|
on Artificial Intelligence |
| By: | Drew Fudenberg; David K Levine |
| Date: | 2026–05–17 |
| URL: | https://d.repec.org/n?u=RePEc:cla:levarc:735347000000000055 |
| By: | Thomas Renault; Antonin Bergeaud; Cl\'ement Bosquet |
| Abstract: | Kusumegi et al. (2025) study whether researchers' preprint output rises after adopting large language models (LLMs), dating adoption as the first month in which at least one submitted abstract exceeds an LLM-detection threshold. We show that this treatment-timing rule is mechanically related to output. The probability that at least one paper is flagged in a month is increasing in the number of papers submitted in that month, so detected-adoption months are disproportionately high-output months. An event study centered on first detection can therefore display positive post-event dynamics even when the flagging rule contains no information about true LLM adoption, because the omitted pre-treatment period is selected from months with no prior detection. We demonstrate this in a simulation: with i.i.d. productivity and no causal effect, first-detection timing generates a spurious positive post-treatment path. We also replicate the stacked event study of Kusumegi et al. (2025) and show that three placebo exercises (random paper-level assignment, neutral keyword flags, and a pre-ChatGPT observation window) each produce a similarly positive post-treatment pattern. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2605.17979 |
| By: | Ali Aouad; Thodoris Lykouris; Huiying Zhong |
| Abstract: | Generative Artificial Intelligence (AI) tools are rapidly adopted in the workplace and in education, yet the empirical evidence on AI's impact remains mixed. We propose a model of human-AI interaction to better understand and analyze several mechanisms by which AI affects productivity. In our setup, human agents with varying skill levels exert utility-maximizing effort to produce certain task outcomes with AI assistance. We find that incorporating either endogeneity in skill development or in AI unreliability can induce a productivity paradox: increased levels of AI assistance may degrade productivity, leading to potentially significant shortfalls. Moreover, we examine the long-term distributional effect of AI on skill, and demonstrate that skill polarization can emerge in steady state when accounting for heterogeneity in AI literacy -- the agent's capability to identify and adapt to inaccurate AI outputs. Our results elucidate several mechanisms that may explain the emergence of human-AI productivity paradoxes and skill polarization, and identify simple measures that characterize when they arise. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2605.11350 |
| By: | Joshua S. Gans |
| Abstract: | Empirical measures of AI's wage effect typically hold fixed the bundle of activities a worker is paid for at its pre-AI shape. We argue that this assumption hides much of the action. When automation breaks a job apart, firms decide how to recombine the surviving activities; whether they rebundle them into one broad role or split them into specialist roles changes which surviving skills the labour market actually rewards. A skill that played no role in the pre-AI wage can become the dominant component of the post-AI wage, while a skill that anchored the pre-AI wage can disappear from the schedule. We develop an assignment model in which the priced human bundle is endogenous, and we use it to show that a fixed-bundle wage regression can mis-sign the effect of AI exposure. In general, the omitted-redesign bias has no unconditional sign: it is the residual covariance between exposure and role-specific redesign terms. Under explicit sufficient conditions, exposure-correlated unbundling loads specialist comparative-advantage premia onto the exposure coefficient, while exposure-correlated rebundling loads a different, often opposite, omitted term. The sign must therefore be measured from local post-AI partition changes rather than assumed from exposure alone. |
| JEL: | D23 J23 J24 J31 O33 |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:35211 |
| By: | Dmitry Dagaev; Egor Ivanov; Petr Parshakov; Alexey Savvateev; Gleb Vasiliev |
| Abstract: | The emergence of large language models (LLMs) has spurred economists to study how humans and LLMs behave in strategic settings. We organized a series of round-robin tournaments in the Colonel Blotto game. This game attracts game theorists' attention due to high-dimensional action space and the absence of pure strategy Nash equilibria. In the first tournament, more than 200 human participants competed against one another. In the second tournament, several popular LLMs were invited to submit strategies. In the third tournament, we matched the number of LLM strategies to the number submitted by humans. We find that humans more often employ better-calibrated intermediate-level allocation heuristics and outperform the simpler, more stereotyped strategies submitted by LLMs. Strategic sophistication is key to success if and only if the necessary level of reasoning depth is reached, while lower and higher levels of reasoning offer no clear advantage over the primitive strategies. Among humans, field of study weakly predicts success: participants with STEM backgrounds perform better in the first tournament. Surprisingly, humans almost do not adjust their strategies across tournaments with different sets of opponents. This result suggests that humans base their choices primarily on the game's rules rather than on the identity of their opponents, treating LLMs much like human competitors. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2605.22095 |
| By: | Dominik Atella-Suri (University of Bonn); Sebastian Kube (University of Bonn) |
| Abstract: | Artificial intelligence (AI) is increasingly becoming part of economic decision-making. Yet, in many strategic interactions, individuals may not know whether others rely on AI when forming their decisions. We examine whether decision-makers who are themselves not allowed to use AI behave differently when other group members may consult AI. In an incentivized experiment with a public goods game and a weakest-link game, we exogenously vary whether group members are allowed to use AI to inform their decisions. We find that AI can affect strategic interaction even when it is not directly used by the decision-maker: merely knowing that others may use AI reduces cooperation in the public goods game and effort provision in the weakest-link game. Participants also perceive group members who may use AI as socially more distant and report lower beliefs about appropriate and expected contributions and effort levels. At the same time, the shares of conditional cooperators and conditional coordinators remain largely stable across treatments. These findings suggest that AI is not only a private decision aid but can also shape the social and strategic environment in which economic decisions are made. |
| Keywords: | Cooperation, Coordination, Human-AI Interaction, Artificial Intelligence, Experiment |
| JEL: | C71 D83 D91 |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:ajk:ajkdps:407 |
| By: | Ze Wang; Guobin Shen; Michael Thaler |
| Abstract: | Humans increasingly delegate decisions to language models, yet whether these systems reproduce or reshape human patterns of discrimination remains unclear. Here we run a large-scale study to analyse whether language models use demographic information in hiring decisions. We show, across 27 models and 177 occupations, that language models give female and Black candidates hiring advantages relative to otherwise-comparable male and white candidates, while giving disabled candidates disadvantages. The differences are meaningful in magnitude: the role of race, gender, and disability status is comparable to six months to one year of additional education. Post-training alignment is the primary driver: relative to matched pre-trained models, alignment amplifies advantages for female and Black candidates by 325% and 330%, and disadvantages for disabled candidates by 171%. Compared with previous human correspondence studies, language models reverse the direction of racial discrimination, attenuate the disability penalty, and amplify the female advantage by 190%. Alignment changes how models use qualification signals: alignment increases returns to skills and work experience overall, but relatively more so for female and Black candidates. Meanwhile, the absence of qualification signals harms marginalised groups more, particularly for disabled candidates, differences that may explain the asymmetry of alignment effects across groups we observe. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2605.13866 |
| By: | Nick Huntington-Klein |
| Abstract: | In educational settings, AI can be used as a learning aid, but can also be used to avoid schoolwork, thereby passing classes while learning little. Many existing studies on the impact of AI on education focus on AI use in controlled settings or with specialized tools. In this paper, the dropoff in ChatGPT activity during non-school summer months in 2023 and 2024 is used to identify areas with heavy educational AI use and thus estimate the educational impact of AI as it is actually used. I find no meaningful impact of AI usage on high school test score averages in either direction. These results imply that, to the extent that high school students use AI to avoid learning, it either does not matter much for their test performance or is cancelled out by positive uses of AI in the aggregate. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2605.08812 |
| By: | Bauer, Kevin; Langenbucher, Katja |
| Abstract: | Agentic artificial intelligence (AI) is increasingly used in corporate communication with investors, including drafting disclosures, answering queries, and summarizing financial information. While these systems can improve the accessibility and efficiency of public corporate communication, they also create risks for market integrity, such as inaccurate statements, unintended disclosure of sensitive information, and unequal access through personalized responses. This paper argues that existing disclosure and market-abuse frameworks remain substantively adequate but require clearer operational expectations for AI-driven communication. AI outputs delivered through issuer-controlled channels should be treated as corporate communications attributable to the issuer. A risk-based regulatory approach should therefore require governance oversight, separation of public and confidential data, safeguards against manipulation, auditable records of AI outputs, and human oversight for market-sensitive communications. Properly governed, agentic AI can enhance investor access to public information while preserving the principles of accurate, timely, and equal disclosure. |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:safepl:341102 |
| By: | Hyunso Kim; Hyo Kang; Jaeyong Song |
| Abstract: | Recent advances in generative artificial intelligence (AI) are reshaping who enters entrepreneurship, but not who reaches the top of the quality distribution. Using data on over 160, 000 product launches on Product Hunt, we find that entrepreneurial entry increased sharply following the public release of ChatGPT-3.5, driven disproportionately by solo entrepreneurs. This shift toward solo entry is particularly pronounced in categories that historically favored team-based ventures. However, much of this growth reflects low-commitment, experimental entry and does not translate into greater representation among the highest-quality outcomes. Team-based ventures are increasingly dominant in the top tiers of platform rankings. These findings suggest that generative AI lowers barriers to solo entrepreneurship while reinforcing team-based advantages. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2605.10291 |
| By: | Nuriye Melisa Bilgin; Gianmarco Ottaviano |
| Abstract: | Do the determinants of technology adoption depend on technological architecture? Using administrative data on Turkish firms from 2021 to 2024, we compare the adoption of traditional and generative artificial intelligence (GenAI).We show that GenAI adoption is driven by workforce skill intensity and is not positively associated with firm size, whereas traditional AI depends on both scale and skills. Firms that adopt both technologies are distinct and represent the most persistent adoption mode. Conditional on adoption, the skill-to-size ratio governs technology choice, and transition dynamics indicate a sequential process in which firms adopt GenAI before expanding to hybrid use. Exploiting the release of ChatGPT as a quasi-experimental reduction in access costs, we find that high-skill firms differentially increased GenAI adoption, while firm size played a limited role. These results suggest that the canonical size-based diffusion pattern is not universal but depends on the cost structure of technologies, with implications for innovation policy and productivity dispersion. |
| Keywords: | artificial intelligence, generative AI, technology adoption, firm heterogeneity |
| Date: | 2026–05–21 |
| URL: | https://d.repec.org/n?u=RePEc:cep:cepdps:dp2184 |
| By: | Michelle Yin; Burhan Ogut |
| Abstract: | A growing literature uses artificial intelligence platform conversation logs to measure occupation exposure. We show that these scores partly measure platform user base rather than the workforce. Holding outcome, sample, controls, and estimator fixed while varying only the platform input changes the post-ChatGPT employment coefficient by a factor of 1.9, and within-vendor consumer-versus-enterprise channels produce estimates that disagree in sign. Reweighting to Bureau of Labor Statistics workforce shares attenuates estimates by 42 to 93 percent. We formalize the non-classical measurement error, derive probability limits and partial-identification bounds for employment elasticities. The bias understates substitution more than augmentation. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2605.21743 |
| By: | Elliott Ash; Francesco Capozza; Sergio Galletta |
| Abstract: | We estimate demand for five conversational AI use cases — administrative assistance, studying, wellness, friendship, and romance — using a preregistered, within-respondent conjoint experiment in a U.S. online sample (N=1, 989). Purpose dominates interest: instrumental uses rank highest, romance lowest. Free pricing raises interest sharply but is attenuated for relational purposes, revealing non-monetary signaling costs that price reductions cannot eliminate. Privacy and personalization shift demand; interaction modality does not. Second-order beliefs exceed own interest in proportion to the purpose penalty, consistent with a single stigma parameter suppressing stated below latent demand. Corrected consumer surplus for a free romantic companion turns positive once suppression costs are recovered — the welfare mirror image of social media’s collective trap. |
| Keywords: | AI, conjoint experiment, WTP, second-order beliefs |
| JEL: | C81 C93 D82 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_12673 |
| By: | Anton Korinek (Peterson Institute for International Economics); Patrick McKelvey (Bank of Canada) |
| Abstract: | The authors construct a macroeconomic estimate of total AI production for the United States, combining inference and research and development/training activities and applying quality adjustments based on the evolution of API prices at fixed performance levels and the pace of algorithmic progress. They estimate that nominal AI compute spending grew over 140 percent per year each in 2024 and 2025, raw compute capacity grew over 200 percent per year, and quality-adjusted AI output grew over 2, 000 percent per year. These growth rates reflect three compounding forces: expanding data-center capacity, continued improvements in chip efficiency, and rapid algorithmic progress. The authors then employ their estimates to develop a nascent framework for "AI GDP" that tracks the AI economy as a coherent whole rather than dispersed across standard industry classifications. Quality-adjusted AI GDP grew by more than 2, 500 percent each in 2024 and 2025. The measures in the paper complement traditional national accounts by providing visibility into a fast-moving sector whose activity is difficult to isolate in existing statistics. The measures may serve as building blocks for satellite accounts that track AI's growing role in the economy. The conceptual framework and methodology used in this paper are described in detail in the technical appendix. |
| Keywords: | artificial intelligence, national accounts, GDP mismeasurement, AI satellite accounts, quality-adjusted prices, algorithmic progress, AI GDP |
| JEL: | E01 O33 O47 E22 |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:iie:wpaper:wp26-9 |
| By: | Anton Korinek (Peterson Institute for International Economics); Patrick McKelvey (Bank of Canada) |
| Abstract: | The AI economy in the United States has been growing at an unprecedented rate, but this extraordinary growth is largely invisible in conventional statistics. The authors propose developing an "AI GDP" framework to better measure AI's growing role in the economy. -Key Takeaways -Quality-adjusted AI production in the United States grew at over 2, 000 percent per year in 2024 and 2025, driven by three compounding forces: expanding data-center capacity, hardware efficiency gains, and—the largest of the three—algorithmic progress. -Treating the AI sector as a coherent economic entity yields preliminary estimates of nominal AI GDP at approximately $250 billion in 2025, growing at roughly 2, 600 percent per year in quality-adjusted real terms. -National economic statistics accounts were not designed to track this kind of activity. Statistics agencies should begin developing AI-focused satellite accounts now, before the measurement gap becomes a policy gap. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:iie:pbrief:pb26-7 |
| By: | OECD |
| Abstract: | This paper develops a new OECD measure of occupational AI exposure based on the OECD AI Capability Indicators. The measure addresses the need for a forward-looking, transparent and updateable approach to assessing how AI may affect work, skills and education over the next 5 to 10 years. It does so by mapping AI capabilities across nine cognitive, social and physical domains to occupational requirements and constructing an AI Capability Gap index. Lower gap values indicate that current AI systems are closer to the capability profile required for an occupation, implying higher potential exposure. The results show that current AI capabilities are closest to occupations involving routine information processing, administrative work and codifiable tasks, and furthest from occupations requiring contextual judgement, interpersonal understanding, complex decision making and responsibility. The analysis also shows that exposure is multidimensional: some occupations are most exposed to language and reasoning systems, while others are more exposed to robotics, machine vision and embodied AI. The measure provides a transparent foundation for analysing task-level transformation, changing skill demand and future labour-market effects, while recognising that actual impacts will depend on adoption, regulation, organisational change and social choice. |
| Date: | 2026–05–26 |
| URL: | https://d.repec.org/n?u=RePEc:oec:comaaa:59-en |
| By: | Alex Leung; Rex Zhang; Ervin Ling; Kentaroh Toyoda; SiewMei Loh |
| Abstract: | The rapid diffusion of agentic AI has created a new coverage problem for commercial insurance: some AI-mediated losses are now affirmatively insured, some create silent-AI exposure under legacy cyber, technology errors-and-omissions (E&O), directors-and-officers (D&O), employment practices liability (EPLI), crime, and media policies, and others are being actively excluded. This paper maps that emerging boundary by coding 55 AI threat classes against 26 insurance products, endorsements, and exclusion regimes using public carrier materials and OWASP/MITRE threat catalogs. We identify a four-tier insurability frontier: affirmatively insured perils, silent-AI exposures, actively excluded perils, and perils outside conventional private insurance structures. Our coding measures publicly claimed positioning rather than executed contract wording; the headline statistics describe what carriers publicly state about coverage, not what would be paid in any specific claim. Three patterns emerge. First, affirmative AI coverage is beginning to differentiate by primary risk emphasis: public materials often position Munich Re around model performance and drift, Armilla and parts of the Lloyd's market around hallucination and broader AI liability, Tokio Marine Kiln and CFC around IP and technology E&O concerns, Apollo ibott around emerging autonomous system liability, and Coalition around deepfake and AI-enabled cyber response. Second, legacy lines retain silent-AI exposure where AI is an instrumentality rather than the legal cause of loss. Third, foundation model concentration is the clearest genuinely novel insurability frontier because upstream model failure can correlate losses across many cedents at once; the relevant market design question is which insurability constraint each candidate structure relaxes, not merely which systemic risk template exists. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2605.18784 |
| By: | Bektemir Ysmailov (Nazarbayev University, Graduate School of Business) |
| Abstract: | I construct a new measure of financial constraints by applying a large language model to narrative disclosures in firms' Management's Discussion and Analysis from Form 10-K filings. The model evaluates each filing as a finance expert and classifies the firm's external financing difficulty on an ordered scale, producing the GenAI FC Index. The index captures contextual signals - such as nuanced liquidity discussions - that traditional accounting-based and prior text-based proxies often miss. It behaves sensibly in both the time series and cross-section and shows only moderate correlations with existing measures, indicating that it contains distinct information. Behavioral tests reveal that firms classified as constrained recycle far less equity and are substantially more likely to omit dividends, and less likely to initiate or increase them. Across these settings, the GenAI FC Index yields stronger and more consistent behavioral separation than benchmark text-based measures. The results demonstrate that generative AI can extract economically meaningful information about firms' financing frictions at scale. |
| Keywords: | financial constraints, generative AI (GenAI), large language models (LLMs), textual analysis, MD&A disclosures, corporate finance |
| JEL: | G30 G32 M41 C81 |
| Date: | 2026–01 |
| URL: | https://d.repec.org/n?u=RePEc:asx:nugsbw:2026-01 |
| By: | Congluo Xu; Jiuyue Liu; Xiangsheng Zheng; Ziyang Li |
| Abstract: | The rapid development of artificial intelligence motivates firms to engage in AI washing. This study examines whether strategic policy shocks increase debt financing costs for such firms. Leveraging China's 14th Five Year Plan as a quasi natural experiment, we identify AI washing through the residual between AI narrative intensity and patent output. External validation confirms this decoupling reflects strategic deception evidenced by subsidy extraction and future regulatory violations rather than benign ambition, supporting its validity as an AI washing proxy. Difference in differences estimations reveal that AI washing firms experience a 12.5 basis point relative increase in debt financing cost afterward. Joint estimation confirms simultaneous adjustments across financing and innovation margins. Management shareholding and analyst attention amplify the penalty while supply chain concentration and bank proximity attenuate it. Results remain robust across checks. Our findings illuminate how macro level policy shocks activate market discipline in emerging market debt markets. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2605.16808 |
| By: | Hung Q. Tran |
| Abstract: | Divergent regulatory regimes for data, driven by different motivations, ranging from privacy protection in the European Union to information control in China, could eventually produce distinctively different, and possibly contradictory, bodies of data. Artificial-intelligence models trained on those datasets could produce differing and possibly even conflicting outputs. To the extent that AI outputs start to shape human perception and to influence decisions, in governments and businesses, and among the public, antagonistic AI models would reinforce the mutual mistrust and hostility inherent in the current geopolitical environment, potentially making it harder to resolve conflicts. As a consequence, the fragmentation of data is becoming an important issue in the evolution of AI and its potential impact on human society. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:ocp:pbtrad:pb08_26 |
| By: | Jagdish Tripathy; Marcus Buckmann |
| Abstract: | Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconnect: models show no output-level bias, yet retain and amplify demographic representations across model layers. Through activation steering and novel cross-layer interventions, we demonstrate that this suppressed information is decision-relevant: when reinjected at critical layers, it produces near-complete decision reversals. Critically, this latent bias is asymmetric - steering interventions affect decisions in one demographic direction, while producing minimal effects in reverse - and susceptible to adversarial prompt engineering and parameter-efficient fine-tuning. These findings demonstrate that behavioural audits focused on outputs are insufficient: fair outputs can mask exploitable internal biases. They also motivate dual-layer testing frameworks combining output evaluation with representational analysis for AI governance in high-stakes decisions. |
| Date: | 2026–05 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2605.15217 |