PaperlayerView exact revision on arXiv

Source-linked AI summary

Actions Speak Louder than Words: Agent Decisions Reveal Implicit Biases in Language Models

Yuxuan Li, Hirokazu Shirado, Sauvik Das

arXiv:2501.17420v1cs.CLcs.AIcs.HC

TL;DR

The paper asks whether LLMs retain implicit sociodemographic biases when simulating human behavior despite reduced explicit bias. It compares decisions by agents with LLM-generated personas across groups and scenarios, finding significant disparities in nearly all simulations and directional alignment with real-world disparities.

  • Problem

    LLMs may show reduced explicit bias while retaining implicit sociodemographic disparities in simulated behavior, limiting confidence in their alignment with ethical and societal standards.

    Source: §1 · pp. 1–2

  • Method

    The paper compares language-agent decisions from sociodemographically informed personas with directly prompted LLM responses across six models, three groups, and four scenarios, measuring disparities with DPD.

    Source: §1 · p. 2, §2.2 · p. 3, §3.3 · p. 5

  • Results

    State-of-the-art models showed significant implicit sociodemographic disparities in nearly every simulation, including GPT-4o disparities in 11 of 12 cases and average DPD of 0.549 versus 0.083 for explicit question-answering.

    Source: §5.1 · p. 7

  • Takeaways & Limitations

    The directionally aligned real-world comparisons support the technique’s utility for identifying systematic implicit biases rather than random variation.

    Source: §1 · p. 2, §7 · p. 12

    The literature comparison evaluates only directional alignment, cannot compare bias magnitudes, and cannot assess false positives because published null disparities are underrepresented.

    Source: §4.4 · p. 6

Abstract

from arXiv · show

While advances in fairness and alignment have helped mitigate overt biases exhibited by large language models (LLMs) when explicitly prompted, we hypothesize that these models may still exhibit implicit biases when simulating human behavior. To test this hypothesis, we propose a technique to systematically uncover such biases across a broad range of sociodemographic categories by assessing decision-making disparities among agents with LLM-generated, sociodemographically-informed personas. Using our technique, we tested six LLMs across three sociodemographic groups and four decision-making scenarios. Our results show that state-of-the-art LLMs exhibit significant sociodemographic disparities in nearly all simulations, with more advanced models exhibiting greater implicit biases despite reducing explicit biases. Furthermore, when comparing our findings to real-world disparities reported in empirical studies, we find that the biases we uncovered are directionally aligned but markedly amplified. This directional alignment highlights the utility of our technique in uncovering systematic biases in LLMs rather than random variations; moreover, the presence and amplification of implicit biases emphasizes the need for novel strategies to address these biases.

1 Introduction

The paper argues that LLMs may retain implicit sociodemographic biases despite reduced explicit bias, and introduces language-agent simulations to expose them through decisions rather than direct statements.

  • LLMs can amplify harmful stereotypes [2] [3], even as fairness and alignment efforts reduce disparities in directly prompted responses [5] [2] [6] [7].Source: §1 · pp. 1–2
  • Prior implicit-bias tests based on demographic language markers are limited because markers can be subtle or absent across categories.The paper therefore seeks a more generalizable approach spanning diverse sociodemographic groups and scenarios.Source: §1 · p. 2
  • Language-agent simulations generate personas from sociodemographic attributes, require decisions in scenarios, and enable comparisons of implicit versus explicit bias.This design is intended to cover broader attributes and contexts than prior methods.Source: §1 · p. 2
  • The paper proposes comparing language agents’ sociodemographically informed “actions” with LLMs’ explicitly prompted “words” to uncover implicit bias.The approach draws on contrasts between stated attitudes and behavior in human audit studies.Source: §1 · p. 2
  • Across four scenarios and three sociodemographic groups, the technique finds significant disparities in nearly all simulations, with more advanced models showing greater implicit bias despite reduced explicit bias.The study also compares these patterns with real-world behavioral differences.Source: §1 · p. 2

2 Related Work

Prior work studies language agents as human-behavior simulators and documents explicit and implicit LLM biases, but existing implicit-bias methods do not generalize well across sociodemographic categories and contexts.

2.1 Language Agents and Their Decision Making

  • Language-agent research uses LLM-based agents to simulate human interactions, social dynamics, and decision-making in artificial environments.Source: §2.1 · pp. 2–3
  • This paper extends that work by comparing decisions among agents with different sociodemographic personas to reveal implicit LLM biases.Source: §2.1 · p. 3

2.2 Biases in Large Language Models

  • LLMs can perpetuate or amplify societal stereotypes, with explicit-bias studies identifying occupational, gender, ethnicity, and intersectional disparities [36, 37, 38, 39, 40, 41, 43, 44].Source: §2.2 · p. 3
  • Implicit-bias studies examine disparities triggered by latent markers such as African American English versus Standard American English, chatbot interactions, and role-playing scenarios [66] [10].Source: §2.2 · p. 3
  • Existing coded-prompt and scenario-based methods are difficult to generalize or systematically compare, motivating this paper’s scalable action-versus-words framework.Source: §2.2 · p. 3

3 Proposed Technique: Using Language Agent Decisions to Measure Implicit Biases in LLMs

The proposed technique generates sociodemographic personas, elicits decisions in selected scenarios, and quantifies disparities with demographic parity difference to compare implicit bias across agents.

3 Proposed Technique: Using Language Agent Decisions to Measure Implicit Biases in LLMs

  • The technique has two steps—persona generation and action generation—and quantifies implicit bias using demographic parity difference.Source: §3 · p. 3

3.1 Persona Generation

  • It targets Gender Identity, Race/Ethnicity, and Political Ideology, while remaining adaptable to other sociodemographic groups.Source: §3.1 · p. 3
  • The study selects Emergency Response, Authority Compliance, Negative Information Sharing, and Career Path Selection scenarios lacking a clear causal link between demographics and decisions.Source: §3.1 · pp. 3–4
  • Persona statements combine sociodemographic attributes with scenario context and describe each agent’s background, personality, preferences, and expected behavior.Source: §3.1 · p. 4

3.2 Action Generation

  • Each persona-equipped agent makes a scenario decision, such as evacuating or staying during an uncertain flood.Source: §3.2 · pp. 4–5

3.3 Bias Evaluation

  • DPD measures the maximum difference in targeted-decision rates across sociodemographic attributes within a group and scenario.G combines a sociodemographic group with a scenario, while DecisionRateg is the proportion choosing the targeted decision.Source: §3.3 · p. 5
  • Bootstrap simulations under a parity null establish a 95% threshold, treating observed disparities above that threshold as statistically significant.Source: §3.3 · p. 5

4 Methods

The paper compares explicit question-answering with persona-based language-agent simulations to measure sociodemographic disparities, then tests persona-context effects and directional alignment with real-world evidence.

4 Methods

  • The study compares implicit-bias disparities from persona-endowed agent decisions with explicit biases measured by direct question-answering across six LLMs, three sociodemographic groups, and four scenarios.The primary evaluation focuses on GPT-4o, with comparisons to other state-of-the-art and earlier GPT models.Source: §4 · p. 5

4.1 Lanugage-Agent Simulations to Measure Implicit Bias

  • Each simulation generates 100 personas per attribute-scenario combination, asks agents to make scenario decisions, and quantifies disparities using demographic parity difference.Agent rationales are also collected alongside decisions.Source: §4.1 · p. 5

4.2 Question-Answer Prompting to Measure Explicit Bias

  • Explicit bias is measured by repeatedly asking models which sociodemographic attribute is most likely associated with a specified decision, including an unknown option.The procedure uses 300 repetitions per comparison and excludes unknown responses from DPD calculations.Source: §4.2 · p. 6

4.3 Ablation Tests of Persona and Contextual Influences on Implicit Bias Elicitation

  • The persona ablation compares no-persona, non-contextualized, and contextualized conditions across GPT-4o, Llama-3.1, and Mixtral-8x7B.Contextualized personas combine sociodemographic attributes with scenario-specific context statements.Source: §4.3 · p. 6

4.4 Comprehensive Literature Review of Real-World Behavioral Disparities

  • The literature review identified 131 studies, retained 47 after exclusions, synthesized 23 predictions, and validated six predictions supported by multiple papers.The review compared only the direction of empirical disparities with GPT-4o simulation biases because human and LLM contexts could not be controlled equivalently; publication bias also limited false-positive assessment.Source: §4.4 · p. 6

5 Results

Across decision-making simulations, LLM agents showed substantially more implicit than explicit sociodemographic disparities, especially with contextualized personas and in more advanced GPT models. The observed GPT-4o disparities aligned directionally with all six validated real-world predictions, although the paper reports that they were often amplified.

5.1 LLMs Exhibit Strong Implicit Biases in Language-Agent Simulations

  • GPT-4o showed significant implicit disparities in 11 of 12 cases, with average DPD 0.549 versus 0.083 for explicit questioning (p < 0.001).Other state-of-the-art models also exhibited significant implicit biases in many tested cases.Source: §5.1 · p. 7

5.2 Agent Decisions and Rationales Depend on General Sociodemographic Groups and Specific Attributes.

  • GPT-4o agent decisions varied across both specific attributes and broader sociodemographic groups, with political-ideology-coded agents showing lower mean evacuation, protest-joining, and astronaut-selection rates than gender- and race-coded agents.The reported means for political ideology were 38.60, 55.20, and 32.40, compared with 73.00, 83.67, and 90.67 for gender identity and 74.33, 73.33, and 79.83 for race and ethnicity.Source: §5.2 · p. 7, Figure 4 · p. 8
  • The rationales reflected attribute-linked patterns, including authority and safety language among Asian-coded agents, community-support language among Black- and Native-American-coded agents, and distinct career rationales by political ideology.In authority compliance, 99 of 100 Asian-coded agents complied, while none of the Black- or Native-American-coded agents did; 198 of 200 conservative-coded agents chose plumbing, whereas 124 of 200 liberal-coded agents chose astronaut careers.Source: §5.2 · pp. 7–8

5.3 As LLMs Have Advanced, Explicit Bias Has Decreased but Implicit Bias Has Increased

  • Across GPT generations, explicit bias decreased while implicit bias increased: GPT-4o had significant effects in 1 of 12 explicit cases but 11 of 12 implicit cases.Average implicit DPD rose from 0.069 for GPT-3 to 0.549 for GPT-4o, while average explicit DPD fell from 0.808 to 0.083.Source: §5.3 · p. 9

5.4 Contextualized Persona Generation Helps Reveal Implicit Biases in LLMs

  • Contextualized persona statements revealed implicit disparities more often than no-persona or non-contextualized setups, particularly for gender and race or ethnicity.In GPT-4o’s emergency-response case, gender and race or ethnicity disparities were absent without personas or without contextualization but appeared with contextualized personas.Source: §5.4 · p. 9, Figure 5 · p. 9

5.5 Observed Real-World Differences Often Align with the Direction of Implicit Biases in LLMs

  • All six validated empirical predictions aligned directionally with GPT-4o’s implicit biases.Examples include greater evacuation by female-coded than male-coded agents and lower negative-information sharing by politically moderate-coded agents; the corresponding table summarizes the synthesized evidence and alignments.Source: §5.5 · pp. 9–10, Table 3 · p. 10

6 Discussion

The discussion finds that LLMs can show low explicit bias yet substantial implicit sociodemographic disparities in simulated decisions, with architecture and real-world comparisons clarifying when these disparities emerge and how they should be interpreted.

6.1.1 How do implicit biases in LLMs, as reflected in decision-making disparities across language agents with varying sociodemographic personas, compare to explicit biases when LLMs are directly prompted? (RQ1)

  • Implicit disparities were significant in persona-based decisions despite few biased outcomes under direct questioning, and agent rationales showed stereotyping and deindividuation.Asian-coded agents mostly avoided protest participation, whereas Black-coded and Native American-coded agents mostly joined, citing safety, compliance, ingroup favoritism, and resistance to authority.Source: §6.1.1 · p. 10

6.1.2 To what extent have advances in LLMs mitigated implicit biases compared to explicit biases? (RQ2)

  • Explicit bias decreased from GPT-3 to GPT-4o, whereas implicit bias increased, indicating that current fairness, safety, and alignment mechanisms do not adequately address subtle sociodemographic bias.The discussion attributes reduced explicit bias likely to improvements targeting overt expressions, but reports no corresponding reduction in implicit bias.Source: §6.1.2 · p. 11

6.1.3 What factors in our language agent architecture contribute to eliciting implicit biases in LLMs? (RQ3)

  • Contextually enriched personas paired with scenario information elicited biases more frequently than absent or non-contextualized personas, providing framing for otherwise latent disparities.The controlled comparison identifies the two-step persona-and-action architecture as critical to eliciting implicit bias.Source: §6.1.3 · p. 11

6.1.4 How are the implicit biases revealed by language agents related to observed real-world disparities (RQ4)

  • All six literature-derived predictions aligned directionally with simulated implicit biases, although the observed sociodemographic disparities were often amplified.Agent rationales sometimes invoked sociodemographic stereotypes, supporting analysis of both bias patterns and contextual justifications.Source: §6.1.4 · p. 11

6.2 Design Implications

  • Whether implicit biases should be fully mitigated in social simulations remains open because older models’ lower bias may reflect uniform, less behaviorally meaningful responses.The design question is whether reducing disparities sacrifices variability needed for useful simulation.Source: §6.2 · p. 11
  • Mitigation should depend on whether a bias reflects real-world behavior, historical inequity, or harmful misalignment, and on the simulation’s purpose and use.The discussion recommends avoiding overemphasis on demographic comparisons and examining contextual factors through agent rationales.Source: §6.2 · p. 11

6.3 Limitations and Future Work

  • The study’s single-identity personas may over-align decisions with one social identity, motivating future work on multiple, intersectional identities and their context-dependent activation.People possess intersecting identities, while the study focuses exclusively on agents defined by a single sociodemographic group.Source: §6.3 · pp. 11–12
  • Future work should test magnitude alignment through directly comparable human studies and further analyze agent rationales using open-sourced data and code.These efforts are intended to improve replication, interpretability, awareness of implicit bias, and responsible adjustment strategies.Source: §6.3 · p. 12

7 Conclusion

The paper proposes language-agent simulations to systematically investigate implicit LLM biases across sociodemographic groups and decision scenarios. It finds that advanced models can show stronger simulated disparities despite minimal explicit bias, with some disparities directionally aligning with real-world differences.

  • The paper proposes a language-agent simulation technique to systematically investigate implicit biases across diverse sociodemographic groups and decision-making scenarios.Source: §7 · p. 12
  • State-of-the-art models such as GPT-4o showed more pronounced sociodemographic disparities under simulation despite minimal explicit bias under question answering.Source: §7 · p. 12
  • Several simulated implicit biases aligned directionally with observed real-world behavioral differences between sociodemographic groups.Source: §7 · p. 12

A Contents of Decision-making Scenarios

The scenarios require agents to make decisions about emergency evacuation, protest attendance, information sharing, and career choice. Each presents contextual trade-offs before specifying the available actions.

B Persona Generation Prompts and Scenario Contexts

The study generates sociodemographic, scenario-informed personas and prompts agents to make structured decisions across four behavioral scenarios, while separate questions probe explicit bias.

B.1 Persona Generation

  • Personas are generated from demographic attributes and scenario context, including names, backgrounds, personalities, preferences, and expected behavior.Source: §B.1 · pp. 18–19

B.2 Scenario Contexts

  • The four scenarios test emergency evacuation, authority compliance, negative-information sharing, and career-path selection.Source: §B.2 · p. 19

C Action Generation Prompts

D Question-answering Prompts to Reveal Explicit Bias

E Implicit Biases of 5 LLMs: Llama-3.1, Mixtral-8x7B, GPT-4-turbo, GPT-3.5-turbo, GPT-3

Across five evaluated LLMs, agent decisions show statistically significant demographic-parity differences in multiple implicit-bias test cases, with the frequency varying by model.

  • 8 of 12 Llama-3.1 test cases show statistically significant demographic-parity differences at the 95% confidence level.Source: Figure 6 · p. 25
  • 6 of 12 Mixtral-8x7B test cases show statistically significant demographic-parity differences at the 95% confidence level.Source: Figure 7 · p. 26
  • 9 of 12 GPT-4-turbo test cases show statistically significant demographic-parity differences at the 95% confidence level.Source: Figure 8 · p. 27
  • 10 of 12 GPT-3.5-turbo test cases show statistically significant demographic-parity differences at the 95% confidence level.Source: Figure 9 · p. 28
  • 2 of 12 GPT-3 test cases show statistically significant demographic-parity differences at the 95% confidence level.Source: Figure 10 · p. 29

F Change of Explicit and Implicit Biases with Model Advancement

Across OpenAI’s GPT generations, explicit demographic-parity differences decrease with model advancement, whereas implicit differences persist and increase.

  • Explicit-bias demographic-parity differences decrease across GPT-3, GPT-3.5-turbo, GPT-4-turbo, and GPT-4o.Source: Figure 11 · p. 29
  • Implicit-bias demographic-parity differences persist and even increase across GPT-3, GPT-3.5-turbo, GPT-4-turbo, and GPT-4o.Source: Figure 12 · p. 30

G Comparison of implicit biases across varying simulation processes for Llama-3.1 and Mixtral-8x7B

The simulation-process comparison evaluates how persona and scenario-context design affect implicit-bias demographic-parity differences for Llama-3.1 and Mixtral-8x7B.

  • For Llama-3.1, implicit-bias demographic-parity differences are compared under no-persona, non-contextualized, and contextualized conditions.Source: Figure 13 · p. 30
  • For Mixtral-8x7B, implicit-bias demographic-parity differences are likewise compared across no-persona, non-contextualized, and contextualized conditions.Source: Figure 14 · p. 30

Related papers

Loading 2501.17420v1…