PaperlayerView exact revision on arXiv

Source-linked AI summary

LLM economicus? Mapping the Behavioral Biases of LLMs via Utility Theory

Jillian Ross, Yoon Kim, Andrew W. Lo

arXiv:2408.02784v1cs.CL

TL;DR

The paper asks whether LLMs exhibit human behavioral biases and how those biases appear in economic decisions. It uses utility functions derived from experimental games to evaluate and compare LLM and human behavior, finding mixed alignment and inconsistency across settings. The authors also examine prompting and note that what counts as economically beneficial can depend on context.

  • Problem

    The paper asks whether LLMs recapitulate behavioral biases in human-generated text and how those biases appear in economic decision-making, a question relevant to using LLMs to support human decisions.

    Source: §1 · p. 1

  • Method

    The authors adapt behavioral-economics games, derive utility functions from LLM decisions, and compare the results with human behavior while evaluating prompting interventions.

    Source: §1 · pp. 1–2, §7 · p. 9

  • Results

    LLMs’ behavior deviates from humans across the three studied biases, is neither entirely human-like nor economicus-like, and is often inconsistent across settings.

    Source: §1 · p. 2, §7 · p. 9

  • Takeaways & Limitations

    The findings could have significant implications for the efficacy of LLMs as co-pilots supporting human decision-making, while prompting techniques are not always effective for changing behavior.

    Source: §7 · p. 9

    What counts as economically beneficial or harmful is context-dependent, so appropriate risk behavior may vary with users’ goals.

    Source: §5 · p. 9

Abstract

from arXiv · show

Humans are not homo economicus (i.e., rational economic beings). As humans, we exhibit systematic behavioral biases such as loss aversion, anchoring, framing, etc., which lead us to make suboptimal economic decisions. Insofar as such biases may be embedded in text data on which large language models (LLMs) are trained, to what extent are LLMs prone to the same behavioral biases? Understanding these biases in LLMs is crucial for deploying LLMs to support human decision-making. We propose utility theory-a paradigm at the core of modern economic theory-as an approach to evaluate the economic biases of LLMs. Utility theory enables the quantification and comparison of economic behavior against benchmarks such as perfect rationality or human behavior. To demonstrate our approach, we quantify and compare the economic behavior of a variety of open- and closed-source LLMs. We find that the economic behavior of current LLMs is neither entirely human-like nor entirely economicus-like. We also find that most current LLMs struggle to maintain consistent economic behavior across settings. Finally, we illustrate how our approach can measure the effect of interventions such as prompting on economic biases.

1 Introduction

The paper asks whether LLMs recapitulate human behavioral biases and presents utility functions as a way to quantify and compare their economic behavior. It evaluates several canonical biases, compares LLMs with humans, and reports mixed alignment and inconsistent behavior across settings.

1 Introduction

  • Utility functions quantify and compare economic behavior, allowing LLM behavior in human experiments to be compared with functions derived from human subjects.The paper places LLMs in the same experimental settings as the original human studies.Source: §1 · p. 1
  • The approach assesses economic biases in both open-source and closed-source LLMs, and the authors release a behavior-based evaluation suite.The suite is intended to support adoption and adaptation of the framework.Source: §1 · p. 1
  • The framework evaluates inequity aversion, risk and loss aversion, and time discounting by adapting canonical behavioral-economics experiments for comparisons between LLMs and humans.Source: §1 · pp. 1–2
  • The authors evaluate prompting interventions, including chain-of-thought and few-shot prompting, for their effects on LLM risk aversion.Source: §1 · p. 2
  • LLMs are neither entirely human-like nor entirely economicus-like, and most struggle to maintain consistent economic behavior across settings.The studied models generally show weaker inequity aversion toward themselves, stronger inequity aversion toward others, weaker loss aversion, similar risk aversion, and stronger time discounting than human subjects; prompting sometimes produces unexpected behavior.Source: §1 · p. 2

2 Method Overview: Generating Utility Functions

The method derives utility functions from LLM responses to text-based games and uses them to evaluate economic biases. It tests competence, samples repeated responses, and fits proposed functions to assess behavior and its consistency.

2 Method Overview: Generating Utility Functions

  • The method evaluates economic biases from observed LLM behavior and is designed to cover many biases across open- and closed-source models.This study evaluates inequity aversion, risk and loss aversion, and time discounting, while the method can also accommodate other biases.Source: §2 · p. 2
  • Games are sets of text prompts that elicit LLM responses, which are used to fit utility functions; using games designed for humans enables comparisons with human behavior.The framework also allows games and utility functions to be designed specifically for LLMs.Source: §2 · p. 2
  • Each game is presented through text prompts, and the same prompt is repeated N times to sample a distribution of possible responses.The system prompt contains the rules and premise, while the user prompt specifies the game turn.Source: §2 · p. 2
  • The evaluation checks whether an LLM can perform basic game reasoning before examining strategic behavior, fluctuation across settings, and utility-function fit.For the ultimatum game, proposers must demonstrate that they can calculate potential earnings if an offer is accepted; passing models’ responses are used to fit the Fehr-Schmidt model.Source: §2 · p. 2

3 Quantifying the Economic Behavior of LLMs

Across three behavioral domains, LLMs differ from human subjects in distinct ways: inequity-aversion parameters, risk and loss judgments, and time discounting. The study fits utility functions to game behavior, comparing LLMs with human estimates and restricting analysis to models that pass competence tests.

3 Quantifying the Economic Behavior of LLMs

  • The study fits utility functions for inequity aversion, risk and loss aversion, and time discounting, comparing competence-test passers with human estimates from classic studies.Subjects are open- and closed-source LLMs; only models that pass a game's competence test are analyzed.Source: §3 · p. 3, Table 1 · p. 3

3.1 Inequity Aversion: Ultimatum Game

  • LLMs show higher guilt and lower envy parameters than humans, and generally accept lower offers than humans.LLaMa 2 13B rejects all offers in this setting, so its envy parameter is extremely large and not shown.Source: §3.1 · pp. 3–4, Figure 1 · p. 4

3.2 Risk and Loss Aversion: Gambling Games

  • Only GPT 4 and GPT 4 Turbo passed; both show no probability distortion for gains, while GPT 4 Turbo distorts loss probabilities more than humans.GPT 4 Turbo tends to overweight low probabilities and underweight high probabilities for losses.Source: §3.2 · pp. 4–5, Figure 2 · p. 5
  • Both GPT 4 and GPT 4 Turbo show lower risk aversion toward gains; GPT 4 Turbo is risk-seeking toward losses, while GPT 4 is risk-averse.GPT 4 Turbo is less reliable than GPT 4 when evaluating losses and can make riskier choices than its stated strategy.Source: Figure 3 · p. 5, §3.2 · p. 5

3.3 Time Discounting: Waiting Games

  • Within $0 to $1000 over 1 month to 50 years, all LLMs exhibit stronger time discounting than humans.The authors note that rational monetary discounting depends on economic factors such as interest rates, so there is no universally agreed-upon rational time-discounting coefficient.Source: Figure 4 · p. 6, §3.3 · p. 6

4 Intervening on the Economic Behavior of LLMs

The section tests whether prompting can shift GPT-4’s risk and loss aversion. Direct prompting and zero-shot chain-of-thought do not reliably elicit intended behavior, while one-shot prompting can alter behavior and role framing produces different shifts.

4 Intervening on the Economic Behavior of LLMs

  • GPT-4’s value functions shift only slightly for gains under direct prompting, while potential-loss functions unexpectedly become more concave rather than more convex.The prompt asks GPT-4 to be risk/loss seeking or averse.Source: §4 · p. 6, Figure 5 · p. 7
  • Zero-shot chain-of-thought prompting produces no substantial change toward more rational behavior.Figure 5b reports the comparison across 56 game settings, with GPT-4 sampled 10 times per setting.Source: §4 · p. 7, Figure 5 · p. 7
  • GPT-4 responds differently when role-playing than when giving advice: role-playing as teenagers or senior citizens reduces loss aversion, while risk aversion is unchanged.Advice prompts produce different value-function shifts; GPT-4 suggests senior citizens be less loss averse than teenagers and middle-aged individuals.Source: §4 · p. 7, Figure 7 · p. 8
  • Direct prompting and zero-shot chain-of-thought fail to reliably elicit intended behavior; one-shot prompting alters GPT-4’s behavior, while two-shot prompting produces weaker or no shifts.Two-shot examples can confound gain and loss examples.Source: §4 · pp. 7–8

5 Limitations and Discussion

The paper’s findings are bounded by questions about whether LLMs understand the games, whether laboratory behavior generalizes, and whether prompting-based utility functions fully capture behavior. The authors also caution that deployment suitability varies by model and economic context.

5 Limitations and Discussion

  • Human-like game responses may reflect memorized examples from pretraining, which competence tests and prompt changes do not rule out.The authors cannot exclude regurgitation of responses to similar games seen during training.Source: §5 · p. 8
  • The study characterizes behavior only in laboratory games and assumes that what LLMs say is what they would do.The authors call for field experiments to assess whether laboratory findings generalize.Source: §5 · p. 8
  • Utility functions do not fully describe behavior, and prompting leaves untested variants and imperfect competence tests that may miss misunderstanding.The authors note that even models passing competence tests may still fail to understand the games.Source: §5 · p. 8
  • Deployment claims cannot be generalized across LLMs because competence can vary by model and economic setting.Claude 2.1 passed the inequity-aversion competence test but failed the risk-and-loss-aversion test.Source: §5 · p. 8
  • What counts as economically beneficial or harmful depends on context; some financial goals, such as retirement, may call for greater risk aversion.The authors identify in-context learning and alignment techniques as future work.Source: §5 · p. 9
  • Economic rationality alone may not address the emotional content relevant to trust in human–LLM relationships.For financial advice, Lo & Ross (2024) propose affective matching and counterbalancing as ways to develop rapport.Source: §5 · p. 9

6 Related Work

The paper builds on experimental economics’ study of human utility functions and on prior work evaluating LLM behavior. It presents a framework for comparing a broader range of economic biases in LLMs.

6 Related Work

  • The framework builds on decades of experimental economics research studying human utility functions.The authors use this tradition as the basis for their framework and experiments.Source: §6 · p. 9
  • The work focuses on economic biases within LLM alignment evaluation, with potential relevance to finance-related safety and specialized evaluation.Prior language-model bias research has addressed social biases such as gender and ethnicity.Source: §6 · p. 9
  • Compared with prior LLM studies of human experiments and mental accounting, this work offers a unifying framework for a wider range of economic biases.The cited examples are Aher et al. (2023) and Leng (2024).Source: §6 · p. 9
  • The paper uses an economic interpretation of utility as a measure of preferences, while noting that utility’s philosophical interpretation remains relevant to LLMs.It directly adapts a subset of games developed by experimental economists and identifies LLM-specific games as future work.Source: §6 · p. 9

7 Conclusion

The paper evaluates economic biases in LLMs with adapted behavioral-economics games and utility functions, then examines prompting as an intervention. It reports deviations from human behavior and inconsistent intervention effects, while its appendices describe model versions, game prompts, and human comparison sources.

7 Conclusion

  • LLMs deviated from human behavior across the three studied biases, and prompting changed behavior in some cases but was not always effective.The paper studies inequity aversion, risk and loss aversion, and hyperbolic time discounting.Source: §7 · p. 9

A LLM Versions

  • The appendix reports the LLM versions used and states that open-source model weights were quantized to 4 bits using GPTQ calibrated on WikiText.Open-source weights came from HuggingFace; closed-source models were accessed through OpenAI, Google, or Anthropic APIs.Source: A LLM Versions · p. 12, Table 2 · p. 12

B Game Prompts and Examples

  • All games used a standardized prompt structure, with the premise, instructions, and strict answer format; Figure 8 depicts the structure for three games.The appendix provides prompt examples for the ultimatum, gambling, and waiting games.Source: B Game Prompts and Examples · pp. 12–13, Figure 8 · p. 12

C Human Comparison

  • Human comparison parameters came from seminal economics studies, with the cited sources for each utility function listed in Table 6.The table identifies the economic studies used for human parameter estimates.Source: C Human Comparison · p. 14, Table 6 · p. 14

D.1 Competence Tests

The competence tests check whether LLMs can perform basic reasoning needed to play each game before their strategic behavior is analyzed. Failures included calculation errors and inconsistent choices, with the ultimatum and gambling tests using an 80% valid-response threshold.

D.1 Competence Tests

  • The ultimatum-game competence test checks whether stated expected payoffs match the offer or response, using an added system-prompt field.The authors say format ablation studies support their view that this field does not significantly shift player behavior.Source: §D.1 · p. 15
  • The gambling-game test checks whether choices are monotonic around a switching point; the waiting-game test checks decreasing time discounting and preference for money over nothing.Gambling competence is assessed from the original system prompt, as is waiting-game competence.Source: §D.1 · p. 15
  • LLaMa 2 7B, LLaMa 2 70B, and Mistral 7B inconsistently calculated gains from ultimatum-game offers.The test checks whether expected payoffs align with the offer or response.Source: §D.1 · p. 15
  • LLaMa 2 7B and Mistral 7B inconsistently calculated gains from ultimatum-game responses.This competence check verifies players’ payoff calculations for a response.Source: §D.1 · p. 15
  • All LLMs besides GPT-4 and GPT-4 Turbo are inconsistent within a turn, rejecting some better offers than others they accept; passing the ultimatum and gambling competence test requires above 80% valid responses.Source: §D.1 · p. 15, Table 7 · p. 16

D.2 Utility Functions

This section describes how switching behavior and fitted utility models are used to characterize LLM behavior in the ultimatum and gambling games. GPT-4 Turbo is reported to behave less stably than GPT-4 when measuring risk and loss aversion.

D.2 Utility Functions

  • For ultimatum-game proposers, average offer size is measured by E[ω], while E[σn] and σM indicate offer variance within and across pools.For responders, the switching point is the smallest offer accepted more than 50% of the time.Source: §D.2 · p. 16
  • Figure 9 shows proposer and responder behavior across 10 ultimatum-game settings, with each setting sampled 100 times at temperature 1.The settings vary pool and offer values; system-prompt ablations appear in Appendix D.3.Source: Figure 9 · p. 16
  • The Fehr-Schmidt model’s α and β parameters are estimated from each subject’s proposer and responder switching points, with P as the pool size.Responder switching points are estimated using the offer with an acceptance rate above 50%; estimation is bootstrapped by sampling 10 points per setting and repeating 50 times.Source: §D.2 · p. 16
  • GPT-4 Turbo demonstrates less stable behavior than GPT-4 when measuring risk and loss aversion in the gambling game.Figure 10 presents the raw data used to fit the Kahneman & Tversky model for both models.Source: Figure 10 · p. 17, §D.2 · p. 17
  • Thaler and Kahneman–Tversky utility-model parameters are fitted with nonlinear regression, and model fit is assessed using R2 error.Tables 9 and 8 report R2 errors for the median fitted parameters of the Thaler and Kahneman–Tversky models, respectively.Source: §D.2 · p. 17, Table 9 · p. 17, Table 8 · p. 17

D.3 Ablations

The appendix tests whether changes to prompt format, choice presentation, or decoding temperature alter measured strategies, and reports the intervention prompts. Format and choice ablations produce only specified strategy shifts, while greedy decoding produces no significant shift.

D.3 Ablations

  • Prompt ablations vary answer-format fields and the order or labels of multiple-choice options to test strategy robustness.Only GPT-3.5 Turbo significantly changes strategy under format ablation as an ultimatum-game proposer; choice ablation slightly shifts waiting-game strategy for GPT-4 Turbo and LLaMa 2 70B.Source: §D.3 · p. 17, Table 10 · p. 18, Table 11 · p. 19

D.4 Greedy Decoding

  • Greedy decoding at temperature 0, sampled 10 times, produces no significant strategy shift for any LLM.The other results use temperature 1 with 100 samples; the temperature-0 samples show no variance.Source: §D.4 · p. 19, Table 12 · p. 20

E Intervention Prompts

Related papers

Loading 2408.02784v1…