|
on Artificial Intelligence |
| By: | Growiec, Jakub; Prettner, Klaus |
| Abstract: | Recent advances in artificial intelligence (AI) have led to a diverse set of predictions about its long-term impact on humanity. A central Focus is the potential emergence of transformative AI (TAI), eventually capable of outperforming humans in all economically valuable tasks and fully automating labor. Discussed scenarios range from human extinction after a misaligned TAI takes over ("AI doom") to unprecedented economic growth and abundance ("post-scarcity"). However, the probabilities and implications of these scenarios remain highly uncertain. Here, we organize the various scenarios and evaluate their associated existential risks and economic outcomes in terms of aggregate welfare. Our analysis shows that even low-probability catastrophic outcomes justify large investments in AI safety and alignment research. We find that the optimizing representative individual would rationally allocate substantial resources to mitigate extinction risk; in some cases, she would prefer not to develop TAI at all. This result highlights that current global efforts in AI safety and alignment research are vastly insufficient relative to the scale and urgency of existential risks posed by TAI. Our findings therefore underscore the need for stronger safeguards to balance the potential economic benefits of TAI with the prevention of irreversible harm. Addressing these risks is crucial for steering technological progress toward sustainable human prosperity. |
| Keywords: | Transformative Artificial Intelligence (TAI); Economic Growth; Technological Singularity; Growth Explosion; AI Takeover; AI Alignment; AI Doom |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:wiw:wus005:73284705 |
| By: | Fabrizio Dell'Acqua; Charles Ayoubi; Hila Lifshitz; Raffaella Sadun; Ethan Mollick; Lilach Mollick; Yi Han; Jeff Goldman; Hari Nair; Stewart Taub; Karim Lakhani |
| Abstract: | We examine how artificial intelligence transforms the core pillars of collaboration—performance, expertise sharing, and social engagement—through a pre-registered field experiment with 776 professionals at Procter & Gamble, a global consumer packaged goods company. Working on real product innovation challenges, professionals were randomly assigned to work either with or without AI, and either individually or with another professional in new product development teams. Our findings reveal that AI significantly enhances performance: individuals with AI matched the performance of teams without AI, demonstrating that AI can effectively replicate certain benefits of human collaboration. Moreover, AI breaks down functional silos. Without AI, R&D professionals tended to suggest more technical solutions, while Commercial professionals leaned towards commercially-oriented proposals. Professionals using AI produced balanced solutions, regardless of their professional background. Finally, AI’s language-based interface prompted more positive self-reported emotional responses among participants, suggesting it can fulfill part of the social and motivational role traditionally offered by human teammates. Our results suggest that AI adoption at scale in knowledge work reshapes not only performance but also how expertise and social connectivity manifest within teams, compelling organizations to rethink the very structure of collaborative work. |
| JEL: | M15 M2 O3 O31 O33 |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:33641 |
| By: | Wang, Yuyan (Stanford U); Li, Pan (Georgia Institute of Technology); Chen, Minmin (Google, Inc) |
| Abstract: | Modern recommender systems use machine learning (ML) models to predict consumer preferences based on consumption history. Although these “black-box†models achieve impressive predictive performance, they often suffer from a lack of transparency and explainability. While explainable AI research suggests a tradeoff between the two, we demonstrate that combining large language models (LLMs) with deep neural networks (DNNs) can improve both. We propose LR-Recsys, which augments state-of-the-art DNN-based recommender systems with LLMs’ reasoning capabilities. LR-Recsys introduces a contrastive-explanation generator that leverages LLMs to produce human-readable positive explanations (why a consumer might like a product) and negative explanations (why they might not). These explanations are embedded via a fine-tuned AutoEncoder and combined with consumer and product features as inputs to the DNN to produce the final predictions. Beyond offering explainability, LR-Recsys also improves learning efficiency and predictive accuracy. To understand why, we provide insights using high-dimensional multi-environment learning theory. Statistically, we show that LLMs are equipped with better knowledge of the important variables driving consumer decision-making, and that incorporating such knowledge can improve the learning efficiency of ML models. Extensive experiments on three real-world recommendation datasets demonstrate that the proposed LR-Recsys framework consistently outperforms state-of-the-art black-box and explainable recommender systems, achieving a 3–14\% improvement in predictive performance. This performance gain could translate into millions of dollars in annual revenue if deployed on leading content recommendation platforms today. Our additional analysis confirms that these gains mainly come from LLMs’ strong reasoning capabilities, rather than their external domain knowledge or summarization skills. LR-RecSys presents an effective approach to combine LLMs with traditional DNNs, two of the most widely used ML models today. Specifically, we show that LLMs can improve both the explainability and predictive performance of traditional DNNs through their reasoning capability. Beyond improving recommender systems, our findings emphasize the value of combining contrastive explanations for understanding consumer preferences and guiding managerial strategies for online platforms. These explanations provide actionable insights for consumers, sellers, and platforms, helping to build trust, optimize product offerings, and inform targeting strategies. |
| Date: | 2025–02 |
| URL: | https://d.repec.org/n?u=RePEc:ecl:stabus:4234 |
| By: | Ma, Ji (The University of Texas at Austin) |
| Abstract: | Large language models (LLMs) increasingly serve as human-like decision-making agents in social science and applied settings. These LLM-agents are typically assigned human-like characters and placed in real-life contexts. However, how these characters and contexts shape an LLM's behavior remains underexplored. This study proposes and tests methods for probing, quantifying, and modifying an LLM's internal representations in a Dictator Game -- a classic behavioral experiment on fairness and prosocial behavior. We extract ``vectors of variable variations'' (e.g., ``male'' to ``female'') from the LLM's internal state. Manipulating these vectors during the model's inference can substantially alter how those variables relate to the model's decision-making. This approach offers a principled way to study and regulate how social concepts can be encoded and engineered within transformer-based models, with implications for alignment, debiasing, and designing AI agents for social simulations in both academic and commercial applications. |
| Date: | 2025–04–18 |
| URL: | https://d.repec.org/n?u=RePEc:osf:osfxxx:8p7wg_v1 |
| By: | Alejandro Lopez-Lira; Yuehua Tang; Mingyin Zhu |
| Abstract: | Large language models (LLMs) cannot be trusted for economic forecasts during periods covered by their training data. We provide the first systematic evaluation of LLMs' memorization of economic and financial data, including major economic indicators, news headlines, stock returns, and conference calls. Our findings show that LLMs can perfectly recall the exact numerical values of key economic variables from before their knowledge cutoff dates. This recall appears to be randomly distributed across different dates and data types. This selective perfect memory creates a fundamental issue -- when testing forecasting capabilities before their knowledge cutoff dates, we cannot distinguish whether LLMs are forecasting or simply accessing memorized data. Explicit instructions to respect historical data boundaries fail to prevent LLMs from achieving recall-level accuracy in forecasting tasks. Further, LLMs seem exceptional at reconstructing masked entities from minimal contextual clues, suggesting that masking provides inadequate protection against motivated reasoning. Our findings raise concerns about using LLMs to forecast historical data or backtest trading strategies, as their apparent predictive success may merely reflect memorization rather than genuine economic insight. Any application where future knowledge would change LLMs' outputs can be affected by memorization. In contrast, consistent with the absence of data contamination, LLMs cannot recall data after their knowledge cutoff date. |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2504.14765 |
| By: | Ben Weidmann; Yixian Xu; David J. Deming |
| Abstract: | We show that leadership skill with artificially intelligent (AI) agents predicts leadership skill with human groups. In a large pre-registered lab experiment, human leaders worked with AI agents to solve problems. Their performance on this “AI leadership test” was strongly correlated (ρ=0.81) with their causal impact as leaders of human teams, which we estimate by repeatedly randomly assigning leaders to groups of human followers and measuring team performance. Successful leaders of both humans and AI agents ask more questions and engage in more conversational turn-taking; they score higher on measures of social intelligence, fluid intelligence, and decision-making skill, but do not differ in gender, age, ethnicity or education. Our findings indicate that AI agents can be effective proxies for human participants in social experiments, which greatly simplifies the measurement of leadership and teamwork skills. |
| JEL: | J24 M54 O30 |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:33662 |
| By: | James J. Choi; Dong Huang; Zhishu Yang; Qi Zhang |
| Abstract: | How good is AI at persuading humans to perform costly actions? We study calls made to get delinquent consumer borrowers to repay. Regression discontinuity and a randomized experiment reveal that AI is substantially less effective than human callers. Replacing AI with humans six days into delinquency closes much of the gap. But borrowers initially contacted by AI have repaid 1% less of the initial late payment one year later and are more likely to miss subsequent payments than borrowers who were always called by humans. AI’s lesser ability to extract promises that feel binding may contribute to the performance gap. |
| JEL: | D14 G4 G51 J24 |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:33669 |
| By: | Vaccaro, Michelle Anna; Caosun, Michael; Ju, Harang; Aral, Sinan; Curhan, Jared |
| Abstract: | Despite the rapid proliferation of artificial intelligence (AI) negotiation agents, there has been limited integration of computer science research and established negotiation theory to develop new theories of AI negotiation. To bridge this gap, we conducted an International AI Negotiations Competition in which participants with a diverse range of experience in negotiation, artificial intelligence (AI), and computer science (CS) iteratively designed and refined prompts for large language model (LLM) negotiation agents. We then facilitated ~120, 000 negotiations between these agents across multiple scenarios with diverse characteristics and objectives. Our findings revealed that fundamental principles from established human-human negotiation theory remain crucial in AI-AI negotiations. Specifically, agents exhibiting high “warmth” reached deals more frequently, created more value in integrative settings, claimed more value in distributive settings, and fostered higher subjective value. In addition, highly “dominant” agents performed better at claiming value in both distributive and integrative settings. These results align with classic negotiation theory emphasizing the importance of relationship-building, assertiveness, and preparation. However, our analysis also revealed unique dynamics in AI-AI negotiations not fully explained by negotiation theory, particularly regarding the effectiveness of AI-specific technical strategies like chain-of-thought reasoning and prompt injection. In fact, the agent that demonstrated the best combined performance across our key metrics—value creation, value claiming, and counterpart subjective value—implemented a sophisticated approach that blended traditional negotiation preparation frameworks with AI-specific technical methods. Together, these results suggest the importance of establishing a new theory of AI negotiations which integrates established negotiation theory with AI-specific negotiating strategies to optimize agent performance. Our research suggests this new theory must account for the unique characteristics of autonomous agents and establish the conditions under which traditional negotiation theory applies in automated settings. |
| Date: | 2025–03–11 |
| URL: | https://d.repec.org/n?u=RePEc:osf:osfxxx:b3v9e_v1 |
| By: | Keppeler, Florian; Borchert, Jana; Pedersen, Mogens Jin; Nielsen, Vibeke Lehmann |
| Abstract: | Artificial Intelligence (AI) applications are transforming public sector decision-making. However, most research conceptualizes AI as a form of specialized algorithmic decision support tool. In contrast, this study introduces the concept of human-AI ensembles, where humans and AI tackle the same tasks together, rather than specializing in certain parts. We argue that this is particularly relevant for many public sector decisions, where neither human nor AI-based decision-making has a clear advantage over the other in terms of legitimacy, efficacy, or legality. We illustrate this design theory within access to public employment, focusing on two key areas: (a) the potential of ensembling human and AI to reduce biases and (b) the inclinations of public managers to use AI advice. Study 1 presents evidence from the assessment of real-life job candidates (n = 2, 000) at the intersection of gender and ethnicity by public managers compared to AI. The results indicate that ensembled decision- making may alleviate ethnic biases. Study 2 examines how receptive public managers are to AI advice. Results from a pre-registered survey experiment involving managers (n = 538 with 4 observations each) show that decision-makers, when reminded of the unlawfulness of hiring discrimination, prioritize AI advice over human advice. |
| Date: | 2025–03–17 |
| URL: | https://d.repec.org/n?u=RePEc:osf:osfxxx:2yf6r_v2 |
| By: | Dirk Leffrang (Paderborn University); Nina Passlack (University of Bamberg); Oliver Müller (Paderborn University); Oliver Posegga (University of Bamberg) |
| Abstract: | Despite the increasing proliferation of Generative Artificial Intelligence (GenAI), systems like large language models (LLMs) can sometimes present misleading or false information as true – a problem known as "hallucinations." As GenAI systems become more widespread and accessible to the general public, understanding how AI literacy influences advice-taking from imperfect GenAI advice is crucial. Drawing on the correspondence bias, we study how individuals with varying AI literacy levels react to GenAI providing bad advice. Gathering empirical evidence through an online programming experiment, we find that AI-literate individuals take less advice, especially while receiving bad advice, but not exclusively. We outline how correspondence bias can explain these variations, reconciling mixed findings of prior studies on AI literacy. Our research thus contributes a holistic perspective on the beneficial and detrimental mistrust through AI literacy to education, integration, and evaluation programs of AI, highlighting the dangers of naive evaluation strategies. |
| Keywords: | AI literacy, artificial intelligence, AI education, algorithm aversion |
| JEL: | D83 D91 O33 C91 |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:pdn:dispap:136 |
| By: | Dirk Leffrang (Paderborn University); Oliver Müller (Paderborn University) |
| Abstract: | Despite the impressive capabilities of Artificial Intelligence (AI), concerns about its environmental impact continue to grow. However, organizations rarely implement sustainable AI practices in real-world settings. Prior research has predominantly focused on promoting sustainability-related information for non-AI products or exploring technical approaches for AI applications. This paper identifies performance uncertainty as a key factor distinguishing AI from prior technologies. Drawing on goal-setting theory, we investigate the impact of sustainability information and an unmet performance goal on AI retraining decisions. We conducted three incentivized online between-subjects experiments with 343 individuals with data science experience. Our results indicate that visualizing sustainability information increases the likelihood of sustainable AI choices. However, presenting an unmet performance goal decreases the likelihood and offsets the beneficial impact of sustainability information. These findings support sustainability initiatives, AI evaluations, and future research by emphasizing that while sustainability matters, achieving it requires appropriate performance goals for AI. |
| Keywords: | sustainability, artificial intelligence, goal-setting, laboratory experiment |
| JEL: | Q01 D91 Q55 C91 |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:pdn:dispap:135 |
| By: | Kumar Rishabh (University of Basel); Roxana Mihet (Swiss Finance Institute - HEC Lausanne); Julian Jang-Jaccard (Swiss Federal Office for Defence Procurement) |
| Abstract: | Does AI make firms vulnerable or resilient to cyber risk? To answer this, we develop a novel measure identifying AI-intensive U.S. public firms using publicly available patents and business-description data. While cyber threats typically suppress innovation, AI-intensive firms neutralize this effect. This protective effect strengthens with greater AI experience. Moreover, firms combining AI innovation and implementation exhibit a stronger buffer protecting their innovation and financial outcomes under cyber stress, whereas firms merely implementing AI without internal innovation gain no such resilience. Our results emphasize internal AI innovation as fundamental in enabling firms to effectively withstand cyber threats. |
| Keywords: | Cyberrisk, artificial intelligence, innovation, resilience, economics of AI, economics of cybercrime |
| JEL: | D8 O3 O4 G3 L1 L2 M1 |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:chf:rpseri:rp2539 |
| By: | Egana-delSol, Pablo (Universidad Adolfo Ibañez); Bravo-Ortega, Claudio (Universidad Adolfo Ibañez) |
| Abstract: | This study examines the implications of artificial intelligence (AI) on employment, wages, and inequality in Latin America and the Caribbean (LAC). The paper identifies tasks and occupations most exposed to AI using comprehensive individual-level data alongside AI exposure indices. Unlike traditional automation, AI exposure correlates positively with higher education levels, ICT, and STEM skills. Notably, younger workers and women with high-level ICT and managerial skills face increased AI exposure, underscoring unique opportunities. A comparison of LAC with the OECD countries reveals greater impacts of AI in the former, with physical and customer-facing tasks showing divergent correlations to AI exposure. The findings indicate that while AI contributes to employment growth at the top and bottom of wage quintiles, its wage impact strongly depends on the movement of workers from the middle class to below the wage mean of the high-level quintile of wages, hence decreasing the average income of the top quintile. |
| Keywords: | artificial intelligence, automation, labor market, developing economies, AI exposure, inequality, non cognitive skills, cognitive skills |
| JEL: | J23 J24 J31 |
| Date: | 2025–02 |
| URL: | https://d.repec.org/n?u=RePEc:iza:izadps:dp17746 |
| By: | Joshua S. Gans |
| Abstract: | This paper examines the role of artificial intelligence as a strategist in organizational decision-making by extending van den Steen's formal theory of strategy. A mathematical model is developed comparing AI and human strategists across different decision contexts, focusing on how each type generates confidence, achieves agreement, and implements decisions through control versus influence. The analysis presumes that AI excels in data-rich domains but faces credibility challenges in judgment-intensive contexts, creating a counterintuitive result where AI requires less formal authority precisely where it demonstrates superior analytical capabilities. The paper identifies distinct mechanisms through which strategic value is created: direct decision quality improvement and enhanced coordination. The authors propose domain-contingent approaches to AI integration, including differentiated authority systems across decision types and progressive control models that evolve as AI demonstrates effectiveness. These findings contribute to strategy theory while providing practical guidance for organizations seeking to effectively integrate AI into strategic processes, highlighting that organizations must adapt to their strategists' capabilities as much as strategists must match their organizations. |
| JEL: | D81 L20 O32 |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:33650 |
| By: | Joshua S. Gans |
| Abstract: | This paper examines how the introduction of artificial intelligence (AI), particularly generative and large language models capable of interpolating precisely between known data points, reshapes scientists' incentives for pursuing novel versus incremental research. Extending the theoretical framework of Carnehl and Schneider (2025), we analyse how decision-makers leverage AI to improve precision within well-defined knowledge domains. We identify conditions under which the availability of AI tools encourages scientists to choose more socially valuable, highly novel research projects, contrasting sharply with traditional patterns of incremental knowledge growth. Our model demonstrates a critical complementarity: scientists strategically align their research novelty choices to maximise the domain where AI can reliably inform decision-making. This dynamic fundamentally transforms the evolution of scientific knowledge, leading either to systematic “stepping stone” expansions or endogenous research cycles of strategic knowledge deepening. We discuss the broader implications for science policy, highlighting how sufficiently capable AI tools could mitigate traditional inefficiencies in scientific innovation, aligning private research incentives closely with the social optimum. |
| JEL: | D82 O30 O34 |
| Date: | 2025–03 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:33566 |
| By: | Roxana Mihet (Swiss Finance Institute - HEC Lausanne); Kumar Rishabh (University of Lausanne - Faculty of Business and Economics (HEC Lausanne); University of Basel, Faculty of Business and Economics); Orlando Gomes (Lisbon Polytechnic Institute - Lisbon Accounting and Business School) |
| Abstract: | Artificial intelligence (AI) is transforming productivity and market structure, yet the roots of firm dominance in the modern economy remain unclear. Is market power driven by AI capabilities, access to data, or the interaction between them? We develop a dynamic model in which firms learn from data using AI, but face informational entropy: without sufficient AI, raw data has diminishing or even negative returns. The model predicts two key dynamics: (1) improvements in AI disproportionately benefit data-rich firms, reinforcing concentration; and (2) access to processed data substitutes for compute, allowing low-AI firms to compete and reducing concentration. We test these predictions using novel data from 2000–2023 and two exogenous shocks—the 2006 launch of Amazon Web Services (AWS) and the 2017 introduction of transformer-based architectures. The results confirm both mechanisms: compute access enhances the advantage of data-intensive firms, while access to processed data closes the performance gap between AI leaders and laggards. Our findings suggest that regulating data usability—not just AI models—is essential to preserving competition in the modern economy. |
| JEL: | L13 L41 O33 D83 E22 L86 |
| Date: | 2025–03 |
| URL: | https://d.repec.org/n?u=RePEc:chf:rpseri:rp2537 |
| By: | Xavier Gabaix; Ralph S. J. Koijen; Robert J. Richmond; Motohiro Yogo |
| Abstract: | Firm characteristics, based on accounting and financial market data, are commonly used to represent firms in economics and finance. However, investors collectively use a much richer information set beyond firm characteristics, including sources of information that are not readily available to researchers. We show theoretically that portfolio holdings contain all relevant information for asset pricing, which can be recovered under empirically realistic conditions. Such guarantees do not exist for other data sources, such as accounting or text data. We build on recent advances in artificial intelligence (AI) and machine learning (ML) that represent unstructured data (e.g., text, audio, and images) by high-dimensional latent vectors called embeddings. Just as word embeddings leverage the document structure to represent words, asset embeddings leverage portfolio holdings to represent firms. Thus, this paper is a bridge from recent advances in AI and ML to economics and finance. We explore various methods to estimate asset embeddings, including recommender systems, shallow neural network models such as Word2Vec, and transformer models such as BERT. We evaluate the performance of these models on three benchmarks that can be evaluated using a single quarter of data: predicting relative valuations, explaining the comovement of stock returns, and predicting institutional portfolio decisions. We also estimate investor embeddings (i.e., representations of investors and their strategies), which are useful for investor classification, performance evaluation, and detecting crowded trades. We discuss other applications of asset embeddings, including generative portfolios, risk management, and stress testing. Finally, we develop a framework to give an economic narrative to a group of similar firms, by applying large language models to firm-level text data. |
| JEL: | C53 G12 G23 |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:33651 |
| By: | Jingru Wang; Wen Ding; Xiaotong Zhu |
| Abstract: | In the modern financial sector, the exponential growth of data has made efficient and accurate financial data analysis increasingly crucial. Traditional methods, such as statistical analysis and rule-based systems, often struggle to process and derive meaningful insights from complex financial information effectively. These conventional approaches face inherent limitations in handling unstructured data, capturing intricate market patterns, and adapting to rapidly evolving financial contexts, resulting in reduced accuracy and delayed decision-making processes. To address these challenges, this paper presents an intelligent financial data analysis system that integrates Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) technology. Our system incorporates three key components: a specialized preprocessing module for financial data standardization, an efficient vector-based storage and retrieval system, and a RAG-enhanced query processing module. Using the NASDAQ financial fundamentals dataset from 2010 to 2023, we conducted comprehensive experiments to evaluate system performance. Results demonstrate significant improvements across multiple metrics: the fully optimized configuration (gpt-3.5-turbo-1106+RAG) achieved 78.6% accuracy and 89.2% recall, surpassing the baseline model by 23 percentage points in accuracy while reducing response time by 34.8%. The system also showed enhanced efficiency in handling complex financial queries, though with a moderate increase in memory utilization. Our findings validate the effectiveness of integrating RAG technology with LLMs for financial analysis tasks and provide valuable insights for future developments in intelligent financial data processing systems. |
| Date: | 2025–03 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2504.06279 |
| By: | Guy Stephane Waffo Dzuyo (Forvis Mazars, LORIA - Laboratoire Lorrain de Recherche en Informatique et ses Applications - Inria - Institut National de Recherche en Informatique et en Automatique - UL - Université de Lorraine - CNRS - Centre National de la Recherche Scientifique, SYNALP - Natural Language Processing : representations, inference and semantics - LORIA - NLPKD - Department of Natural Language Processing & Knowledge Discovery - LORIA - Laboratoire Lorrain de Recherche en Informatique et ses Applications - Inria - Institut National de Recherche en Informatique et en Automatique - UL - Université de Lorraine - CNRS - Centre National de la Recherche Scientifique); Gaël Guibon (LORIA - Laboratoire Lorrain de Recherche en Informatique et ses Applications - Inria - Institut National de Recherche en Informatique et en Automatique - UL - Université de Lorraine - CNRS - Centre National de la Recherche Scientifique, LIPN - Laboratoire d'Informatique de Paris-Nord - CNRS - Centre National de la Recherche Scientifique - Université Sorbonne Paris Nord, SYNALP - Natural Language Processing : representations, inference and semantics - LORIA - NLPKD - Department of Natural Language Processing & Knowledge Discovery - LORIA - Laboratoire Lorrain de Recherche en Informatique et ses Applications - Inria - Institut National de Recherche en Informatique et en Automatique - UL - Université de Lorraine - CNRS - Centre National de la Recherche Scientifique); Christophe Cerisara (SYNALP - Natural Language Processing : representations, inference and semantics - LORIA - NLPKD - Department of Natural Language Processing & Knowledge Discovery - LORIA - Laboratoire Lorrain de Recherche en Informatique et ses Applications - Inria - Institut National de Recherche en Informatique et en Automatique - UL - Université de Lorraine - CNRS - Centre National de la Recherche Scientifique); Luis Belmar-Letelier (Forvis Mazars) |
| Abstract: | The identification of the financial characteristics of industry sectors has a large importance in accounting audit, allowing auditors to prioritize the most important area during audit. Existing company classification standards such as the Standard Industry Classification (SIC) code allow to map a company to a category based on its activity and products. In this paper, we explore the potential of machine learning algorithms and language models to analyze the relationship between those categories and companies' financial statements. We propose a supervised company classification methodology and analyze several types of representations for financial statements. Existing works address this task using solely numerical information in financial records. Our findings show that beyond numbers, textual information occurring in financial records can be leveraged by language models to match the performance of dedicated decision tree-based classifiers, while providing better explainability and more generic accounting representations. We think this work can serve as a preliminary work towards semi-automatic auditing. Models, code, and a preprocessed dataset are publicly available for further research at https://github.com/WaguyMz/hybrid company classification |
| Keywords: | Machine Learning, Industry Sectors, Large Language Models, LLM Applications, Audit, Financial Statement |
| Date: | 2025–02–25 |
| URL: | https://d.repec.org/n?u=RePEc:hal:journl:hal-05031499 |
| By: | Dimitrios Kanelis; Lars H. Kranzmann; Pierre L. Siklos |
| Abstract: | We analyze how financial stability concerns discussed during Federal Open Market Committee (FOMC) meetings influence the Federal Reserve’s monetary policy implementation and communication. Utilizing large language models (LLMs) to analyze FOMC minutes from 1993 to 2022, we measure both mandate-related and financial stability-related sentiment within a unified framework, enabling a nuanced examination of potential links between these two objectives. Our results indicate an increase in financial stability concerns following the Great Financial Crisis, particularly during periods of monetary tightening and the COVID-19 pandemic. Outside the zero lower bound (ZLB), heightened financial stability concerns are associated with a reduction in the federal funds rate, while within the ZLB, they correlate with a tightening of unconventional measures. Methodologically, we introduce a novel labeled dataset that supports a contextualized LLM interpretation of FOMC documents and apply explainable AI techniques to elucidate the model’s reasoning. |
| Keywords: | explainable artificial intelligence, financial stability, FOMC deliberations, monetary policy communication, natural language processing |
| JEL: | E44 E52 E58 |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:een:camaaa:2025-26 |
| By: | Vasant Dhar; Jo\~ao Sedoc |
| Abstract: | Long-term investing was previously seen as requiring human judgment. With the advent of generative artificial intelligence (AI) systems, automated systematic long-term investing is now feasible. In this paper, we present DBOT, a system whose goal is to reason about valuation like Aswath Damodaran, who is a unique expert in the investment arena in terms of having published thousands of valuations on companies in addition to his numerous writings on the topic, which provide ready training data for an AI system. DBOT can value any publicly traded company. DBOT can also be back-tested, making its behavior and performance amenable to scientific inquiry. We compare DBOT to its analytic parent, Damodaran, and highlight the research challenges involved in raising its current capability to that of Damodaran's. Finally, we examine the implications of DBOT-like AI agents for the financial industry, especially how they will impact the role of human analysts in valuation. |
| Date: | 2025–04 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2504.05639 |
| By: | Geiger, Felix; Kanelis, Dimitrios; Lieberknecht, Philipp; Sola, Diana |
| Abstract: | Central bank communication has become a crucial tool for steering the monetary policy stance and shaping the outlook of market participants. Traditionally, analyzing central bank communication required substantial human effort, expertise, and resources, making the process time-consuming. The recent introduction of artificial intelligence (AI) methods has streamlined and enhanced this analysis. While fine-tuned language models show promise, their reliance on large annotated datasets is a limitation that the use of large language models (LLMs) combined with prompt engineering overcomes. This paper introduces the Monetary-Intelligent Language Agent (MILA), a novel framework that leverages advanced prompt engineering techniques and LLMs to analyze and measure different semantic dimensions of monetary policy communication. MILA performs granular classifications of central bank statements conditional on the macroeconomic context. This approach enhances transparency, integrates expert knowledge, and ensures rigorous statistical calculations. For illustration, we apply MILA to the European Central Bank's (ECB) monetary policy statements to derive sentiment and hawkometer indicators. Our findings reveal changes in the ECB's communication tone over time, reflecting economic conditions and policy adaptions, and demonstrate MILA's effectiveness in providing nuanced insights into central bank communication. A model evaluation of MILA shows high accuracy, flexibility, and strong consistency of the results despite the stochastic nature of language models. |
| Keywords: | Central bank communication, monetary policy, sentiment analysis, artificial intelligence, large language models |
| JEL: | C45 E31 E44 E52 E58 |
| Date: | 2025 |
| URL: | https://d.repec.org/n?u=RePEc:zbw:bubtps:316448 |
| By: | Stella Jia; Abhishek Nagaraj |
| Abstract: | Large Language Models (LLMs) have demonstrated remarkable capabilities in text generation, but their performance may be influenced by the datasets on which they are trained, including potentially unauthorized or pirated content. We investigate the extent to which data access through pirated books influences LLM responses. We test the performance of leading foundation models (GPT, Claude, Llama, and Gemini) on a set of books that were and were not included in the Books3 dataset, which contains full-text pirated books and could be used for LLM training. We assess book-level performance using the “name cloze” word-prediction task. To examine the causal effect of Books3 inclusion we employ an instrumental variables strategy that exploits the pattern of book publication years in the Books3 dataset. In our sample of 12, 916 books, we find significant improvements in LLM name cloze accuracy on books available within the Books3 dataset compared to those not present in these data. These effects are more pronounced for less popular books as compared to more popular books and vary across leading models. These findings have crucial implications for the economics of digitization, copyright policy, and the design and training of AI systems. |
| JEL: | K24 O36 |
| Date: | 2025–03 |
| URL: | https://d.repec.org/n?u=RePEc:nbr:nberwo:33598 |