|
on Artificial Intelligence |
| By: | Zac, Amit; Gal, Michal S. |
| Abstract: | The integration of large language models (LLMs) into recommender systems (RS) has given rise to a new generation of Conversational RS (CRS). This study asks how CRS systems shape consumer behavior, and, in particular, their spending. Despite the rapid proliferation of such systems, including widely used tools like OpenAI's ChatGPT and Google's Gemini, we still lack evidence on their behavioral effects. This study provides the first controlled empirical test of CRS influence on real purchasing decisions. In a laboratory experiment, complemented by large-scale API studies, participants were randomly assigned to one of four conditions: a traditional search baseline, GPT, Gemini, or a customized GPT designed to steer users toward more expensive products. CRS consistently increase consumer expenditures, with Customized GPT producing the highest average spending. Importantly, these effects are not driven by differences in perceived product quality, prior shopping experience, or generalized trust. Rather, they stem from subtle linguistic framing and increased exposure to premium brands. Taken together, the findings position LLM-based CRS as novel and potent choice architects with downstream implications for consumer protection, market design, and regulatory oversight. |
| Keywords: | recommendation systems, conversational recommender systems (CRS), largelanguage models (LLMS), LLM-empowered recommendation systems, consumerbehavior, algorithmic choice architecture, algorithmic bias, human-AI shoppinginteraction, trust in CRS, algorithmic regulation |
| Date: | 2025 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:cbscwp:336742 |
| By: | Jana Friedrichsen; Julia Schwarz; Michel Clement |
| Abstract: | Artificial intelligence (AI) is rapidly reshaping society, including the music industry. Recent advancements in generative AI enable users to create music from text-based prompts, raising questions about public perception and valuation of AI-generated music. We conducted three studies with German-speaking participants (Study 1: N=2000, Study 2: N=425; Study 3: N=1248) to explore awareness, enjoyment, and willingness to pay for AI music. After finding no clear rejection of AI composed music in Study 1, Study 2 varied whether listeners knew the music was AI-generated. Study 3 involved regular listeners of pop and electronic dance music, manipulating song origin (human vs. AI) and disclosure. Results show that listeners generally could not distinguish between AI and human-made songs. When unaware, participants slightly preferred AI music and valued it equally. However, disclosing that AI had been used to create compositions reduced appreciation and willingness to pay. We explore how reactions differ by genre and individual attitudes toward AI and discuss implications for the music industry and for regulatory initiatives. |
| Keywords: | artificial intelligence, music, willingness-to-pay, ethics of AI |
| JEL: | D12 Z11 O33 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_12405 |
| By: | Dominik Suri (University of Bonn); Simon Gächter (University of Nottingham); Sebastian Kube (University of Bonn) |
| Abstract: | AI-driven systems are rapidly moving from decision support to directing human behavior through rules, recommendations, and compliance requests. This shift expands everyday human–AI interaction and raises the possibility that AI may function as an authority figure. However, the behavioral consequences of AI as an authority figure remain poorly understood. We investigate whether individuals differ in their willingness to comply with arbitrary rules depending on whether these rules are attributed to an AI agent (ChatGPT) or to a fellow human. In a between-subject design, 977 US Prolific users completed the coins task: they could earn a monetary payoff by stopping the disappearance of coins at any time, but a rule instructed them to wait for a signal before doing so. There are no conventional reasons to follow this rule: complying is costly and nobody is harmed by non-compliance. Despite this, we find high rule-following rates: 64.3% followed the rule set by ChatGPT and 63.9% complied with the human-set rule.Descriptive and normative beliefs about rule following, aswell as compliance conditional on these beliefs, are also largely unaffected by the rule’s origin. However, subjective social closeness to the rule setter significantly predicts how participants condition their behavior on social expectations: when participants perceive the rule setter as subjectively closer, conditional compliance is higher and associated beliefs are stronger, irrespective of whether the rule setter is human or AI. |
| Keywords: | Artificial intelligence, AI-human interaction, ChatGPT, rule-following, coins task, CRISP framework, social expectations, conditional rule conformity, social closeness, IOS11, online experiments. |
| JEL: | C91 D91 Z13 |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:ajk:ajkdps:391 |
| By: | Thomas R. Cook; Sophia Kazinnik; Zach Modig; Nathan M. Palmer |
| Abstract: | Large language models (LLMs) are now used for economic reasoning, but their implicit "preferences" are poorly understood. We study these preferences by analyzing revealed choices in canonical allocation games and a sequential job-search environment. In dictator-style allocation games, most models favor equal splits, consistent with inequality aversion. Structural estimation of Fehr-Schmidt parameters suggests this aversion exceeds levels typically observed in human experiments. However, LLM preferences prove malleable. Interventions such as prompt framing (e.g., masking social context) and control vectors reliably shift models toward more payoff-maximizing behavior, while persona-based prompting has more limited impact. We then extend our analysis to a sequential decision-making environment based on the McCall job search model. Here, we recover implied discount factors from accept/reject behavior, but find that responses are less consistently rationalizable and preferences more fragile. Our findings highlight two core insights: (i) LLMs exhibit structured, latent preferences that often align with human behavioral norms, and (ii) these preferences can be steered, albeit more effectively in simple settings than in complex, dynamic ones. |
| Keywords: | Behavioral economics; Game theory; Search and matching models |
| JEL: | C63 C68 C61 D14 D83 D91 E20 E21 |
| Date: | 2026–01–30 |
| URL: | https://d.repec.org/n?u=RePEc:fip:fedgfe:102439 |
| By: | Engberg, Erik; Görg, Holger; Hellsten, Mark; Javed, Farrukh; Lodefalk, Magnus; Längkvist, Martin; Monteiro, Natália Pimenta; Nordås, Hildegunn Kyvik; Pulito, Giuseppe; Schroeder, Sarah; Tang, Aili |
| Abstract: | • Occupations that are highly cognitive, non-physical, and low in social interaction - typically higher-skill white-collar roles such as data analysts, software developers, and translators - turn out to be highly AI-exposed • Occupations requiring manual dexterity or intensive interpersonal contact - such as construction labourers or nursing aides - remain among the least exposed to current AI technologies • Aggregate occupational exposure to AI has risen markedly since 2010, with especially rapid gains in the late 2010s and early 2020s • Our baseline estimates show no detectable effect of AI exposure on total firm employment, while it is associated with clear skill upgrading |
| Abstract: | • Berufe, die in hohem Maße kognitiv, nicht körperlich und mit geringen sozialen Interaktionen verbunden sind - typischerweise höher qualifizierte Angestelltenberufe wie Datenanalysten, Softwareentwickler und Übersetzer - sind offenbar in hohem Maße von KI betroffen • Berufe, die manuelle Geschicklichkeit oder intensiven zwischenmenschlichen Kontakt erfordern - wie Bauarbeiter oder Pflegehelfer - gehören nach wie vor zu den Berufen, die am wenigsten von aktuellen KI-Technologien betroffen sind • Die aggregierte berufliche Exposition gegenüber KI ist seit 2010 deutlich gestiegen, wobei die Zuwächse Ende der 2010er und Anfang der 2020er Jahre besonders rasch waren • Unsere Basisschätzungen zeigen keine erkennbaren Auswirkungen der KI-Exposition auf die Gesamtbeschäftigung in Unternehmen, während sie mit einer deutlichen Verbesserung der Qualifikationen einhergeht |
| Keywords: | Artificial intelligence, Labour demand, Multi-country firm-level evidence, Künstliche Intelligenz, Arbeitskräftenachfrage, Daten auf Unternehmensebene aus mehreren Ländern |
| JEL: | E24 J23 J24 N34 O33 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:ifwkpb:336753 |
| By: | Zhengyi Yu |
| Abstract: | This paper studies the impact of AI on productivity and inequality by focusing on the introduction of AlphaFold2. This AI algorithm can accurately predict protein structures, which were traditionally characterized by structural biologists through experiments. To capture the impact of AI on structural biologists at scale, I implement a difference-in-differences strategy comparing them to life scientists in other fields. While structural biologists did not change their overall number of publications with the availability of AlphaFold2, they experienced a 10% increase in citations to their new projects, a 4% rise in publications in high-impact journals, and a shift from their original research trajectory. However, the emergence of AI intensifies citation polarization between highly cited and less-cited researchers. Consistent with this growing inequality, highly cited scientists are twice as likely to incorporate AlphaFold2 successfully into their research as their less-cited peers. In addition, AI affects the next generation of researchers: the average years of experience of leading authors in protein structure papers increase after the emergence of AI. |
| Keywords: | AI, technology, labor productivity, inequality |
| JEL: | J21 J24 O33 D63 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_12462 |
| By: | Matthias Mertens; Natalia Fischl-Lanzoni; Neil Thompson |
| Abstract: | Do leading LLM developers possess a proprietary ``secret sauce'', or is LLM performance driven by scaling up compute? Using training and benchmark data for 809 models released between 2022 and 2025, we estimate scaling-law regressions with release-date and developer fixed effects. We find clear evidence of developer-specific efficiency advantages, but their importance depends on where models lie in the performance distribution. At the frontier, 80-90% of performance differences are explained by higher training compute, implying that scale--not proprietary technology--drives frontier advances. Away from the frontier, however, proprietary techniques and shared algorithmic progress substantially reduce the compute required to reach fixed capability thresholds. Some companies can systematically produce smaller models more efficiently. Strikingly, we also find substantial variation of model efficiency within companies; a firm can train two models with more than 40x compute efficiency difference. We also discuss the implications for AI leadership and capability diffusion. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.07238 |
| By: | Daron Acemoglu; David Autor; Simon Johnson |
| Abstract: | This paper defines pro-worker technologies, including Artificial Intelligence, as technologies that make human skills and expertise more valuable by expanding worker capabilities. Our conceptual framework distinguishes among five categories of technological change: labor-augmenting, capital-augmenting, automating, expertise-leveling, and new task-creating. Only the last category is unambiguously pro-worker, generating demand for novel human expertise rather than commodifying it. We illustrate these distinctions through hypothetical and real-world examples spanning aviation maintenance, electrical services, custodial work, education, patent examination, and gig delivery. While AI’s capacity to automate work is substantial, we argue that its potential to serve as a collaborator, by extending human judgment, enabling new tasks, and accelerating skill acquisition, is equally transformative and currently underexploited. We identify market failures, including misaligned firm and developer incentives, path dependence, and a pervasive pro-automation ideology, that may lead to underinvestment in pro-worker AI. We consider nine policy directions that would change incentives, including targeted investments in health care and education, tax code reform, antitrust enforcement, and intellectual property protections for worker expertise. |
| JEL: | J23 J24 J31 M50 O33 |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:34854 |
| By: | Christos A. Makridis |
| Abstract: | Using longitudinal data from the Gallup Panel covering more than 30, 000 U.S. employees surveyed from 2023 to 2025, I document heterogeneous adoption of generative AI and its "organizational transmission" within firms. First, I show that the share of active (occasional) AI users grew from 9% to 24% (10% to 23%). I subsequently document these patterns by organizational class of work, organizational hierarchy, income, industry, and occupation. Second, I show that perceived strategic clarity is the dominant correlate of frequent adoption: employees in clear-strategy organizations are roughly 26 percentage points more likely to report frequent use. Strategic clarity itself is tightly linked to managerial communication and credibility, including meaningful feedback and trust in leadership. Third, exploiting within-person variation, income-by-time controls, and a job-switcher design, I show that the relationship between AI use and worker outcomes is strongly contingent on organizational context. Frequent AI use is associated with substantially higher engagement and job satisfaction and with markedly lower burnout when organizations communicate a clear AI strategy, while these benefits are muted or reversed in low-clarity environments. These results are consistent with the predictions from a stylized theoretical model whereby workers are more likely to adopt and experiment with AI when they experience psychological safety. |
| Keywords: | generative AI, technology adoption, managers, trust, communication, organizational complementarities, panel data |
| JEL: | O33 M15 M54 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_12373 |
| By: | Florian Misch; Ben Park; Carlo Pizzinelli; Galen Sher |
| Abstract: | The discussion on Artificial Intelligence (AI) often centers around its impact on productivity, but macroeconomic evidence for Europe remains scarce. Using the Acemoglu (2024) approach we simulate the medium-term impact of AI adoption on total factor productivity for 31 European countries. We compile many scenarios by pooling evidence on which tasks will be automatable in the near term, using reduced-form regressions to predict AI adoption across Europe, and considering relevant regulation that restricts AI use heterogeneously across tasks, occupations and sectors. We find that the medium-term productivity gains for Europe as a whole are likely to be modest, at around 1 percent cumulatively over five years. While economically still moderate, these gains are still larger than estimates by Acemoglu (2024) for the US. They vary widely across scenarios and countries and are substantially larger in countries with higher incomes. Furthermore, we show that national and EU regulations around occupation-level requirements, AI safety, and data privacy combined could reduce Europe’s productivity gains by over 30 percent if AI exposure were 50 percent lower in tasks, occupations and sectors affected by regulation. |
| Keywords: | artificial intelligence, productivity, technology, regulation |
| JEL: | E24 J24 O30 O47 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_12401 |
| By: | Ping Wang; Russell Wong |
| Abstract: | How large are the effects of artificial intelligence (AI) on labor productivity and unemployment? We develop a labor-search model of technological unemployment where AI learns from workers, raises productivity, and displaces them if renegotiation fails. The model admits three steady states: no AI; some AI with limited capability, more job creation but higher unemployment; unbounded AI with endogenous growth and employment gains. Calibrated to U.S. data, the model implies a threefold productivity gain but a 23% employment loss, half within five years. Plausible parameters give rise to global and local indeterminacy with endogenous cycles in productivity and unemployment, underscoring the uncertainty of AI's impacts in line with a wide range of empirical findings. Equilibria are inefficient despite the Hosios condition; subsidizing jobs at risk of AI displacement is constrained optimal. |
| Keywords: | generative artificial intelligence; technological unemployment; search and bargaining; en dogenous growth; constrained efficiency; indeterminacy |
| JEL: | E20 J20 J64 L20 O30 O40 |
| Date: | 2026–02–23 |
| URL: | https://d.repec.org/n?u=RePEc:fip:fedrwp:102794 |
| By: | Arntz, Melanie; Baum, Myriam; Brüll, Eduard; Dorau, Ralf; Hartwig, Matthias; Matthes, Britta; Meyer, Sophie-Charlotte; Schlenker, Oliver; Tisch, Anita; Wischniewski, Sascha |
| Abstract: | Artificial intelligence (AI) is diffusing rapidly in the workplace, yet aggregate productivity gains remain limited. This paper examines the dual diffusion of AI - through both formal, employer-led and informal, employee-initiated adoption - as potential explanation. Using a representative survey of nearly 10, 000 employees in Germany, we document a high extensive but low intensive margin of usage: while 64 percent use AI tools, only 20 percent use them frequently. This diffusion is strongly skill-biased and depends less on establishment and regional characteristics. While formality is associated with more frequent usage, training, AI-based supervision, and higher perceived productivity gains, it does not broaden access. These patterns suggest that widespread informal usage can coexist with limited productivity effects when complementary investments and organizational integration lag behind. |
| Keywords: | artificial intelligence, AI, technology diffusion, formal and informal adoption, training, algorithmic management, productivity, inequality |
| JEL: | O33 O32 J24 J81 C83 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:zewdip:336765 |
| By: | Guillermo Cruces; Diego Fernández Meijide; Sebastian Galiani; Ramiro H. Gálvez; María Lombardi |
| Abstract: | Does generative artificial intelligence (AI) reinforce or reduce productivity differences across workers? Existing evidence largely studies AI within firms and occupations, where organizational selection compresses educational heterogeneity, leaving unclear whether AI narrows productivity gaps across individuals with substantially different levels of formal education. We address this question using a randomized online experiment conducted outside firms, in which 1, 174 adults ages 25–45 with heterogeneous educational backgrounds complete an incentivized, workplace-style business problem-solving task. The task is a general (not domain specific) exercise, and participants perform it either with or without access to a generative-AI assistant. Unlike prior work that studies heterogeneity within relatively homogeneous worker samples, our design targets the between–education-group productivity gap as the primary estimand. We find that AI increases productivity for all participants, with substantially larger gains for lower-education individuals. In the absence of AI access, higher-education participants outperform lower-education participants by 0.548 standard deviations; with AI access, this gap falls to 0.139 standard deviations, implying that generative AI closes about three quarters of the initial productivity gap. We interpret this pattern as evidence that generative AI narrows effective productivity differences in task execution by relaxing cognitive constraints that are more binding for lower-education individuals, even though underlying skill differences remain, as reflected in persistent education gaps in task performance and in a follow-up exercise without AI assistance. |
| JEL: | J24 O33 |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:34851 |
| By: | Mert Demirer; John J. Horton; Nicole Immorlica; Brendan Lucier; Peyman Shahidi |
| Abstract: | Production is a sequence of steps that can be executed (1) manually, (2) augmented with AI, or (3) fully automated within contiguous AI-executed steps called “chains.” Firms optimally bundle steps into tasks and then jobs, trading off specialization gains against coordination costs. We characterize the optimal assignment of humans and AI to steps and the firm’s resulting job structure, showing that comparative advantage logic can fail with AI chaining. The model implies non-linear productivity gains from AI quality improvements and admits a CES representation at the macro level. Empirical evidence supports the model’s key predictions that (1) AI-executed steps co-occur in chains, (2) dispersion of AI-exposed steps lowers AI execution at the job level, and (3) adjacency to AI-executed steps increases the likelihood that a step is AI-executed. |
| JEL: | D24 J23 J24 O33 |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:34859 |
| By: | Lukas Althoff; Hugo Reichardt |
| Abstract: | Artificial intelligence is changing which tasks workers do and how they do them. Predicting its labor market consequences requires understanding how technical change affects workers’ productivity across tasks, how workers adapt by changing occupations and acquiring new skills, and how wages adjust in general equilibrium. We introduce a dynamic task-based model in which workers accumulate multidimensional skills that shape their comparative advantage and, in turn, their occupational choices. We then develop an estimation strategy that recovers (i) the mapping from skills to task-specific productivity, (ii) the law of motion for skill accumulation, and (iii) the determinants of occupational choice. We use the quantified model to study generative AI’s impact via augmentation, automation, and a third and new channel — simplification — which captures how technologies change the skills needed to perform tasks. Our key finding is that AI substantially reduces wage inequality while raising average wages by 21 percent. AI’s equalizing effect is fully driven by simplification, enabling workers across skill levels to compete for the same jobs. We show that the model’s predictions line up with recent labor market data. |
| Keywords: | artificial intelligence, technology, labor markets, growth, inequality, wages, employment |
| JEL: | J24 J31 O33 J23 E24 D31 I26 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_12403 |
| By: | Anton Korinek; Lee Lockwood |
| Abstract: | Transformative artificial intelligence (TAI) - machines capable of performing virtually all economically valuable work - may gradually erode the two main tax bases that underpin modern tax systems: labor income and human consumption. We examine optimal taxation across two stages of artificial intelligence (AI)-driven transformation. First, if AI displaces human labor, we find that consumption taxation may serve as a primary revenue instrument, with differential commodity taxation gaining renewed relevance as labor distortions lose their constraining role. In the second stage, as autonomous artificial general intelligence (AGI) systems both produce most economic value and absorb a growing share of resources, taxing human consumption may become an inadequate means of raising revenue. We show that the taxation of autonomous AGI systems can be framed as an optimal harvesting problem and find that the resulting tax rate on AGI depends on the rate at which humans discount the future. Our analysis provides a theoretically grounded approach to balancing efficiency and equity in the Age of AI. We also apply our insights to evaluate specific proposals such as taxes on robots, compute, and tokens, as well as sovereign wealth funds and windfall clauses. |
| JEL: | H21 H24 O33 |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:34873 |
| By: | Leone de Castris, Arcangelo; Laher, Shakir; Ostmann, Florian |
| Abstract: | A significant barrier to AI adoption in the business world is the scarcity of clear, accessible information on how to leverage AI to enhance organisational productivity. Understanding the practical applications of AI is a prerequisite for companies to identify relevant opportunities and develop a strategy to operationalise them. To address this need, the AI Governance and Regulatory Innovation team at The Alan Turing Institute is pursuing a research project to illuminate how businesses in the four BridgeAI priority sectors of agriculture, forestry and fishing, construction, creative industries, and transportation and storage can leverage AI to be more productive. The first milestone of this project is the publication of a framework for categorising and analysing business applications of AI and a brief analysis of sector-specific AI use cases. Our findings are published as a series of five documents: four sector-specific briefings complemented by this paper presenting a framework to categorise and analyse AI use cases. The sector-specific briefings can be accessed from here. The paper on the framework presents the tool that we developed to categorise and analyse AI use cases in a business context. In addition to providing the hermeneutical structure underpinning our research, this tool provides a valuable resource for businesses trying to identify relevant AI opportunities. Companies can use this framework as a starting point to build on and develop their bespoke methodology to identify, select, and implement the right AI solutions. The second milestone of this project will be to refine the framework based on feedback collected after the publication of this first exploratory version and expand its scope to include information about the risks connected to each AI use case and the mitigation strategies that can be adopted to address those risks in that specific context. |
| JEL: | R14 J01 |
| Date: | 2025–01–23 |
| URL: | https://d.repec.org/n?u=RePEc:ehl:lserod:137290 |
| By: | Ivan Yotzov; Jose Maria Barrero; Nicholas Bloom; Philip Bunn; Steven J. Davis; Kevin M. Foster; Aaron Jalca; Brent H. Meyer; Paul Mizen; Michael A. Navarrete; Pawel Smietanka; Gregory Thwaites; Ben Zhe Wang |
| Abstract: | We present the first representative international data on firm-level AI use. We survey almost 6000 CFOs, CEOs and executives from stratified firm samples across the US, UK, Germany and Australia. We find four key facts. First, around 70% of firms actively use AI, particularly younger, more productive firms. Second, while over two thirds of top executives regularly use AI, their average use is only 1.5 hours a week, with one quarter reporting no AI use. Third, firms report little impact of AI over the last 3 years, with over 80% of firms reporting no impact on either employment or productivity. Fourth, firms predict sizable impacts over the next 3 years, forecasting AI will boost productivity by 1.4%, increase output by 0.8% and cut employment by 0.7%. We also survey individual employees who predict a 0.5% increase in employment in the next 3 years as a result of AI. This contrast implies a sizable gap in expectations, with senior executives predicting reductions in employment from AI and employees predicting net job creation. |
| JEL: | E0 |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:34836 |
| By: | Sumin Kim; Jihoon Kwon; Yoon Kim; Nicole Kagan; Raffi Khatchadourian; Wonbin Ahn; Alejandro Lopez-Lira; Jaewon Lee; Yoontae Hwang; Oscar Levy; Yongjae Lee; Chanyeol Choi |
| Abstract: | Mention markets, a type of prediction market in which contracts resolve based on whether a specified keyword is mentioned during a future public event, require accurate probabilistic forecasts of keyword-mention outcomes. While recent work shows that large language models (LLMs) can generate forecasts competitive with human forecasters, it remains unclear how input context should be designed to support accurate prediction. In this paper, we study this question through experiments on earnings-call mention markets, which require forecasting whether a company will mention a specified keyword during its upcoming call. We run controlled comparisons varying (i) which contextual information is provided (news and/or prior earnings-call transcripts) and (ii) how \textit{market probability}, (i.e., prediction market contract price) is used. We introduce Market-Conditioned Prompting (MCP), which explicitly treats the market-implied probability as a prior and instructs the LLM to update this prior using textual evidence, rather than re-predicting the base rate from scratch. In our experiments, we find three insights: (1) richer context consistently improves forecasting performance; (2) market-conditioned prompting (MCP), which treats the market probability as a prior and updates it using textual evidence, yields better-calibrated forecasts; and (3) a mixture of the market probability and MCP (MixMCP) outperforms the market baseline. By dampening the LLM's posterior update with the market prior, MixMCP yields more robust predictions than either the market or the LLM alone. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.21229 |
| By: | Alexis Antoniades; Chuan He; Zheming Liang; Mingzhi Jimmy Xu |
| Abstract: | Early adoption of artificial intelligence (AI) reshaped how firms responded to market dynamics by enhancing data collection and analysis. Linking China's universe of customs shipments to millions of online job ads, we tracked AI hiring in sales, marketing, and analytics to build a firm-level proxy for non-production AI and map its exposure across products within firms. To structure our analysis, we introduce a model in which firms confront asymmetric information about heterogeneous consumer preferences and show that AI mitigates these frictions by sharpening firms' ability to learn demand patterns across markets. The model predicts — and the data confirm — that AI-intensive firms fine-tune their product mix and prices with greater precision: they are more likely to export, expand their product lines, and adjust market choices. Crucially, this refinement appears as narrower price dispersion but wider quantity dispersion across destinations — effects strongest for differentiated goods, larger firms, and sales to high-income economies. Together, the evidence shows that AI eases demand-side information frictions, allowing firms to optimize their global reach strategies. |
| Keywords: | artificial intelligence, export behavior, product differentiation |
| JEL: | F14 O14 J24 |
| Date: | 2026 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_12456 |
| By: | Kunihiro Miyazaki; Takanobu Kawahara; Stephen Roberts; Stefan Zohren |
| Abstract: | The advancement of large language models (LLMs) has accelerated the development of autonomous financial trading systems. While mainstream approaches deploy multi-agent systems mimicking analyst and manager roles, they often rely on abstract instructions that overlook the intricacies of real-world workflows, which can lead to degraded inference performance and less transparent decision-making. Therefore, we propose a multi-agent LLM trading framework that explicitly decomposes investment analysis into fine-grained tasks, rather than providing coarse-grained instructions. We evaluate the proposed framework using Japanese stock data, including prices, financial statements, news, and macro information, under a leakage-controlled backtesting setting. Experimental results show that fine-grained task decomposition significantly improves risk-adjusted returns compared to conventional coarse-grained designs. Crucially, further analysis of intermediate agent outputs suggests that alignment between analytical outputs and downstream decision preferences is a critical driver of system performance. Moreover, we conduct standard portfolio optimization, exploiting low correlation with the stock index and the variance of each system's output. This approach achieves superior performance. These findings contribute to the design of agent structure and task configuration when applying LLM agents to trading systems in practical settings. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.23330 |
| By: | Yaxuan Kong; Hoyoung Lee; Yoontae Hwang; Alejandro Lopez-Lira; Bradford Levy; Dhagash Mehta; Qingsong Wen; Chanyeol Choi; Yongjae Lee; Stefan Zohren |
| Abstract: | Large Language Models (LLMs) are increasingly integrated into financial workflows, but evaluation practice has not kept up. Finance-specific biases can inflate performance, contaminate backtests, and make reported results useless for any deployment claim. We identify five recurring biases in financial LLM applications. They include look-ahead bias, survivorship bias, narrative bias, objective bias, and cost bias. These biases break financial tasks in distinct ways and they often compound to create an illusion of validity. We reviewed 164 papers from 2023 to 2025 and found that no single bias is discussed in more than 28 percent of studies. This position paper argues that bias in financial LLM systems requires explicit attention and that structural validity should be enforced before any result is used to support a deployment claim. We propose a Structural Validity Framework and an evaluation checklist with minimal requirements for bias diagnosis and future system design. The material is available at https://github.com/Eleanorkong/Awesome-F inancial-LLM-Bias-Mitigation. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.14233 |
| By: | Zeping Li; Guancheng Wan; Keyang Chen; Yu Chen; Yiwen Zhao; Philip Torr; Guangnan Ye; Zhenfei Yin; Hongfeng Chai |
| Abstract: | Recent works have increasingly applied Large Language Models (LLMs) as agents in financial stock market simulations to test if micro-level behaviors aggregate into macro-level phenomena. However, a crucial question arises: Do LLM agents' behaviors align with real market participants? This alignment is key to the validity of simulation results. To explore this, we select a financial stock market scenario to test behavioral consistency. Investors are typically classified as fundamental or technical traders, but most simulations fix strategies at initialization, failing to reflect real-world trading dynamics. In this work, we assess whether agents' strategy switching aligns with financial theory, providing a framework for this evaluation. We operationalize four behavioral-finance drivers-loss aversion, herding, wealth differentiation, and price misalignment-as personality traits set via prompting and stored long-term. In year-long simulations, agents process daily price-volume data, trade under a designated style, and reassess their strategy every 10 trading days. We introduce four alignment metrics and use Mann-Whitney U tests to compare agents' style-switching behavior with financial theory. Our results show that recent LLMs' switching behavior is only partially consistent with behavioral-finance theories, highlighting the need for further refinement in aligning agent behavior with financial theory. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.07023 |
| By: | Nguyen, Manh-Hung |
| Abstract: | I develop a growth model in which AI-generated content contaminates the knowledge commons, creating two nested irreversibilities. A derivative trap arises when recombinative output crosses a threshold in the corpus, degrading frontier productivity faster than talent reallocation or R&D subsidies can offset. A governance trap arises because the institutional capacity to distinguish frontier from derivative knowledge–epistemic capital–is itself a depletable stock. In the baseline simulation, the governance trap preempts the derivative trap by roughly nine years, closing the window for effective policy while measured innovation remains positive. The competitive equilibrium features a double wedge: frontier knowledge is undervalued and derivative output overvalued, driving a strict instrument hierarchy in which epistemic investment is a precondition for governance, which is a precondition for R&D subsidies. The welfare cost of inaction is 6.8% in consumption-equivalent terms. |
| Keywords: | Derivative trap; data quality, epistemic capital; governance trap; innovation policy; forward invariance |
| JEL: | O31 O33 O38 D83 |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:tse:wpaper:131486 |
| By: | Alex Farach |
| Abstract: | Task-based models of AI and labor hold organizational structure fixed, analyzing how technology shifts task assignments within a given firm architecture. Yet emerging evidence shows firms flattening hierarchies in response to AI adoption -- a phenomenon these models cannot generate. We extend the task-based framework by introducing agent capital (K_A): AI systems that reduce coordination costs within organizations, expanding managerial spans of control and enabling endogenous task creation. We derive five propositions characterizing how coordination compression affects output, hierarchy, manager demand, wage dispersion, and the task frontier. The model generates a regime fork: depending on whether agent capital complements all workers broadly (general infrastructure) or high-skill managers disproportionately (elite complementarity), the same technology produces either broad-based productivity gains or superstar concentration, with sharply divergent distributional consequences. Numerical simulations with heterogeneous managers and workers across a 2x2 parameter space (elite complementarity x endogenous task creation) confirm sharp regime divergence: in settings where coordination compression substantially expands employment, economy-wide inequality falls in all regimes, but the rate of reduction is regime-dependent and the manager-worker wage gap widens universally. The distributional impact of AI hinges not on the technology itself but on the elasticity of organizational structure -- and on who controls that elasticity. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.16078 |
| By: | Daniel W. O'Neill; Stefano Vrizzi; Noemi Luna Carmeno; Felix Creutzig; Jefim Vogel |
| Abstract: | Artificial intelligence (AI) is advancing exponentially and is likely to have profound impacts on human wellbeing, social equity, and environmental sustainability. Here we argue that the "alignment problem" in AI research is also an economic alignment problem, as developing advanced AI inside a growth-based system is likely to increase social, environmental, and existential risks. We show that post-growth research offers concepts and policies that could substantially reduce AI risks, such as by replacing optimisation with satisficing, using the Doughnut of social and planetary boundaries to guide development, and curbing systemic rebound with resource caps. We propose governance and business reforms that treat AI as a commons and prioritise tool-like autonomy-enhancing systems over agentic AI. Finally, we argue that the development of artificial general intelligence (AGI) may require a new economics, for which post-growth scholarship provides a strong foundation. |
| Date: | 2026–02 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2602.21843 |