Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat
Sign Up
Cited by 81

The DOI matches its Crossref record; 26 of the 26 references it cites were found in a registry or an index.

Do LLMs Exhibit Human-like Response Biases? A Case Study in Survey Design

Tjuatja, Lindia, Chen, Valerie, Wu, Tongshuang, Talwalkwar, Ameet, Neubig, Graham·2024·Transactions of the Association for Computational Linguistics
DOI 10.1162/tacl_a_00685
Open this paper in CoChat — free account

Abstract

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.1

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1162/tacl_a_00685

Cites

26 of 26 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 Our evaluation framework consists of three steps: (1) generating a dataset of original and modified questions given a response bias of interest, (2) collecting LLM responses, and (3) evaluating whether the change in the distribution of LLM responses aligns with known trends about human behavior. We directly apply the same workflow to evaluate LLM behavior on non-bias perturbations (i.e., question modifications that have been shown to not elicit a change in response in humans).
Figure 1 Our evaluation framework consists of three steps: (1) generating a dataset of original and modified questions given a response bias of interest, (2) collecting LLM responses, and (3) evaluating whether the change in the distribution of LLM responses aligns with known trends about human behavior. We directly apply the same workflow to evaluate LLM behavior on non-bias perturbations (i.e., question modifications that have been shown to not elicit a change in response in humans).CC-BY · From the PDF
Figure 2 We compare LLMs' behavior on bias types ( ¯ ∆ b ) with their respective behavior on the set of perturbations ( ¯ ∆ p ). We color cells that have statistically significant changes by the directionality of ¯ ∆ b (blue indicates a positive effect and orange indicates a negative effect), using p = 0 . 05 cut-off, and use hatched cells to indicate non-significant changes. A full table with ¯ ∆ b and ¯ ∆ p values and p-values is in Table 3. While we would ideally observe that models are only responsive to the bias modifications and are not responsive to the other perturbations, as shown in the top-right the 'most human-like' depiction, the results do not generally reflect the ideal setting.
Figure 2 We compare LLMs' behavior on bias types ( ¯ ∆ b ) with their respective behavior on the set of perturbations ( ¯ ∆ p ). We color cells that have statistically significant changes by the directionality of ¯ ∆ b (blue indicates a positive effect and orange indicates a negative effect), using p = 0 . 05 cut-off, and use hatched cells to indicate non-significant changes. A full table with ¯ ∆ b and ¯ ∆ p values and p-values is in Table 3. While we would ideally observe that models are only responsive to the bias modifications and are not responsive to the other perturbations, as shown in the top-right the 'most human-like' depiction, the results do not generally reflect the ideal setting.CC-BY · From the PDF
Figure 3 Representativeness is a metric based on the Wasserstein distance which measures the extent to which each model reflects the opinions of a population, in this case Pew U.S. survey respondents (the higher the better) [5]. Colors indicate model groupings, with red for the Llama2 base models, green for Solar (instruction fine-tuned Llama2 70b), blue for Llama2 chat models, and purple for GPT 3.5.
Figure 3 Representativeness is a metric based on the Wasserstein distance which measures the extent to which each model reflects the opinions of a population, in this case Pew U.S. survey respondents (the higher the better) [5]. Colors indicate model groupings, with red for the Llama2 base models, green for Solar (instruction fine-tuned Llama2 70b), blue for Llama2 chat models, and purple for GPT 3.5.CC-BY · From the PDF

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Research works printed but not linked: 41

  1. Aher, 2023, Using large language models to simulate multiple humans and replicate human subject studies, 337
  2. Al-Abri, 2014, Patient satisfaction survey as a tool towards quality improvement, Oman Medical Journal, 29, 3, 10.5001/omj.2014.02 · DOI · Paper page
  3. Ayidiya, 1990, Response effects in mail surveys, Public Opinion Quarterly, 54, 229, 10.1086/269200 · DOI
  4. Belinkov, 2017, Synthetic and natural noise both break neural machine translation, arXiv preprint arXiv:1711.02173 · arXiv
  5. Brace, 2018, Questionnaire Design: How to Plan, Structure and Write Survey Material for Effective Market Research
  6. Chen, 2022, Use-case-grounded simulations for explanation evaluation, 1764
  7. Choi, 2005, Peer reviewed: A catalog of biases in questionnaires, Preventing Chronic Disease, 2
  8. Chu, 2023, Language models trained on media diets can predict public opinion, arXiv preprint arXiv:2303.16779 · arXiv
  9. Cox, 1980, The optimal number of response alternatives for a scale: A review, Journal of Marketing Research, 17, 407, 10.1177/002224378001700401 · DOI
  10. Dasgupta, 2022, Language models show human-like content effects on reasoning, arXiv preprint arXiv:2207.07051 · arXiv
  11. Durmus, 2023, Towards measuring the representation of subjective global opinions in language models, arXiv preprint arXiv:2306.16388 · arXiv
  12. Fernandes, 2023, Bridging the gap: A survey on integrating (human) feedback for natural language generation, Transactions of the Association for Computational Linguistics, 11, 1643, 10.1162/tacl_a_00626 · DOI
  13. Gordon, 1987, Social desirability bias: A demonstration and technique for its reduction, Teaching of Psychology, 14, 40, 10.1207/s15328023top1401_11 · DOI
  14. Hauser, 1980, Intensity measures of consumer preference, Operations Research, 28, 278, 10.1287/opre.28.2.278 · DOI
  15. Hippler, 1987, Response effects in surveys, 102, 10.1007/978-1-4612-4798-2_6 · DOI
  16. Hämäläinen, 2023, Evaluating large language models in generating synthetic HCI research data: A case study, 1, 10.1145/3544548.3580688 · DOI · Paper page
  17. Kalton, 1982, The effect of the question on survey responses: A review, Journal of the Royal Statistical Society Series A: Statistics in Society, 145, 42, 10.2307/2981421 · DOI
  18. Kim, 2023, AI-augmented surveys: Leveraging large language models for opinion prediction in nationally representative surveys, arXiv preprint arXiv: 2305.09620 · arXiv
  19. Maus, 2023, Black box adversarial prompting for foundation models
  20. McClendon, 1991, Acquiescence and recency response-order effects in interview surveys, Sociological Methods & Research, 20, 60, 10.1177/0049124191020001003 · DOI
  21. McFarland, 1981, Effects of question order on survey responses, Public Opinion Quarterly, 45, 208, 10.1086/268651 · DOI
  22. Michaelov, 2022, Collateral facilitation in humans and language models, 13, 10.18653/v1/2022.conll-1.2 · DOI
  23. Morwitz, 1996, Do polls reflect opinions or do opinions reflect polls? The impact of political polling on voters’ expectations, preferences, and behavior, Journal of Consumer Research, 23, 53, 10.1086/209466 · DOI
  24. O’Halloran, 2014, Response order effects in the youth tobacco survey: Results of a split-ballot experiment, Survey Practice, 7, 10.29115/SP-2014-0013 · DOI
  25. O’Muircheartaigh, 2001, Middle Alternatives, Acquiescence, and the Quality of Questionnaire Data
  26. Park, 2022, Social simulacra: Creating populated prototypes for social computing systems, 1, 10.1145/3526113.3545616 · DOI · Paper page
  27. Park, 2023, Artificial intelligence in psychology research, arXiv preprint arXiv: 2302.07267 · arXiv
  28. Perez, 2022, Ignore previous prompt: Attack techniques for language models, arXiv preprint arXiv:2211.09527 · arXiv
  29. Pruthi, 2019, Combating adversarial misspellings with robust word recognition, arXiv preprint arXiv:1905.11268 · arXiv
  30. Rawlinson, 2007, The significance of letter position in word recognition, IEEE Aerospace and Electronic Systems Magazine, 22, 26, 10.1109/MAES.2007.327521 · DOI
  31. Sakaguchi, 2017, Robsut wrod reocginiton via semi-character recurrent neural network, 10.1609/aaai.v31i1.10970 · DOI
  32. Scherrer, 2023, Evaluating the moral beliefs encoded in LLMs
  33. Schuman, 1996, Questions and Answers in Attitude Surveys: Experiments on Question Form, Wording, and Context
  34. Schwarz, 1992, A cognitive model of response-order effects in survey measurement, 187, 10.1007/978-1-4612-2848-6_13 · DOI
  35. Sinclair, 2022, Structural persistence in language models: Priming as a window into abstract language representations, Transactions of the Association for Computational Linguistics, 10, 1031, 10.1162/tacl_a_00504 · DOI
  36. Tjuatja, 2023, Syntax and semantics meet in the “middle”: Probing the syntax-semantics interface of LMs through agentivity, 10.18653/v1/2023.starsem-1.14 · DOI
  37. Wallace, 2019, Universal adversarial triggers for attacking and analyzing NLP, 2153, 10.18653/v1/D19-1221 · DOI · Paper page
  38. Webson, 2023, Are language models worse than humans at following prompts? It’s complicated, 7662, 10.18653/v1/2023.findings-emnlp.514 · DOI
  39. Wei, 2022, Finetuned language models are zero-shot learners
  40. Weisberg, 1996, An Introduction to Survey Research, Polling, and Data Analysis
  41. Zheng, 2023, On large language models’ selection bias in multi-choice questions, arXiv preprint arXiv:2309.03882 · arXiv

Cited by

See all 81

Metadata

Authors
Tjuatja, L., Chen, V., Wu, T., Talwalkwar, A., & Neubig, G.
Affiliations
  • Carnegie Mellon University, USA. lindiat@andrew.cmu.edu
  • Carnegie Mellon University, USA. vchen2@andrew.cmu.edu
  • Carnegie Mellon University, USA
Venue
Transactions of the Association for Computational Linguistics
Publisher
MIT Press
Published
2024
Volume
12
Pages
1011-1026
ISSN
2307-387X
DOI
10.1162/tacl_a_00685
Cited by
81
References
26
Language
en
Licence
creativecommons.org/licenses/by/4.0/ (vor, 247 d)

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.