PaperlayerView exact revision on arXiv

Source-linked AI summary

The impact of generative artificial intelligence on socioeconomic inequalities and policy making

Valerio Capraro, Austin Lentsch, Daron Acemoglu, Selin Akgun, Aisel Akhmedova, Ennio Bilancini, Jean-François Bonnefon, Pablo Brañas-Garza, Luigi Butera, Karen M. Douglas, Jim A. C. Everett, Gerd Gigerenzer, Christine Greenhow, Daniel A. Hashimoto, Julianne Holt-Lunstad, Jolanda Jetten, Simon Johnson, Chiara Longoni, Pete Lunn, Simone Natale, Iyad Rahwan, Neil Selwyn, Vivek Singh, Siddharth Suri, Jennifer Sutcliffe, Joe Tomlinson, Sander van der Linden, Paul A. M. Van Lange, Friederike Wall, Jay J. Van Bavel, Riccardo Viale

arXiv:2401.05377v2cs.CY

TL;DR

Generative AI may reshape socioeconomic inequalities across information, work, education, and healthcare, while creating benefits and risks that current policies do not fully address. This paper synthesizes interdisciplinary research and concludes that dynamic regulation and further research are needed to harness benefits while mitigating harms.

  • Problem

    Research has insufficiently examined how generative AI affects information plurality, inequalities, and policy challenges across information-intensive social domains.

    Source: Impact on (mis)information · p. 7, Regulation using AI · p. 34

  • Method

    The paper provides an interdisciplinary overview, evaluates existing research, identifies gaps, and develops research directions on generative AI’s socioeconomic impacts.

    Source: Impact on work · p. 14

  • Results

    Generative AI offers tailored and democratic information processing and potential benefits across work, education, and healthcare, but also raises risks including misinformation, inequality, and unequal access.

    Source: Conclusion · p. 35

  • Takeaways & Limitations

    Current regulatory approaches in the EU, US, and UK may inadequately address these challenges, supporting the need for dynamic regulation that keeps pace with AI advances.

    Source: Conclusion · p. 35

    Existing interventions may be less effective against subtle or personalized AI-generated misinformation, requiring adapted or new intervention strategies.

    Source: Impact on (mis)information · p. 6

Abstract

from arXiv · show

Generative artificial intelligence has the potential to both exacerbate and ameliorate existing socioeconomic inequalities. In this article, we provide a state-of-the-art interdisciplinary overview of the potential impacts of generative AI on (mis)information and three information-intensive domains: work, education, and healthcare. Our goal is to highlight how generative AI could worsen existing inequalities while illuminating how AI may help mitigate pervasive social problems. In the information domain, generative AI can democratize content creation and access, but may dramatically expand the production and proliferation of misinformation. In the workplace, it can boost productivity and create new jobs, but the benefits will likely be distributed unevenly. In education, it offers personalized learning, but may widen the digital divide. In healthcare, it might improve diagnostics and accessibility, but could deepen pre-existing inequalities. In each section we cover a specific topic, evaluate existing research, identify critical gaps, and recommend research directions, including explicit trade-offs that complicate the derivation of a priori hypotheses. We conclude with a section highlighting the role of policymaking to maximize generative AI's potential to reduce inequalities while mitigating its harmful effects. We discuss strengths and weaknesses of existing policy frameworks in the European Union, the United States, and the United Kingdom, observing that each fails to fully confront the socioeconomic challenges we have identified. We propose several concrete policies that could promote shared prosperity through the advancement of generative AI. This article emphasizes the need for interdisciplinary collaborations to understand and address the complex challenges of generative AI.

Introduction

Generative AI is a qualitative departure from traditional automation whose effects may span society, producing both benefits and harms that are unevenly distributed. This article examines those trade-offs across information, work, education, healthcare, and policymaking.

Motivation

  • The article evaluates positive and negative societal effects to help scholars and policymakers harness generative AI’s benefits while mitigating costs to individuals and society.This follows Kranzberg’s principle that technology is neither inherently good nor bad nor neutral.Source: Introduction · p. 3

Information

  • Generative AI can democratize information and content creation while increasing misinformation and eroding trust in digital content.Its ability to create original output distinguishes it from traditional automation and may transform how knowledge and skills are valued.Source: Introduction · p. 3

Work, education, and healthcare

  • In work, education, and healthcare, generative AI may improve productivity, learning, patient outcomes, and accessibility, but benefits and access may remain uneven.Risks include unequal effects across firms, sectors, and worker demographics; unequal access to educational tools; and deeper inequalities for under-resourced and marginalized healthcare communities.Source: Introduction · pp. 3–4

Policymaking

  • Existing policy approaches in the European Union, United States, and United Kingdom fall short of addressing generative AI’s socioeconomic risks.The article recommends studying policies to combat AI-generated misinformation, prevent job-market inequalities, and bridge digital divides in education and healthcare.Source: Introduction · p. 4

Impact on (mis)information

Generative AI can expand access to and personalize information, but it also enables misinformation, surveillance, market concentration, and degraded information quality. These risks can alter political behavior, human-machine interactions, and knowledge production, requiring targeted interventions, organizational responsibility, and government regulation.

Access and market power

  • Generative AI broadens information access through personalization, translation, accessibility formats, and automated fact-checking, while excessive data collection enables privacy violations, bias exploitation, surveillance capitalism, and anticompetitive market power.Companies may optimize advertising and user engagement in ways that harm individual and societal well-being.Source: Impact on (mis)information · pp. 4–5

Misinformation production

  • AI-generated misinformation can convincingly imitate human text, audio, images, and video, with manipulated political images comprising approximately 20% of visual misinformation and deepfakes increasingly relevant during elections and conflicts.People often cannot distinguish AI-generated from human-generated text, and AI can produce more convincing misinformation and propaganda than humans.Source: Impact on (mis)information · p. 5

Misinformation effects

  • Personalized AI misinformation can spread rapidly, micro-target audiences, influence political attitudes and voting, deepen polarization and distrust, affect health decisions, and fuel intergroup conflict.AI can generate thousands of persuasive targeted messages at scale, making misinformation especially consequential for susceptible groups.Source: Impact on (mis)information · pp. 5–6

Interventions and regulation

  • Accuracy prompts and prebunking can improve misinformation discernment, but their effects are often modest and may be weaker against subtle AI-generated content, personalized communications, and visual misinformation.Prebunking exposes people to weakened misinformation techniques and may work better than debunking, while AI could itself help counter conspiracy beliefs through dialogue.Source: Impact on (mis)information · p. 6

Human-machine interactions

  • Even without malicious actors, AI hallucinations can create realistic falsehoods and self-fulfilling prophecies, while uncertainty about whether users interact with machines can reduce emotional responsiveness, reciprocity, prosociality, and honesty.People may feel less heard when they know a response comes from a machine and may behave more selfishly or dishonestly in human-machine interactions.Source: Impact on (mis)information · pp. 6–7

Information quality and plurality

  • Generative AI search and mass-produced bot or search-optimized content may restrict source plurality, flood the web with mediocre “average” material, reduce novice participation in knowledge-sharing platforms, and undermine innovation and creativity.AI-mediated search may provide fewer sources, while reduced basic-question participation can weaken the continued usefulness of Q&A sites.Source: Impact on (mis)information · pp. 7–8

Future research directions

Future research should test how generative AI can improve information accessibility, reduce harms, and support ethical, pluralistic information systems. The proposed studies emphasize experimental evaluation of competing benefits and risks, including accessibility versus accuracy, transparency versus trust, and availability versus content diversity.

Information accessibility

  • Research should test whether AI summarization improves comprehension for people with cognitive disabilities while avoiding omissions that reduce content accuracy.A proposed comparison evaluates complex-news comprehension with and without AI summarization.Source: Future research directions · p. 9

Competition and monopolization

  • Studies should examine whether open-source AI enables smaller firms to compete with dominant firms, while comparing growth and success under open-source and proprietary models.Open-source AI may democratize development, but smaller firms may lack the capacity to benefit equally.Source: Future research directions · p. 9

Data governance

  • Future work should test whether standardized transparency protocols reduce data misuse and privacy breaches, while measuring effects on user trust and data sharing.More demanding protocols could reduce trust and data sharing even if they improve privacy outcomes.Source: Future research directions · p. 9

Misinformation

  • Experiments should compare misinformation beliefs after dialoguing with a misinformation-countering AI versus a neutral AI, accounting for hallucinations that may generate new misinformation.AI personalization, fluency, and reasoning could diminish misinformation beliefs, but hallucinations remain a risk.Source: Future research directions · pp. 9–10

Human-machine ethics

  • Research should test whether delegating morally ambiguous decisions to AI changes ethical standards, balancing AI’s emotional detachment against its capacity to document decisions.Participants could choose delegation to AI or another human while researchers observe changes in ethical standards and accountability.Source: Future research directions · p. 10

Information plurality and digital responsibility

  • Future studies should assess whether transparent search sources increase engagement with alternative information, while testing how AI-content ratios affect forum diversity and quality.Source transparency may encourage pluralism but cause information overload; higher AI-content ratios may increase availability while homogenizing or banalizing content.Source: Future research directions · p. 10

Impact on work

Generative AI can reduce workplace inequality by disproportionately improving less-skilled workers’ productivity, but unequal access, firm practices, and infrastructure constraints may instead widen disparities. Its effects depend on whether organizations use AI to augment human capabilities and expand opportunity or to replace workers and commodify expertise.

Productivity and inequality

  • 34% more issues per hour were resolved by less-skilled or inexperienced customer-service workers using a generative AI assistant, versus 14% across all workers (Brynjolfsson and et al., 2023).Workers with two months’ tenure performed as well as unaided workers with more than six months’ tenure.Source: Impact on work · p. 12
  • Generative AI shows an inverse skill-bias, often benefiting less-skilled workers more, but performance can decline on tasks requiring imperfect information, omitted data, or experience-based intuition.Below-average consultants improved by 43% with AI, compared with 17% for above-average consultants.Source: Impact on work · p. 13
  • 71 minutes versus 161 minutes enabled programmers with GitHub Copilot to complete a task, with the largest gains among less-experienced and older programmers (Peng et al., 2023).ChatGPT similarly accelerated writing and improved quality, job satisfaction, and self-efficacy, benefiting lower-ability workers more.Source: Impact on work · p. 13

Access and opportunity

  • Translation and remote-work technologies could broaden digital-economy access, but near-term benefits for the Global South are constrained by insufficient infrastructure, researchers, and digital-skills training.Generative search may also lower research costs and help low-resource users access information traditionally available only in high-resource environments.Source: Impact on work · pp. 13–15

Workplace inequalities

  • Unequal access to internet-enabled devices, technical inputs, and training may leave less-resourced workers, firms, and industries further behind.Even when AI raises less-skilled workers’ capabilities, firms may exploit worker interchangeability rather than share resulting wage gains.Source: Impact on work · p. 14

Future research

  • Future research should examine human-complementary AI design, automation versus hiring, marginalized-group access to remote work, and long-term effects on job satisfaction and burnout.The central trade-off is between automation’s efficiency and human knowledge workers’ contributions to innovation, productivity, and worker well-being.Source: Impact on work · p. 15

Impact on education

Generative AI could expand personalized, scalable educational support, but it may also reproduce biases, widen participation gaps, and undermine independent learning. Realizing its benefits requires redesigned curricula that teach critical thinking, fact-checking, prompt engineering, and AI literacy alongside transparent, equitable implementation.

Benefits and opportunities

  • Chatbot tutors could provide skill-adaptive, personalized teaching and on-demand support at scale, potentially bridging persistent educational gaps, especially in large classes.Unlike earlier online courses, generative AI can interact dynamically with students and reduce educators’ need to continually repackage content.Source: Impact on education · p. 17
  • Personalized learning systems have produced higher test scores, while assessment automation can accelerate grading and deliver timely feedback; students also report benefits for writing, brainstorming, research, and analysis.Students simultaneously raised concerns about accuracy, privacy, ethics, personal development, and career prospects (Chan & Hu, 2023).Source: Impact on education · p. 17

Equity and bias

  • Generative AI may reinforce racial and gender stereotypes in personalized learning, scoring, admissions, and translation, amplifying inequalities affecting historically marginalized groups (Akgun and Greenhow, 2022; Baker and Hawn, 2022; Johnson, 2021).Bias can enter through training data derived from historically biased human decisions.Source: Impact on education · p. 17
  • AI could be engineered to reduce and audit educational bias, but transparent systems and stakeholder monitoring are necessary because marginally less discriminatory technology cannot replace eliminating discrimination.Relevant stakeholders include students, teachers, and parents.Source: Impact on education · p. 18
  • Unequal engagement may widen group-based inequalities: female students reported using ChatGPT less frequently than male students, potentially affecting academic achievement and future workforce gender gaps.The passage therefore calls for fair distribution of generative AI’s benefits across student segments.Source: Impact on education · p. 18

Implementation risks

  • Teachers may gain or lose workload as they curate content, monitor student-AI interactions, and provide technical support, while overreliance, inaccurate outputs, and weak evaluation skills could undermine independent student thinking.Technology proficiency may also create a generational divide among educators.Source: Impact on education · p. 18

Curricular transformation

  • Curricula should teach idea organization, writing, critical thinking, prompt engineering, fact-checking, AI operation, and appropriate interaction rather than treating generative AI as mere assistance.Generative AI’s variable text output makes evaluating arguments, structures, and prompts fundamental; students should also understand that language models calculate word correlations rather than think like humans.Source: Impact on education · pp. 18–19
  • 96 percent of more than 3,000 U.S. high school students and undergraduates did not know how to evaluate website trustworthiness, underscoring the need to teach lateral reading and comparable heuristics for AI outputs.Without curricular adaptation, mechanical AI use could reduce higher-order cognitive skills, particularly among students using these tools less analytically.Source: Impact on education · p. 19

Impact on healthcare

Generative AI could make healthcare more accessible, affordable, accurate, and personalized while reducing clinicians’ administrative burdens, but benefits depend on effective human–AI collaboration and equitable data. Its adoption also risks reproducing bias, limiting insurance access, weakening trust, and replacing rather than supplementing human social interaction, requiring targeted research and safeguards.

Clinical decision-making

  • GPT-4 advice improved physicians’ diagnostic accuracy and reduced gender and race bias in randomized chest-pain triage, risk, and treatment scenarios.Clinicians were willing to heed the chatbot’s advice.Source: Impact on healthcare · p. 22
  • GPT-4 shows reasonable accuracy on simple medical visual questions and standardized exams but struggles with diagnostically complex prompts.Generative AI can analyze medical images, including X-rays and MRI scans, while answering image-specific questions.Source: Impact on healthcare · p. 22
  • Human–AI collaboration can outperform either alone in multimodal clinical decisions, but AI may also worsen some expert physicians’ diagnoses, making context and clinician characteristics critical.Endoscopists benefited particularly when AI advice was correct, while another study found AI could cause errors in cases physicians would otherwise assess correctly.Source: Impact on healthcare · pp. 22, 26

Patient access and engagement

  • ChatGPT answers to Reddit medical questions were rated higher quality and more empathetic than physicians’ answers 79% of the time.People trusted ChatGPT for low-risk questions, but trust varied with medical complexity (Nov et al., 2023).Source: Impact on healthcare · p. 23

Equity and bias

  • AI-mediated healthcare may improve engagement among stigmatized groups by reducing the salience of identity-based treatment concerns, but biased training data and clinical criteria can reproduce unequal outcomes.A resource-allocation algorithm favored White over Black patients at equal health risk until its cost-based criterion was changed to focus on health; another model was nearly five times as effective as the Kellgren-Lawrence system for predicting knee pain.Source: Impact on healthcare · p. 23

Insurance

  • AI-driven insurance may either expand coverage and reduce claim denials or increase premiums and remove access for high-risk people if insurers exploit individualized risk estimates.Responsible AI could predict and prevent denials, while more accurate loss probabilities could reduce asymmetric information (Johnson et al., 2021; Eling et al., 2022).Source: Impact on healthcare · p. 24

Social connection and health

  • Generative AI should supplement, not replace, human relationships because overreliance on conversational agents may increase isolation and harm mental and physical health.Chatbots may reduce loneliness for some people, but insufficient human interaction may fail to meet deep social needs and inhibit motivation for human contact.Source: Impact on healthcare · pp. 25, 27

Policymaking in the age of artificial intelligence · Regulation of AI

The EU, US, and UK have developed distinct AI governance approaches, but these frameworks insufficiently address socioeconomic inequalities. The paper proposes interventions spanning taxation, labor voice, human-complementary innovation, training, misinformation, and governmental expertise.

Regulation of AI

  • The EU AI Act uses risk-based categories, transparency, accountability, and rights protections, but broad definitions and strict rules may create uncertainty and reduce global competitiveness.It distinguishes banned, high-risk, and low-risk applications while emphasizing nondiscrimination and privacy.Source: Regulation of AI · p. 29
  • The US relies on flexible but fragmented federal and state initiatives, creating regulatory inconsistency and ethical gaps, while a 2023 Executive Order seeks more cohesive federal policy.The fragmented approach may hinder unified standards for privacy, bias, accountability, and global AI governance.Source: Regulation of AI · p. 29
  • The UK favors pro-innovation, non-statutory principles and existing regulators, but critics argue this overlooks public-sector AI and requires stronger transparency, accountability, consultation, specialist oversight, and redress.The alternative white paper also calls for prohibiting uses that threaten fundamental rights.Source: Regulation of AI · p. 30
  • Existing EU, US, and UK regulations do not sufficiently address socioeconomic inequalities potentially caused by generative AI.The paper identifies missing interventions and summarizes them in Table 5.Source: Regulation of AI · p. 30, Table 5 · p. 32
  • Tax reform should equalize marginal burdens on hiring and training labor with investment in equipment and software, reducing incentives for automation over human capital.The proposed symmetry aims to encourage human-complementary technological choices.Source: Regulation of AI · p. 30
  • Workers and civil society should influence AI-related changes, while updated safety rules and data unions could strengthen consumer control and compensation for information use.Data unions could also support competition against dominant AI companies such as Microsoft and Google.Source: Regulation of AI · p. 30
  • Governments should fund human-complementary AI research in education, healthcare, and craft training, where augmenting workers’ skills could raise productivity and wages.The proposal builds on targeted public investment and competition to pair AI tools with human expertise.Source: Regulation of AI · p. 31
  • Additional priorities include professional training, AI-misinformation detection and education, and expert consultative bodies to support effective human-complementary integration across government.Training should cover AI capabilities, limitations, ethics, and technical interaction; task forces and advisory initiatives can support implementation.Source: Regulation of AI · p. 31

Regulation using AI

Generative AI could strengthen policymaking through data-driven analysis, forecasting, and policy simulation, but its use is constrained by behavioral biases, conflicting human values, implementation difficulties, and unresolved philosophical questions. Reliable deployment therefore requires continuous organizational oversight of alignment with legal and regulatory requirements.

Potential benefits and concerns

  • Generative AI can analyze vast data, detect complex patterns, forecast trends, and simulate policy outcomes, although ethical and practical concerns may currently be prohibitive.These capabilities could provide policymakers with insights and evidence that elude unaided human analysis.Source: Regulation using AI · p. 32

Human-behavior simulation

  • GPT-4 can simulate human decision behavior and explain how people balance self-interest with others’ interests, but it underestimates self-interest and inequity-aversion while overestimating altruism, creating an “optimistic bias” (Chen et al., 2023; Capraro et al., 2023; Capraro et al., 2024).This bias could produce disillusionment, frustration, and social tensions when AI-assisted decisions assume excessive human prosociality.Source: Regulation using AI · p. 33

Value alignment

  • Aligning policy recommendations with beneficence, non-maleficence, justice, autonomy, and culturally diverse values is difficult because these principles can conflict and humans lack consensus on resolving moral trade-offs.Generative AI might maximize aggregate welfare while infringing autonomy or harming a minority, and even diverse preference data cannot establish a shared normative standard (Mill, 1861; Acemoglu and Johnson, 2023).Source: Regulation using AI · pp. 33–34

Alignment mechanisms

  • Convenience-sample human feedback can encode biased views, while representative feedback, consensus-seeking chatbots, or value-diverse chatbot ecosystems still cannot fully ensure cultural sensitivity.The alignment problem is worst when systems reflect a small homogeneous group of informants.Source: Regulation using AI · pp. 33–34

Implementation and oversight

  • Because policymakers often cannot evaluate embedded preferences or systematic biases, they become principals dependent on other agents for scientific assessments of AI behavior and misbehavior (Rahwan et al., 2018).Supporting organizations should frequently evaluate alignment with legal and regulatory requirements and signal changes to AI-using companies through a dedicated monitoring office (Rahwan, 2018).Source: Regulation using AI · p. 34

Philosophical problem

  • Even if preference conflicts, culturally sensitive sampling, and implementation were solved, the philosophical question of whether policymakers should delegate such decisions to generative AI would remain unresolved.This problem extends beyond technical reliability to the justification for using AI in policy decision-making.Source: Regulation using AI · p. 34

Conclusion

Generative AI may reshape socioeconomic inequalities across information, work, education, and healthcare, with outcomes depending on how society designs and deploys the technology. The conclusion calls for research, inclusive debate, and dynamic regulation to capture benefits while mitigating risks.

Conclusion

  • Research and inclusive public discussion are urgently needed because decisions made across governance, science, industry, and society will shape AI’s lasting effects on future generations.Source: Conclusion · p. 35

Information

  • Generative AI could offer more tailored, efficient, and democratic information processing while creating anticompetitive advantages, enabling data misuse, and accelerating misinformation.Source: Conclusion · p. 35

Work, education, and healthcare

  • Across work, education, and healthcare, generative AI may expand opportunity while altering wage distributions, widening the digital divide, and limiting equitable access to AI-driven services.It could automate tasks and create new work, personalize and democratize learning, and improve patient outcomes through large-scale data analysis, while requiring new skills and preserving genuine human interactions.Source: Conclusion · p. 35

Policy and research agenda

  • Current regulatory approaches in the EU, US, and UK may inadequately address these challenges, requiring dynamic frameworks that keep pace with rapid AI advances.Source: Conclusion · p. 35

Supplementary Information

This supplementary section outlines future research directions for examining how generative AI may reshape information, work, education, and healthcare. It emphasizes experimental and longitudinal approaches to assess potential benefits and risks, including inequality, access, privacy, bias, oversight, trust, creativity, and human autonomy.

Accessibility

  • Future research should investigate AI tools that improve information accessibility for people with cognitive, visual, and physical disabilities while balancing assistance against omitted details, distraction, confusion, or overreliance.Proposed studies examine summarization, audio descriptions, and mobility assistants, measuring comprehension, enjoyment, independence, and navigation.Source: Supplementary Information · p. 51

Competition and concentration

  • Research should examine how open-source AI, regulation, and venture capital can help smaller firms compete with dominant firms, while accounting for compliance burdens and short-term investment incentives.Suggested designs compare open-source and proprietary firms, evaluate competition-promoting regulation, and relate venture funding to innovation and market performance.Source: Supplementary Information · p. 52

Data governance

  • Future studies should test transparency protocols, anonymization, and consent models to reduce data misuse and privacy breaches without undermining personalization, trust, or data sharing.The proposed experiments measure misuse, privacy breaches, user trust, sharing behavior, and personalized content-filtering effectiveness.Source: Supplementary Information · pp. 52–53

Misinformation

  • Research should investigate interventions against AI-generated misinformation, including dialoguing systems, detection models, and account verification, while addressing hallucinations, misclassification, and reduced platform trust.Suggested comparisons test misinformation beliefs, classify human and AI-generated texts, and measure bot-driven misinformation after verification.Source: Supplementary Information · pp. 53–54

Ethical cooperation

  • Studies should test how delegation, interface design, system descriptions, and cultural norms shape ethical behavior and cooperation in human-AI interactions.AI may reduce moral responsibility, whereas documenting decisions may increase accountability; cooperation may improve without reaching human-human levels, and conflicting norms complicate universal design.Source: Supplementary Information · pp. 54–55

Information plurality

  • Research should assess whether transparent sources, interactive query guidance, and algorithm education preserve autonomy and information plurality, while recognizing risks of overload, dependence, and unaddressed algorithmic bias.Proposed experiments measure alternative-source seeking, breadth of exploration, and users’ ability to find diverse information.Source: Supplementary Information · p. 55

Human contributions

  • Future work should examine how AI-generated content affects online diversity and human originality, including curation that prioritizes innovative human contributions without discarding valuable AI insights.Higher AI-content ratios may increase availability or creativity but can homogenize platforms, saturate them, or demotivate human contributors.Source: Supplementary Information · p. 56

Corporate Digital Responsibility

  • The section also calls for organizational strategies implementing Corporate Digital Responsibility, while identifying limited awareness and possible over-regulation as implementation challenges.The proposed agenda includes studying organizational barriers and developing a general framework for corporate digital responsibility.Source: Supplementary Information · p. 56

Workplace

  • AI may augment rather than fully automate work because automation offers efficiency while human knowledge workers provide unique insights, though oversight and quality-control demands may reduce freed capacity.Suggested research tracks strategic-task allocation before and after organizational AI pilots and examines automation investment relative to knowledge-worker recruitment.Source: Supplementary Information · p. 58
  • Generative AI may improve access to economic opportunities through remote work, language translation, and job matching, but algorithmic bias, limited technology access, and restrictions could perpetuate inequalities.Proposed studies compare AI-supported job matching, translation, and remote work outcomes with traditional platforms and teams, including low-resource language contexts.Source: Supplementary Information · pp. 58–59
  • AI’s effects on workers may include improved flexibility, innovation, and upskilling, but increased monitoring, workload, digital-literacy gaps, and technology apprehension could worsen burnout, work-life balance, or reskilling disparities.The proposed agenda includes longitudinal studies of satisfaction and burnout, AI-based educational interventions for workers, and digital-literacy programs for older workers and those facing displacement.Source: Supplementary Information · pp. 59–60
  • Intergenerational AI knowledge-sharing could improve team performance, inclusivity perceptions, and wellbeing, but resistance from senior workers or frustration among juniors could produce negative outcomes.A proposed diverse-team study evaluates performance, inclusivity perceptions, and psychological wellbeing following intergenerational knowledge-sharing.Source: Supplementary Information · p. 60

Education

  • Educational applications of generative AI may personalize feedback, simulations, homework support, and problem-solving practice, but overreliance, inaccurate outputs, and reduced initiative could impair independent learning and critical thinking.Proposed comparisons assess writing proficiency, conceptual understanding, engagement, independent learning, and problem-solving against standard teacher or conventional-tool approaches.Source: Supplementary Information · pp. 62–63
  • Teacher training through peer mentoring, collaborative instruction, and just-in-time support may improve AI integration, but differences in teaching styles, learning pace, and training breadth could limit implementation success.Suggested studies compare collaborative, individual, peer-mentoring, and just-in-time training approaches using teacher confidence, usage, and classroom implementation outcomes.Source: Supplementary Information · pp. 63–64
  • Generative AI may broaden students’ ideas and critical thinking with appropriate pedagogy, but unmanaged use could reduce originality or reinforce limited perspectives from training data.Proposed research tests pedagogically advanced uses, alternate viewpoints, divergent thinking, critical thinking, and inclusive classroom practices.Source: Supplementary Information · p. 64

Healthcare: clinical decision-making

  • AI-assisted clinical decision-making, diagnosis, and monitoring could improve accuracy and recovery management, but mistrust, low AI literacy, overreliance, and reduced clinician intuition may harm outcomes.Suggested studies compare AI-supported and traditional decisions, diagnostic methods, and post-operative monitoring while measuring accuracy, bias, and recovery.Source: Supplementary Information · p. 66

Healthcare: administration

  • AI-integrated EHRs and speech-to-text could reduce documentation time, errors, and long-term administrative costs, but implementation, training, and recognition errors may initially increase burdens or reduce record quality.Proposed comparisons measure documentation time, record accuracy and completeness, and direct and indirect costs against traditional systems.Source: Supplementary Information · p. 67

Healthcare: prediction and public health

  • Predictive models could enable earlier, targeted, and personalized interventions that reduce disparities, but inaccurate or biased data and predictions could trigger unnecessary care or miss at-risk patients.Research directions include diabetes-risk prediction, vaccine distribution, readmission reduction, and outbreak forecasting, with equity and response outcomes measured.Source: Supplementary Information · pp. 67–69

Healthcare: accessibility and trust

  • AI could expand healthcare access, trust, and satisfaction in underserved populations, but connectivity, digital literacy, biased data, complex explanations, insincere personalization, and negative media may widen disparities or undermine confidence.Proposed studies examine mobile health applications, outbreak prediction, explanations, personalized communication, and media campaigns across rural and underserved settings.Source: Supplementary Information · pp. 68–69

Healthcare: social connection

  • AI chatbots and pro-social prompts may reduce loneliness and encourage courteous digital interaction, but inauthenticity and overreliance could inhibit human contact, deepen isolation, or fail to meet social needs.Suggested experiments compare chatbot interaction with traditional social programs and human-to-human exchanges with and without AI prompts.Source: Supplementary Information · p. 70

Related papers

Loading 2401.05377v2…