Source-linked AI summary
Do LLMs exhibit human-like response biases? A case study in survey design
Lindia Tjuatja, Valerie Chen, Sherry Tongshuang Wu, Ameet Talwalkar, Graham Neubig
TL;DR
The paper examines whether LLMs can serve as human proxies by asking whether their survey responses exhibit the response biases documented in human studies. It develops a survey-based evaluation framework and applies it to nine models, finding generally non-human-like behavior, especially around RLHF, and little correspondence between opinion-distribution replication and human-like response biases.
Problem
The paper investigates whether LLMs reflect human response biases in surveys, an important question for using them as proxies in tasks involving subjective labels.
Method
The authors create original and modified survey-question pairs, collect LLM response distributions, and compare changes against human bias trends and human-robust perturbations.
Results
Across nine models, LLMs generally fail to reflect human-like behavior; RLHF-ed models differ from vanilla models, and opinion-distribution replication generally does not correspond to human-like response biases.
Takeaways & Limitations
The findings indicate that LLMs should be used as human proxies with care and evaluated using finer-grained characterizations of behavior.
Takeaways & Limitations
The experiments focus on English-based, U.S.-centric survey questions, do not directly compare responses on modified questions with human responses, and cover neither all languages nor all response biases.
Abstract
from arXiv · showhide
As large language models (LLMs) become more capable, there is growing excitement about the possibility of using LLMs as proxies for humans in real-world tasks where subjective labels are desired, such as in surveys and opinion polling. One widely-cited barrier to the adoption of LLMs as proxies for humans in subjective tasks is their sensitivity to prompt wording - but interestingly, humans also display sensitivities to instruction changes in the form of response biases. We investigate the extent to which LLMs reflect human response biases, if at all. We look to survey design, where human response biases caused by changes in the wordings of "prompts" have been extensively explored in social psychology literature. Drawing from these works, we design a dataset and framework to evaluate whether LLMs exhibit human-like response biases in survey questionnaires. Our comprehensive evaluation of nine models shows that popular open and commercial LLMs generally fail to reflect human-like behavior, particularly in models that have undergone RLHF. Furthermore, even if a model shows a significant change in the same direction as humans, we find that they are sensitive to perturbations that do not elicit significant changes in humans. These results highlight the pitfalls of using LLMs as human proxies, and underscore the need for finer-grained characterizations of model behavior. Our code, dataset, and collected samples are available at https://github.com/lindiatjuatja/BiasMonkey
1 Introduction
The paper asks whether LLMs mirror human response biases in surveys, where wording, format, and placement changes can systematically alter human responses. Across nine models, it finds generally non-human-like behavior and weak correspondence between opinion-distribution accuracy and human-like response biases.
- Motivation: LLMs are considered potential participants or proxies in surveys because prior work found they can replicate human results on tasks involving subjective labels.These tasks include preference annotation, social science and psychological studies, and opinion polling.
- Motivation: Human response biases can arise from changes to survey-question wording, format, or placement, producing predictable deviations from original responses.The study investigates whether LLM responses show parallels to these human sensitivities.
- Approach: The evaluation generates original and modified question pairs, collects LLM response distributions, and compares their changes with known human trends, including non-bias perturbations.The same workflow tests modifications such as perturbations known not to change human responses.
- Findings: Across five response biases and three non-bias perturbations, all evaluated models showed behavior notably unlike humans in at least some ways.Observed mismatches included changes opposite to known human biases and significant changes to perturbations that do not affect human responses.
- Findings: RLHF-ed models showed less significant responses to bias modifications but greater sensitivity to non-bias perturbations than vanilla models.The authors identify these differences as potentially undesirable effects of additional training schemes.
- Implications: A model’s ability to replicate human opinion distributions is not indicative of how well it reflects human behavior.The conclusion motivates more extensive evaluations that distinguish different dimensions of human-model similarity.
2 Methodology
The methodology constructs paired survey questions representing human-inducing biases and human-robust perturbations, samples responses from nine varied LLMs, and tests distributional changes against human trends. Its analysis uses bias-specific response-option changes, averages them across questions, and applies statistical tests for consistency.
- Framework: The evaluation framework has three stages: dataset generation, LLM response collection, and response analysis.The framework is designed to compare model behavior with patterns established in human studies.
- Dataset generation: Each question pair contains an original multiple-choice question q and a modified form q′ to test whether prompt changes produce a response bias.This follows the control-versus-treatment structure used in prior social science studies.
- Dataset generation: Qbias contains modifications known to induce specific human biases, whereas Qperturb applies human-robust changes such as typos or randomized letter changes.For Qperturb, the original question is also drawn from Qbias and its modified version applies a non-bias perturbation.
- Dataset generation: The dataset contains 2578 question pairs covering five response biases and three non-bias perturbations, based on expert-curated Pew American Trends Panel questions.The question topics were selected for similarity to domains studied in prior social psychology research.
- Models and sampling: For each question, the study samples 50 valid responses per model to estimate distributions Dq and Dq′.Responses are collected at temperature 1 until they match one of the provided letter options.
- Models and sampling: The study evaluates nine models spanning open and commercial systems, instruction fine-tuning, RLHF, and parameter count.Models include Llama2 base, instruction-tuned, and chat variants across 7b, 13b, and 70b sizes, plus Solar.
- Response analysis: The analysis computes bias-specific changes ∆b between Dq and Dq′, averages them across questions, and uses a Student’s t-test against a zero-change null hypothesis.The same procedure computes ∆p for Qperturb, where the expected average change is not statistically different from 0.
3 Results
Across nine models, LLMs generally failed to reproduce human-like response-bias patterns and were significantly affected by perturbations that humans typically resist. Training scheme mattered: RLHF models were less responsive to bias-inducing changes but more responsive to non-bias perturbations.
- 3.1 General trends in LLM behavior: No model aligned with known human response-bias patterns across all five biases, and all models changed significantly under non-bias perturbations.The evaluation compared nine models across five response biases and three perturbation types.
- 3.1 General trends in LLM behavior: Llama2 70b showed the most human-like responses, yet still changed significantly under non-bias perturbations for three of five bias types.Its residual sensitivity illustrates that matching bias direction did not imply robustness to irrelevant modifications.
- 3.1 General trends in LLM behavior: Model size showed no consistent monotonic relationship with the magnitude of response-bias effects across the 7b, 13b, and 70b base Llama2 models.Only a handful of biases, including allow/forbid and opinion float, increased or decreased with parameter count.
- 3.2 Comparing base models with their modified counterparts: RLHF-ed models were less sensitive to bias-inducing changes than vanilla models, especially for wording changes such as acquiescence and allow/forbid.Odd/even was an exception: all but one RLHF-ed model had a larger positive effect size than the Llama2 base models.
- 3.2 Comparing base models with their modified counterparts: RLHF-ed chat models had larger non-bias perturbation effects in 23 of 29 shared significant settings, averaging 81% larger effects than base models.They also tended to show more significant changes from perturbations, which the authors characterize as less human-like.
4 Examining the effect of uncertainty
The paper tests whether LLM uncertainty predicts reduced sensitivity to question modifications, as prior human studies suggest. Most models show no such correspondence, indicating a mismatch with human behavior.
- Interpretation: The evaluation therefore finds that most models are not well calibrated with respect to this human response-bias pattern.The human pattern is that greater confidence predicts less influence from question modifications.
- Measuring uncertainty: Normalized entropy measures answer-distribution uncertainty across questions with different numbers of response options.A value of 0 indicates maximal confidence, while 1 indicates maximal uncertainty.
- Results: In seven of nine models, uncertainty does not correspond to the magnitude of bias-related response changes.This provides further evidence of dissimilarities between human and LLM behavior.
- Results: Llama2 70b and GPT 3.5 turbo-instruct show weakly positive significant correlations between uncertainty and the magnitude of bias-related changes.The reported correlations satisfy 0.2 ≤r ≤0.5 with p < 0.05.
5 Comparison to other desiderata for LLMs as human proxies
The paper compares human-like response biases with representativeness of population opinion distributions. These evaluations do not consistently rank models in the same way, so matching opinion distributions is not indicative of human-like behavior.
- Comparison: The study evaluates whether models that better replicate population opinions also exhibit more human-like response biases.This separates matching opinion distributions from responding like humans to question modifications.
- Representativeness: Representativeness compares each model’s answer distribution with human responses from the Pew American Trends Panel.The authors aggregate model responses for unmodified questions, construct corresponding human distributions, and compute similarity between them.
- Findings: Llama2 70b has the highest representativeness score and the most human-like behavior relative to the ideal setting, but model rankings are otherwise inconsistent.The comparison uses Figure 3’s representativeness scores alongside behavior on modified questions.
- Findings: Llama2 7b-chat and 13b-chat show similar behavior and representativeness, whereas GPT 3.5 turbo and turbo instruct behave differently despite extremely close representativeness scores.This contrast directly demonstrates that the two desiderata can diverge.
- Conclusion: Replicating human opinion distributions is not indicative of how well an LLM reflects human behavior.The paper treats response-bias behavior and population-level opinion matching as distinct evaluation dimensions.
6 Related Work
Prior work studies prompt sensitivity, human-like behavior, and direct LLM–human comparisons across several task types. This paper addresses limitations in those comparisons by evaluating survey-based response biases using stimuli grounded in established human studies.
- LLM prompt sensitivity: Research has examined LLM sensitivity to adversarial prompts, example order, multiple-choice wording, and few-shot formatting.These studies characterize prompt-construction effects across several settings.
- Human-like behavior: A smaller literature tests whether LLM changes resemble human linguistic priming and cognitive biases, often outside downstream use cases.The paper argues that such settings may provide limited guidance for using LLMs as human proxies.
- Stimulus validation: Claims about human-like behavior depend on stimuli verified in prior human studies, because initial assumptions about human responses can later be disproved.The paper grounds its evaluation in long-standing social-science literature.
- Comparison settings: LLM–human comparisons distinguish open-ended outputs judged by humans from closed-form labels generated from fixed response options.This study focuses on closed-form tasks to enable direct comparisons.
- Prior evaluations: Existing evaluations compare LLM and human outcomes in social science, annotation, and opinion surveys, but are limited to the question forms used to collect human responses.The present work instead studies behavior under systematic question modifications.
7 Conclusion
Using survey design as a case study, the paper evaluates behaviors that could affect LLM suitability as human proxies. Across nine models, it finds broad divergence from human-like response biases and weak correspondence between such biases and opinion representativeness.
- Conclusion: The study provides a comprehensive evaluation of behaviors that could make LLMs more suitable as human proxies.Survey design serves as the case study for these desired behaviors.
- Conclusion: Across nine evaluated models, LLMs are generally not reflective of human-like behavior.The conclusion covers both open and commercial model families evaluated in the study.
- Training effects: Llama2 base and chat models show distinct behavior, highlighting effects associated with additional training schemes, especially RLHF.The paper notes that RLHF may produce greater sensitivity to specific perturbations despite improving helpfulness and harmlessness.
- Evaluation criteria: Replicating human opinion distributions generally does not correspond to exhibiting human-like response biases.The paper presents this as a separate limitation of using opinion matching as a proxy criterion.
- Implications: The findings highlight limitations of using LLMs as human proxies in survey design and motivate more critical evaluations of similarities and differences with humans.The conclusion calls for finer-grained characterization of model behavior.
8 Limitations
The study is limited to English-based, U.S.-centric survey questions and five response biases, and it emphasizes behavioral trends rather than exact human–LLM effect magnitudes.
- The experiments focus on English-based, U.S.-centric survey questions.The authors suggest replicating these evaluations with more diverse languages and users.
- The evaluation compares LLM trends with extensively studied human trends rather than directly comparing responses on every modified question.Consequently, the analysis emphasizes directions of behavioral change instead of specific magnitudes.
- The five response biases studied are neither representative nor comprehensive of all human biases.The work is intended to illustrate an approach for studying similarities between human and LLM behavior, not to exhaustively test human biases.
A Stimuli implementation
The stimuli are drawn from a formatted subset of the ATP dataset and filtered to retain subjective opinion questions rather than factual or daily-habit questions.
- The study uses a subset of the ATP dataset formatted into CSV files from prior work.The source is linked in the paper, and the Pew Research Center disclaims responsibility for analyses or interpretations of the data.
- Questions about daily habits and factual information were filtered out because the study targets subjective questions.Examples include smoking frequency and marital status.
A.2 Qbias and Qperturb Details
The study constructs paired survey questions by implementing five human response biases and three human-robust perturbations through manual or systematic modifications.
- Acquiescence: Acquiescence pairs use two-option ATP questions, with the modified question reworded toward the first original option.The comparison counts responses selecting option ‘a’.
- Allow/forbid and response order: Allow/forbid candidates are identified through keywords such as “allow” or “acceptable,” while response-order pairs reverse the answer-option order.Responses to reversed options are mapped back to the original order for analysis.
- Scale and opinion-floating effects: Opinion-floating pairs add a “don’t know” option, whereas odd/even scale effects use questions requiring odd-numbered scales.Both procedures are based on the relevant survey-question sets described in the implementation.
- Non-bias perturbations: Middle-random perturbations swap an interior character with a neighbor in words of length at least four, while key typos randomly replace letters with 20% probability.These perturbations are applied only under constraints designed to avoid nonsensical outputs.
- Non-bias perturbations: Non-bias perturbations are excluded from words containing numeric values or punctuation.This restriction prevents completely nonsensical modified questions.
A.3 Full results
The full-results appendix identifies the reported metrics, model sources, prompting template, and sampled question tables used to document the evaluation.
- Results table: Table 3 reports average bias changes and t-test p-values for each bias, average perturbation changes and p-values, and Pearson r values for uncertainty relationships.These metrics organize the full stimulus-level results.
- Models: The evaluated model materials include Base Llama2 and Llama2 chat variants, Solar, GPT-3.5-turbo, and GPT-3.5-turbo-instruct.The appendix provides model-weight links where applicable and identifies specific OpenAI model versions.
- Prompting: All models receive a minimal prompt asking them to answer a question using one of its alphabetical options.The main experiments restrict generation to one token.
- Stimulus tables: Sampled-question tables document acquiescence, allow/forbid, response-order, odd/even-scale, opinion-floating, key-typo, and letter-swap stimuli.The appendix also includes the full set of results in Table 3.