|
on Artificial Intelligence |
| By: | Francesca Molinari; Yiqi Liu |
| Abstract: | Algorithms are increasingly used to aid with high-stakes decision making. Yet, their predictive ability frequently exhibits systematic variation across population subgroups. To assess the trade-off between fairness and accuracy using finite data, we propose a debiased machine learning estimator for the fairness-accuracy frontier introduced by Liang, Lu, Mu, and Okumura (2024). We derive its asymptotic distribution and propose inference methods to test key hypotheses in the fairness literature, such as (i) whether excluding group identity from use in training the algorithm is optimal and (ii) whether there are less discriminatory alternatives to a given algorithm. In addition, we construct an estimator for the distance between a given algorithm and the fairest point on the frontier, and characterize its asymptotic distribution. Using Monte Carlo simulations, we evaluate the finite-sample performance of our inference methods. We apply our framework to re-evaluate algorithms used in hospital care management and show that our approach yields alternative algorithms that lie on the fairness-accuracy frontier, offering improvements along both dimensions. |
| Date: | 2025–07–01 |
| URL: | https://d.repec.org/n?u=RePEc:azt:cemmap:13/25 |
| By: | Sanchaita Hazra; Marta Serra-Garcia |
| Abstract: | LLMs are emerging as information sources that influence organizational knowledge, though trust in them varies. This paper combines data from a large-scale experiment and the World Values Survey (WVS) to examine the determinants of trust in LLMs. The experiment measures trust in LLM-generated answers to policy-relevant questions among over 2, 900 participants across 11 countries. Trust in the LLM is significantly lower in high-income countries-especially among individuals with right-leaning political views and lower educational attainment-compared to low- and middle-income countries. Using large-scale data on trust from the WVS, we show that patterns of trust in the LLM differ from those in generalized trust but closely align with trust in traditional information sources. These findings highlight that comparing trust in LLMs to other forms of societal trust can deepen our understanding of the potential societal impacts of AI. |
| Keywords: | information, generative AI, accuracy, trust, experiment |
| JEL: | D83 D91 C72 C91 |
| Date: | 2025 |
| URL: | https://d.repec.org/n?u=RePEc:ces:ceswps:_11954 |
| By: | Marie Obidzinski (Université Paris Panthéon Assas, CRED UR 7321, 75005 Paris, France); Yves Oytana (Université Marie et Louis Pasteur, CRESE UR3190, F-25000 Besançon, France) |
| Abstract: | We study the design of optimal liability sharing rules when the use of an AI prediction by a human user may cause external damage. To do so, we set up a game-theoretic model in which an AI manufacturer chooses the level of accuracy with which an AI is developed (which increases the reliability of its prediction) and the price at which it is distributed. The user then decides whether to buy the AI. The AI’s prediction gives a signal about the state of the world, while the user chooses her effort to discover the payoffs in each possible state of the world. The user may be susceptible to an automation bias that leads her to overestimate the algorithm’s accuracy (overestimation bias). In the absence of an automation bias, we find that full user liability is optimal. However, when the user is prone to an overestimation bias, increasing the share of liability borne by the AI manufacturer can be beneficial for two reasons. First, it reduces the rent that the AI manufacturer can extract by exploiting the user’s overestimation bias by underinvesting or overinvesting in the AI accuracy. Second, due to the nature of the interaction between algorithm accuracy and the user effort, the user may be incentivized to increase her (too low) judgment effort. |
| Keywords: | liability sharing, advisory algorithm, automation bias, prediction, judgment effort |
| JEL: | K13 |
| Date: | 2025–06 |
| URL: | https://d.repec.org/n?u=RePEc:crb:wpaper:2025-08 |
| By: | Xienan Cheng; Mustafa Dogan; Pinar Yildirim |
| Abstract: | This study investigates the effects of artificial intelligence (AI) adoption in organizations. We ask: (1) How should a principal optimally deploy limited AI resources to replace workers in a team? (2) In a sequential workflow, which workers face the highest risk of AI replacement? (3) How does substitution with AI affect both the replaced and non-replaced workers' wages? We develop a sequential team production model in which a principal can use peer monitoring -- where each worker observes the effort of their predecessor -- to discipline team members. The principal may replace some workers with AI agents, whose actions are not subject to moral hazard. Our analysis yields four key results. First, the optimal AI strategy involves the stochastic use of AI to replace workers. Second, the principal replaces workers at the beginning and at the end of the workflow, but does not replace the middle worker, since this worker is crucial for sustaining the flow of information obtained by peer monitoring. Third, the principal may choose not to fully exhaust the AI capacity at her discretion. Fourth, the optimal AI adoption increases average wages and reduces intra-team wage inequality. |
| Date: | 2025–06 |
| URL: | https://d.repec.org/n?u=RePEc:arx:papers:2506.12337 |
| By: | Leland D. Crane; Akhil Karra; Paul E. Soto |
| Abstract: | We evaluate the ability of large language models (LLMs) to estimate historical macroeconomic variables and data release dates. We find that LLMs have precise knowledge of some recent statistics, but performance degrades as we go farther back in history. We highlight two particularly important kinds of recall errors: mixing together first print data with subsequent revisions (i.e., smoothing across vintages) and mixing data for past and future reference periods (i.e., smoothing within vintages). We also find that LLMs can often recall individual data release dates accurately, but aggregating across series shows that on any given day the LLM is likely to believe it has data in hand which has not been released. Our results indicate that while LLMs have impressively accurate recall, their errors point to some limitations when used for historical analysis or to mimic real time forecasters. |
| Keywords: | Artificial intelligence; Forecasting; Large language models; Real-time data |
| JEL: | C53 C80 E37 |
| Date: | 2025–06–25 |
| URL: | https://d.repec.org/n?u=RePEc:fip:fedgfe:2025-44 |
| By: | Ryota IWAMOTO; Takunori ISHIHARA; Takanori IDA |
| Abstract: | This study empirically investigates the differences in risk preferences and loss aversion between humans and generative AI. We conduct a nationwide online survey of 4, 838 individuals and generate AI responses under identical conditions by using personas constructed from demographic attributes. The results show that in gain domains, both humans and the AI select risk-averse options and exhibit similar preference patterns. However, in loss domains, AI shows a stronger risk-loving tendency and responds more sharply to individual attributes such as gender, age, and income. We retrain the AI by fine-tuning it based on human choice data. After fine-tuning, the AI’s preference distribution moves closer to that of humans, with loss-related decisions showing the greatest improvement. Using Wasserstein distance, we also confirm that fine-tuning reduces the behavioral gap between AI and humans. |
| Keywords: | bias, bias, loss aversion, risk preference, generative AI, persona, fine-tuning, Wasserstein distance |
| JEL: | D91 C91 |
| Date: | 2025–07 |
| URL: | https://d.repec.org/n?u=RePEc:kue:epaper:e-25-006 |
| By: | Johannesson, Mikael Poul (NORCE Norwegian Research Centre); Arnesen, Sveinung |
| Abstract: | This paper examines the impact of Artificial Intelligence (AI) on citizens' expectations of street-level bureaucrats. Using survey experiments fielded in The Norwegian Citizen Panel, we assess how AI's role as a decision-support tool affects the importance citizens place on various bureaucratic traits. Our findings suggest that when street-level bureaucrats use AI, citizens will sometimes want bureaucrats that are more similar to themselves, and also tend to consider bureaucrats' technical expertise as less important. This suggests that as AI takes on more of the technical judgments that is part of bureaucratic decision-making, citizens place greater importance on the human elements of bureaucracy, such as shared experiences and empathy. This research highlights the growing need to understand how the use of AI will shift citizens' expectations of public institutions. |
| Date: | 2025–06–11 |
| URL: | https://d.repec.org/n?u=RePEc:osf:socarx:8r36s_v1 |
| By: | Nils H. Lehr (International Monetary Fund); Pascual Restrepo (Yale University) |
| Abstract: | Leading AI firms claim to prioritize social welfare. How should firms with a social mandate price and deploy AI? We derive pricing formulas that depart from profit maximization by incorporating incentives to enhance welfare and reduce labor disruptions. Using US data, we evaluate several scenarios. A welfarist firm that values both profit and welfare should price closer to marginal cost, as efficiency gains outweigh distributional concerns. A conservative firm focused on labormarket stability should price above the profit-maximizing level in the short run, especially when its AI may displace low-income workers. Overall, socially minded firms face a trade-off between expanding access to AI and the resulting loss in profits and labor market risks. |
| Date: | 2025–05–20 |
| URL: | https://d.repec.org/n?u=RePEc:cwl:cwldpp:2445 |