rates do not, creating an apparent neutrality on standard benchmarks while stereotypes remain pervasive.The findings indicate that lower benchmark bias can coexist with persistent implicit associations.
  • Election Misinformation: Reports disagree about whether AI-driven misinformation had the feared electoral impact, with some finding limited effects while others identify continuing risks.The supplied evidence supports uncertainty about impact rather than a definitive conclusion about electoral consequences.
  • AI Misinformation in the US Elections: More than 35,000 deepfake images depicted 26 members of Congress—25 women—on pornographic sites, illustrating the scale of AI-enabled election-related abuse in the United States.The American Sunlight Project documented the content on pornographic sites.
  • Rest of World 2024 AI-Generated Election Content: AI-generated election content was documented in 60 incidents across 15 countries, demonstrating that election misinformation spread globally in 2024.The reported incidents included manipulated audio, video, and images used to influence voters and political perceptions.
  • Election Misinformation: AI-generated political content was used to discourage voting, cast doubt on election integrity, and intensify polarization across platforms including X, Facebook, Instagram, and Reddit.Examples included Pakistan content encouraging an election boycott, a China-linked “spamouflage” campaign targeting the U.S. election, and an AI-generated South African image circulated across platforms.

7. China’s dominance in industrial robotics continues despite slight moderation. In 2023, China installed

China remained dominant in industrial robotics in 2023, installing 276,300 robots—six times Japan’s total and 7.3 times the United States’. Its global installation share reached 51.1%, although its lead over the rest of the world narrowed slightly.

  • 276,300 industrial robots were installed in China in 2023, six times more than Japan and 7.3 times more than the United States.
  • China’s share of global industrial-robot installations rose from 20.8% in 2013, when it surpassed Japan, to 51.1%.
  • China installed more robots than the rest of the world combined in 2023, but this margin narrowed slightly, indicating modest moderation.

8. Collaborative and interactive robot installations become more common. In 2017, collaborative robots · 9. AI is driving significant shifts in energy sources, attracting interest in nuclear energy. Microsoft announced

Collaborative robots grew from 2.8% of new industrial robot installations in 2017 to 10.5% in 2023, alongside broader growth in service robots. AI is also reshaping energy investment through nuclear agreements and is associated with productivity gains that often narrow skill gaps.

  • 8. Collaborative and interactive robot installations become more common. In 2017, collaborative robots: Collaborative robots’ share of new industrial robot installations rose from 2.8% in 2017 to 10.5% in 2023.
  • 8. Collaborative and interactive robot installations become more common. In 2017, collaborative robots: Service robot installations increased across every application category except medical robotics.
  • 8. Collaborative and interactive robot installations become more common. In 2017, collaborative robots: Together, these trends indicate both overall growth in robot installations and greater emphasis on human-facing roles.
  • 9. AI is driving significant shifts in energy sources, attracting interest in nuclear energy. Microsoft announced: Microsoft announced a $1.6 billion deal to revive the Three Mile Island nuclear reactor to power AI, while Google and Amazon secured nuclear energy agreements for AI operations.
  • 9. AI is driving significant shifts in energy sources, attracting interest in nuclear energy. Microsoft announced: Research reinforced earlier findings that AI boosts productivity.
  • 9. AI is driving significant shifts in energy sources, attracting interest in nuclear energy. Microsoft announced: In most cases, AI helps narrow the productivity gap between low- and high-skilled workers.

Chapter 4: Economy

Corporate AI adoption accelerated in 2024, with broad growth in organizational and generative AI use across regions and functions. Reported benefits include productivity gains, cost savings, revenue increases, and workforce changes, although most companies remain early in scaling value.

  • Industry usage trends: 78% of surveyed organizations used AI in at least one business function in 2024, up from 55% in 2023, while generative AI use reached 71% from 33%.The increase followed stagnation between 2017 and 2023 and occurred across regions.
  • AI deployment: Generative AI most commonly supported marketing strategy content at 27%, followed by knowledge management and personalization at 19% each, but only 1% of executives reported mature rollouts.Most companies remain in the early stages of capturing value at scale.

Chapter 4: Economy … 6. Synthetic data shows significant promise in medicine. Studies released in 2024 suggest that AI-generated

AI deployment spans industrial robotics, scientific protein modeling, and clinical medicine, with 2023 robot installations declining slightly despite record operational stock and China retaining global leadership. In science and medicine, 2024 advances included stronger protein models, rapid FDA authorization of AI-enabled devices, and synthetic-data applications for privacy-preserving prediction and drug discovery.

  • Chapter 4: Economy: Industrial robot installations fell 2.2% to 541,000 units in 2023, the first year-over-year decrease since 2019, while operational stock rose to 4,282,000.Since 2012, both industrial robot installation and utilization have steadily increased.
  • Industrial Robots: Traditional vs. Collaborative Robots: Collaborative robots grew from 2.8% of new industrial robot installations in 2017 to 10.5% in 2023.The report distinguishes collaborative robots, which work alongside humans, from traditional robots that operate in place of humans.
  • By Geographic Area: China led 2023 industrial robot installations with 276,300 units, six times Japan’s 46,100 and 7.3 times the United States’ 37,600.China’s share of global installations increased from 20.8% in 2013 to 51.1% in 2023.
  • 1. Bigger and better protein sequencing models emerge: In 2024, larger protein sequencing models such as ESM3 and AlphaFold 3 continued improving protein prediction accuracy and scientific discovery.The chapter also highlights Aviary for biological tasks and FireSat for wildfire prediction.
  • 1. Bigger and better protein sequencing models emerge: OpenAI’s o1 achieved 96.0% on MedQA, a 5.8-percentage-point gain over the best 2023 score and a 28.4-point improvement since late 2022.The report says MedQA may be approaching saturation, indicating a need for more challenging evaluations.
  • 5. The number of FDA-approved, AI-enabled medical devices skyrockets: FDA-authorized AI-enabled medical devices increased from six in 2015 to 223 in 2023.The FDA authorized its first AI-enabled medical device in 1995.
  • 6. Synthetic data shows significant promise in medicine: AI-generated synthetic data can help identify social determinants of health, improve privacy-preserving clinical risk prediction, and support discovery of new drug compounds.Synthetic datasets can preserve statistical fidelity, support exploratory analysis, and develop predictive models without real and identifiable patient data.

2. Governments across the world invest in AI infrastructure. Canada announced a $2.4 billion AI infrastructure

Governments worldwide are making substantial AI infrastructure investments, including major commitments from Canada, China, France, India, and Saudi Arabia. AI mentions in legislative proceedings across 75 major countries also rose sharply in 2024 and have increased more than ninefold since 2016.

  • Government AI infrastructure investment: China launched a $47.5 billion semiconductor-production fund, while Canada announced a $2.4 billion AI infrastructure package.
  • Government AI infrastructure investment: Saudi Arabia’s Project Transcendence represents a $100 billion AI investment, while France committed €109 billion to AI infrastructure.
  • Government AI infrastructure investment: India pledged $1.25 billion to AI, adding to a broad wave of national investment in AI infrastructure.
  • AI in legislative proceedings: AI mentions in legislative proceedings across 75 major countries increased 21.3% in 2024, reaching 1,889 from 1,557 in 2023.
  • AI in legislative proceedings: Since 2016, the total number of AI mentions in legislative proceedings has grown more than ninefold.

4. AI safety institutes expand and coordinate across the globe. In 2024, countries worldwide launched international · 6. U.S. states expand deepfake regulations. Before 2024, only five states—California, Michigan, Washington, Texas,

In 2024, AI safety institutes expanded from the United States and United Kingdom to countries across Asia, Europe, and beyond, while governments worldwide increased AI-related regulation, investment, and coordination. U.S. states also sharply broadened laws targeting deepfakes, with 24 states regulating them by 2024.

  • 4. AI safety institutes expand and coordinate across the globe. In 2024, countries worldwide launched international: AI safety institutes first emerged in the U.S. and U.K. in November 2023, followed by pledged institutes across Japan, France, Germany, Italy, Singapore, South Korea, Australia, Canada, and the European Union.The expansion followed the inaugural AI Safety Summit and the AI Seoul Summit in May 2024.
  • 6. U.S. states expand deepfake regulations. Before 2024, only five states—California, Michigan, Washington, Texas,: Before 2024, five U.S. states had enacted election deepfake laws; in 2024, 15 more states introduced similar measures, bringing the total regulating deepfakes to 24.The newly regulating states included Oregon, New Mexico, and New York.
  • 6.1 Major Global AI Policy News in 2024: Governments announced major AI investments and policy initiatives, including Singapore’s over $1 billion five-year commitment, Abu Dhabi’s $100 billion AI investment firm, and the EU AI Act.The EU AI Act introduced transparency and reporting obligations, risk-based regulation, and bans on certain applications.
  • 6.1 Major Global AI Policy News in 2024: Countries at the AI Seoul Summit signed a letter of intent to establish a collaborative network of institutes, emphasizing global cooperation in advancing AI safety.The summit built on safety measures shared in line with the Bletchley Declaration.
  • Overview: AI-related laws passed worldwide rose from 30 in 2023 to 40 in 2024, with 39 countries having enacted at least one AI-related law and 204 laws passed in total.The 2024 total was the second-highest annual count after 2022, rising from one law in 2016.
  • 4. AI safety institutes expand and coordinate across the globe. In 2024, countries worldwide launched international: 59 U.S. AI-related regulations were introduced in 2024, more than double 2023’s 25, and they came from a record 42 agencies, up from 21.The agency count reflects growing AI-related activity across a wider range of government departments.
  • 6.3 Public Investment in AI: Public AI investment expanded substantially: the U.S. awarded approximately $830 million in AI-related tenders versus $4.5 billion in grants in 2023, while Europe’s 2023 investment was approximately 67 times its 2013 level.From 2013 to 2023, the United States distributed about $5.2 billion across 2,678 AI contracts.

5. The U.S. continues to be a global leader in producing information, technology, and communications

The U.S. remains a global leader in ICT education, producing more graduates than any sampled country across all degree levels and nearly twice as many or more than the next-highest country. However, access and participation remain unequal across K–12 and postsecondary education, while AI education and use are expanding rapidly.

  • K–12 education: K–12 access to computer science varies widely, with 35% of U.S. high schools offering CS and student participation ranging from 26% in South Carolina to 2% in Florida and Arizona.AP CS participation has grown, but Asian students, white boys, and multiracial students are overrepresented while all other groups are underrepresented.
  • K–12 education: AI preparation is advancing unevenly: 26 U.S. states had issued AI guidance by January 2025, but only 46% of high school and 34% of elementary CS teachers felt equipped to teach AI.Although 81% of CS teachers supported including AI in foundational CS education, elementary teachers reported needing more AI resources at 88%.
  • Postsecondary education: AI education is expanding quickly, with CS master’s graduates increasing 26% from 2022 to 2023 and 83% over the past decade.The number of U.S. institutions offering AI-specific bachelor’s degrees nearly doubled between 2022 and 2023, while AI-specific master’s programs also increased sharply.
  • Postsecondary education: The U.S. produces more ICT graduates than any sampled country at the associate, bachelor’s, master’s, and PhD levels.It graduates more than twice as many associate, master’s, and PhD students—and nearly twice as many bachelor’s students—as the next-highest country.
  • Postsecondary education: Participation remains unequal: women comprise approximately one-quarter of ICT graduates at most degree levels, while Black and Hispanic students remain underrepresented in computing degrees.In 2023, nonresidents accounted for 67% of U.S. master’s graduates and 60% of PhD graduates in computing.

4. Regional differences persist regarding AI optimism. First reported in the 2023 AI Index, significant regional · 5. People in the United States remain distrustful of self-driving cars. A recent American Automobile Association · 7. AI optimism registers sharp increase among countries that previously showed the most skepticism.

Public opinion on AI remains sharply divided by region: Asian countries are more optimistic than Europe and the Anglosphere, while U.S. distrust extends to self-driving cars. Optimism has nevertheless risen most in countries that were previously the most skeptical.

  • 5. People in the United States remain distrustful of self-driving cars. A recent American Automobile Association: 61% of people in the U.S. fear self-driving cars, while only 13% trust them; fear declined from 68% in 2023 but remains above 54% in 2021.These figures come from a recent American Automobile Association survey.
  • 7. AI optimism registers sharp increase among countries that previously showed the most skepticism.: Optimism increased most in countries that were previously skeptical: Great Britain rose 8%, Germany 10%, the United States 4%, Canada 8%, and France 10%.In 2022, these countries had among the lowest shares viewing AI as more beneficial than harmful: 38%, 37%, 35%, 32%, and 31%, respectively.
  • 7. AI optimism registers sharp increase among countries that previously showed the most skepticism.: 60% of respondents expect AI to change how they do their job within five years, while 36% believe AI will replace their job.The survey reports broad expectations of workplace change but a smaller share expecting replacement.
  • 7. AI optimism registers sharp increase among countries that previously showed the most skepticism.: 80.4% of local U.S. policymakers support stricter data privacy rules, compared with 76.2% for retraining unemployed workers and 72.5% for AI deployment regulations.Support is much lower for a law enforcement facial recognition ban at 34.2%, wage subsidies at 32.9%, and universal basic income at 24.6%.
  • 7. AI optimism registers sharp increase among countries that previously showed the most skepticism.: 55% expect AI to save time and 51% expect better entertainment, but only 38% foresee improved health and 36% a stronger national economy.Only 31% expect a positive impact on the job market, and 37% expect AI to enhance their own jobs.
  • 5. People in the United States remain distrustful of self-driving cars. A recent American Automobile Association: 73.7% of local U.S. policymakers supported AI regulation in 2023, up from 55.7% in 2022, with stronger support among Democrats than Republicans.Support was 79.2% among Democrats and 55.5% among Republicans.

APPENDIX … Measuring Demand for AI

The appendix documents methods for measuring AI research, patents, notable models, projects, environmental impact, responsible AI, robotics, and labor-market demand. Across these analyses, the AI Index combines large-scale databases, classification systems, surveys, and administrative or commercial data, with explicit handling of attribution, coverage, and estimation limitations.

  • AI Publication Analysis: AI publication analysis combines OpenAlex with the Computer Science Ontology and CSO Classifier to identify and classify AI research topics.OpenAlex provides the primary publication data, while CSO refinement and an unsupervised three-stage classifier support more precise AI-topic identification.
  • AI Publication Analysis: AI publication counts are grouped by year, geography, sector, and citations, with cross-country or cross-sector coauthorship potentially contributing multiple counts.Conference-versus-journal classification is reconciled by mapping OpenAlex records to DBLP using DOIs, then titles and publication years for unmatched papers.
  • Top 100 Publications Analysis: The top-100 publication analysis starts with the 150 most-cited papers per year, refines the list through review, and attributes papers to all countries and regions represented by affiliations.Publication years use the most recent versions, and organizational affiliations are standardized with countries assigned by headquarters location.
  • AI Patent Analysis: AI patents are identified through a hybrid of keyword and classification-code methods using PATSTAT Global, focusing on granted patents from 2010 onward at the DOCDB-family level.Keyword dictionaries are expanded and validated with AI models, Word2Vec, BERTopic, DeBERTA, and manual checks, then merged with IPC- and CPC-code results; marginal duplicates can remain despite aggregation.
  • Epoch Notable Models Analysis: Epoch’s notable-model dataset tracks country contributions to landmark AI research and estimates training costs for the largest-scale ML models from hardware hours and historical cloud rental rates.Cost estimates are limited because developers may not disclose training duration or hardware, some hardware prices are unavailable, and purchased hardware or vendor commitments may differ from rental assumptions.
  • Identifying AI Projects: Public AI projects are identified from GitHub using AI/ML and generative-AI topic labels, snowball-sampled keywords, and dependencies on major AI libraries, then mapped geographically from owners’ IP-address locations.Annual project locations use the modal owner location sampled daily, carrying the last known location forward on inactive days.
  • Environmental Impact Analysis: Environmental-impact analysis estimates training-stage emissions for 50 influential language and vision models using hardware, GPU hours, provider, and region inputs.The estimates exclude embodied hardware production, idle infrastructure, and deployment emissions; newer hardware may use substitutes, while custom infrastructure incorporates carbon efficiency and offsets.

LinkedIn Data … Quid insights prepared by Heather English and Hansen Yang

The section defines LinkedIn-based AI skills, talent, hiring, migration, and career-transition measures, then reports selected global findings while documenting coverage, privacy, and methodological limits. It also describes Quid’s AI/NLP-based analysis of company and other public datasets to identify topics, trends, and hidden patterns.

  • LinkedIn Data; Country Sample; METHODOLOGIES: LinkedIn’s anonymized, aggregated data cover more than 1 billion members, but platform use varies by professional, social, and regional culture; strict quality thresholds prevent disclosure about individuals.The data include information from the corresponding period and are published to support accurate statistics while protecting member privacy.
  • Gender; GLOBAL COMPARISON: BY GENDER; GLOBAL COMPARISON: ACROSS GENDERS; 6. Female Representation in AI: Gender comparisons use inferred or self-identified binary categories, exclude members whose gender cannot be inferred and countries with insufficient coverage, and report female AI Engineering representation of 30.5% globally.Gender-specific penetration rankings should be compared within gender, while the across-gender measure uses the same global average.
  • 1. Top AI Skills; 2. Fastest Growing AI Skills; AI Engineering: The most frequently added AI Engineering skills globally since 2015 are machine learning, AI, and deep learning, while the fastest-growing are custom GPTs, AI productivity, and AI agents.LinkedIn applies thresholds to recent skill-add volumes based on the 50th percentile of each country’s recent distribution.
  • 3. AI Talent Concentration; 4. Relative AI Talent Hiring Rate YoY Ratio: AI Engineering talent represents 0.78% of LinkedIn members in the United States, and AI talent hiring relative to overall hiring there grew 24.7% year over year.Talent concentration divides AI talent by LinkedIn membership, while the hiring metric compares year-over-year AI hiring changes with overall hiring in the same country.
  • 7. AI Talent Migration; 8. Career Transitions Into AI Jobs: LinkedIn defines migration as profile-location changes and normalizes net arrivals minus departures by country membership; the United States recorded 1.07 net AI-talent flow per membership-normalized unit, and 26.9% of transitions into AI engineer came from software engineer.Career-transition estimates pool five years of member-level moves, exclude first occupations and intra-occupation transitions, and migration uses minimum sample-size thresholds.
  • Quid insights prepared by Heather English and Hansen Yang: Quid combines an in-house LLM, smart search, Boolean queries, and AI/NLP across more than 8 million company profiles and other datasets, linking documents by similar language into clusters that reveal topics and trends.Its outputs include network graphs, dashboards, and PostgreSQL delivery; company information is updated weekly, with metadata spanning investment and firmographic attributes.

Data … Ipsos

The AI Index 2025 compiles global company, investment, policy, education, medical, and public-opinion data from diverse sources using structured searches, filtering, processing, and validation procedures. Coverage is broad but constrained by undisclosed company data, uneven procurement and government-spending records, methodological changes, and omitted legal or survey details.

  • Companies: Company and investment data cover global organizations and funding events, combining Capital IQ with Crunchbase to capture early-stage startups, while undisclosed investors or amounts create gaps.The dataset includes private and public organizations, subsidiaries, and out-of-business entities, plus private investments, M&A, public offerings, minority stakes, and venture funding.
  • Target Event Definitions: The report defines private placements, minority investments, and M&A by ownership and transaction structure, distinguishing minority stakes below 50% from acquisitions exceeding 50%.Private placements may involve equity or debt and can produce either minority or majority stakes.
  • FDA-Approved AI Medical Devices: Medical-AI analyses use FDA device records and a PubMedCentral search of 2020–2024 English-language articles intersecting AI, medicine, and ethics, leaving 2,916 eligible articles after exclusions.Articles were filtered by abstract keywords and exclusions including preprints and retracted articles; ethical-issue frequencies in abstracts were then analyzed.
  • State-Level Data: Education and public-opinion appendices draw on Code.org, CSTA, ECEP Alliance, College Board, Cambridge, IPEDS, OECD, and Ipsos sources, but education measures reflect coding choices and the Ipsos methodology is not republished.AI-related completions use multiple CIP codes; Colorado and Virginia standards adopted in late 2024 are excluded; the Ipsos survey directs readers to its original survey for methodological details.
Loading 2504.07139v3…