PaperlayerView exact revision on arXiv

Source-linked AI summary

Out of One, Many: Using Language Models to Simulate Human Samples

Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua Gubler, Christopher Rytting, David Wingate

arXiv:2209.06899v1cs.LGcs.CL

TL;DR

Social-science applications of language models face uncertainty about whether their apparent biases and human-like outputs represent specific populations. The paper conditions GPT-3 on socio-demographic backstories to create silicon samples, finding algorithmic fidelity across U.S. public-opinion studies and proposing uses for human-population research while noting important limitations.

  • Problem

    It is unclear whether large language models can serve as effective proxies for specific human sub-populations rather than reflecting only aggregate, demographically skewed training data.

  • Method

    The paper defines algorithmic fidelity and evaluates GPT-3 by conditioning it on socio-demographic backstories from human surveys, then comparing silicon outputs with human responses.

  • Results

    GPT-3 shows a striking degree of algorithmic fidelity in U.S. public opinion, meeting different criteria across three studies, including human-indistinguishable outputs and granular pattern correspondence.

  • Takeaways & Limitations

    With sufficient algorithmic fidelity, language models can provide information about attitudes and ideas across demographic sub-populations and support social-science research.

  • Takeaways & Limitations

    Silicon sampling does not itself guarantee faithful representation of human sub-populations, and the tools retain dangerous potential alongside their research promise.

Abstract

from arXiv · show

We propose and explore the possibility that language models can be studied as effective proxies for specific human sub-populations in social science research. Practical and research applications of artificial intelligence tools have sometimes been limited by problematic biases (such as racism or sexism), which are often treated as uniform properties of the models. We show that the "algorithmic bias" within one such tool -- the GPT-3 language model -- is instead both fine-grained and demographically correlated, meaning that proper conditioning will cause it to accurately emulate response distributions from a wide variety of human subgroups. We term this property "algorithmic fidelity" and explore its extent in GPT-3. We create "silicon samples" by conditioning the model on thousands of socio-demographic backstories from real human participants in multiple large surveys conducted in the United States. We then compare the silicon and human samples to demonstrate that the information contained in GPT-3 goes far beyond surface similarity. It is nuanced, multifaceted, and reflects the complex interplay between ideas, attitudes, and socio-cultural context that characterize human attitudes. We suggest that language models with sufficient algorithmic fidelity thus constitute a novel and powerful tool to advance understanding of humans and society across a variety of disciplines.

1 Introduction

The paper reframes algorithmic bias as many demographically correlated patterns rather than one uniform property, proposing algorithmic fidelity as a way to emulate human sub-populations. Across three studies, GPT-3 is conditioned on socio-demographic backstories and compared with human responses in U.S. politics and public opinion.

  • Algorithmic bias reflects multiple associations among ideas, attitudes, and contexts rather than a single macro-level model property.
  • Algorithmic fidelity is the model’s ability to produce response distributions aligned with distinct human sub-populations after conditioning on targeted identity and personality profiles.
  • The studies condition GPT-3 on thousands of socio-demographic backstories to evaluate whether its outputs match human patterns across groups and intersecting demographics.
  • Study 1 compares GPT-3 surrogate descriptions of outgroup partisans with human descriptions, while Studies 2 and 3 examine relationships among demographics, attitudes, and reported behaviors.
  • The paper introduces algorithmic fidelity, silicon sampling, first-person demographic conditioning, evidence for GPT-3 fidelity in U.S. politics, and social-science applications.

2 The GPT-3 Language Model

GPT-3 is a large language model trained on extensive text that generates language by predicting tokens sequentially. Its outputs depend strongly on the conditioning context supplied before generation.

  • GPT-3 has 175 billion parameters and was trained on more than 45 terabytes of text, with generated texts often difficult to distinguish from human compositions.
  • A language model represents a conditional probability distribution over tokens and generates text by iteratively sampling from that distribution.
  • Conditioning provides the initial input tokens, called the model’s context, before text generation begins.
  • Changing a single word in the context can reverse which next-token outputs receive high probability.

3 Algorithmic Fidelity

Algorithmic fidelity describes how closely a model’s relationships among ideas, attitudes, and socio-cultural contexts mirror those of human sub-populations. The paper evaluates this concept through continuity, human indistinguishability, and pattern correspondence criteria.

  • Algorithmic fidelity is the degree to which a model mirrors complex relationships among ideas, attitudes, and socio-cultural contexts across human sub-populations.
  • Algorithmic fidelity does not mean GPT-3 simulates a specific individual or produces coherent responses every time.
  • The four criteria are a Social Science Turing Test, Backward Continuity, Forward Continuity, and Pattern Correspondence.
  • The criteria assess human indistinguishability, consistency with conditioning information, natural continuation of context, and human-like relationships among ideas, demographics, and behavior.
  • Missing fidelity in any criterion lowers confidence in reliable surrogate responses, and the paper does not specify numerical thresholds for passing.
  • The paper evaluates GPT-3 against human-produced patterns in U.S. politics using established political-science relationships as a comparison basis.

4 Silicon Sampling: Correcting Skewed Marginals

Silicon sampling addresses demographic imbalance in language-model training by conditioning GPT-3 on backstories drawn from representative human samples. This enables estimation of population-level patterns from conditional outputs, provided the model faithfully represents the relevant conditional distributions.

  • Internet-user demographics and fixed-time training data may not represent or balance the populations social scientists want to study.
  • GPT-3 jointly models voting patterns and its training-data demographics, so skewed demographic backstories can distort marginal voting estimates.
  • Silicon sampling draws backstories from a known nationally representative sample and estimates population patterns from GPT-3 outputs conditioned on those backstories.
  • Conditioned and generic GPT-3 completions can exhibit different sub-population opinions, biases, and voting patterns, analogous to Simpson’s Paradox.
  • Sampling GPT-3’s component distributions alone does not establish that they faithfully represent specific human sub-populations; algorithmic fidelity must first be assessed for the domain and groups.
  • Figure 1 shows conditioning contexts and completions for four silicon individuals, with inserted demographics and four harvested words visually distinguished.

5 Study 1: Free-form Partisan Text

Study 1 tests whether GPT-3 can generate partisan word lists that resemble human lists in source indistinguishability, content, tone, and ideological patterning. Across these tests, GPT-3 lists closely matched human evaluations and retained cues about writers’ partisanship.

  • Silicon sampling: GPT-3 generated four-word partisan lists from first-person backstories constructed for human survey respondents.The conditioning design usually elicited four delineated words, with regular-expression post-processing extracting the final set.
  • Source recognition: Human evaluators identified human-generated lists at 61.7% and GPT-3 lists at 61.2%, with no significant difference (p=0.44).Nearly 40% of both list types were judged computer-generated, and this tendency did not vary by source.
  • Content and tone: Human and GPT-3 lists showed similar content and tone across five characteristics, including personality traits in 72.3% versus 66.5% of lists and extremity in 39.8% versus 41.0%.The similarity extended across the characteristic patterns, with traits the main high-frequency exception in both sources.
  • Ideological patterning: GPT-3 reproduced human-similar use of positive and extreme words across ideological subgroups, supporting pattern correspondence.The subgroup comparison examined underlying patterns rather than only aggregate frequencies.
  • Partisanship cues: Participants guessed list writers’ partisanship above the 33% three-choice chance level for both human and GPT-3 lists.Human lists yielded 60.1% successful guesses versus 52.8% for GPT-3 lists, a significant 7.3-percentage-point difference (p<.001).
  • Overall assessment: Study 1 provides repeated support for the Turing Test and Backward Continuity criteria, with initial evidence for Pattern Correspondence across American subgroups.The evidence comes from multiple measures of source recognition, list evaluation, and subgroup patterning.

6 Study 2: Vote Prediction

Study 2 evaluates whether GPT-3, conditioned on ANES demographic and attitudinal profiles, can reproduce presidential vote-choice distributions across years and subgroups. The generated responses closely corresponded with human data, despite modest aggregate biases and a notable independent-voter exception.

  • Design: GPT-3 was conditioned on ANES participants’ demographic, ideological, behavioral, and political-attitudinal characteristics to predict presidential vote choice.The conditioning variables included race or ethnicity, gender, age, ideology, party identification, political interest, church attendance, political discussion, patriotism, and state.
  • Aggregate vote choice: GPT-3 and ANES two-party vote proportions closely corresponded across 2012, 2016, and 2020, although the model showed modest overall candidate-specific bias.GPT-3 estimated 0.391 versus 0.404 for Romney in 2012, 0.432 versus 0.477 for Trump in 2016, and 0.472 versus 0.412 for Trump in 2020.
  • Overall correspondence: Across all three survey years, GPT-3 and ANES vote reports showed remarkably high correspondence after GPT-3 probabilities were dichotomized for comparison.Table 1 compares tetrachoric correlations and proportion agreement between the binary GPT-3 vote and ANES self-reports.
  • Subgroup correspondence: More than half of subgroup tetrachoric correlations were at least 0.90, while all other correspondence measures exceeded 0.65 and generally approached 0.8 or 0.9.Pure independents, especially in 2020, were the sole reported deviation from the broader subgroup pattern.
  • Implications: The results support GPT-3’s ability to capture voting preferences across groups and time, including possible use beyond its original training corpus with appropriate conditioning.The authors interpret the repeated cross-year and subgroup correspondence as support for Forward Continuity and Pattern Correspondence.

7 Study 3: Closed-ended Questions and Complex Correlations in Human Data

Study 3 tests whether GPT-3 can reproduce complex associations among demographic, attitudinal, behavioral, and political survey variables. Comparing Cramer’s V across ANES and silicon data, the study finds that GPT-3 mirrors both stronger and weaker human relationships.

  • Study design: Study 3 extends vote prediction to a broader set of closed-ended survey responses and evaluates a complex structure of associations in 2016 ANES data.It is presented as the paper’s most rigorous evaluation of Pattern Correspondence.
  • Conditioning method: An interview-style conditioning template elicited short responses from specified options while supplying demographic and attitudinal background information.The format leveraged zero-shot learning and incorporated survey questions into the conditioning text.
  • Association measure: Cramer’s V compared associations among ANES survey items with associations between ANES conditioning values and GPT-3 answers.Cramer’s V summarizes association while accounting for variation in raw-data base rates.
  • Results: The mean difference between human and GPT-3 Cramer’s V values was -0.026, with GPT-3 reproducing both stronger and weaker relationships found in human data.The generated associations were not uniformly high or low; weak human relationships remained weak, and stronger ones were likewise reflected.
  • Conclusion: The study reports compelling, consistent, repeated evidence that GPT-3 reproduces nuanced inter-correlations beyond aggregated topline responses.The authors state that GPT-3 answers closed-ended questions in ways that closely mirror human respondents when given real survey data as inputs.

8 Where do we go from here?

The paper argues that GPT-3 outputs can provide useful social-science evidence even before or without human data. In Study 1, silicon samples reproduced interpretable patterns about partisan stereotypes, affect, extremity, and perceived authorship, while Studies 2 and 3 suggest uses for identifying variables and relationships for further research.

  • GPT-3 data alone suggested that partisan descriptions emphasize distinct stereotypes and that their affective content and extremity track political beliefs and identity.
  • Silicon-sample descriptions contained issue-, group-, and trait-based content, with trait references most common.
  • GPT-3 outputs indicated that others could guess the partisanship of their writers, supporting their use for generating social-science theory.
  • Study 2’s ablation analysis suggests which variables researchers should include to understand Americans’ voting behavior accurately.
  • Study 3 could help social scientists target connections between characteristics and views for further exploration.

9 Discussion

The paper presents algorithmic fidelity as a framework for evaluating language models as aggregate proxies for human cognition. Across studies, GPT-3 showed repeated fidelity in U.S. public opinion, while the authors emphasize both its potential uses and ethical and domain-specific boundaries.

  • Algorithmic fidelity provides a framework for using large-scale language models as aggregate proxies for human cognition and windows into human thinking.
  • The paper proposes four criteria and empirical methods for evaluating algorithmic fidelity.
  • GPT-3 showed a striking degree of algorithmic fidelity in U.S. public opinion across the paper’s studies.
  • The studies provide repeated evidence across data sources, measures, and points in time, including viewpoints from demographically varied U.S. sub-populations.
  • The authors warn that highly faithful models could be used to target groups for misinformation, manipulation, or fraud.
  • The current study is restricted to a specific domain, so further work is needed to quantify algorithmic fidelity and its limitations across social-science fields.

A General details on GPT-3 usage

The appendix describes GPT-3 usage for generating conditioned text and estimating probabilities. It documents model settings, token-probability aggregation, temperature choices, and examples of demographic backstories with sampled continuations.

  • The study accessed GPT-3’s Davinci model through OpenAI’s remote API, using 175 billion parameters and temperature 0.7.
  • For probability modeling, token probabilities were combined with log-sum-exp when multiple token forms represented the same answer.
  • Studies 1 and 3 used temperature 0.7 for text generation, favoring more deterministic answers over response diversity.
  • Figure 5 shows first-person demographic backstories paired with multiple GPT-3 continuations, including voting patterns and related beliefs or opinions.

B Details on Study 1

Study 1 constructs matched human and GPT-3 samples from partisan, demographic, and socioeconomic information, then evaluates generated and human partisan descriptions. The appendix details compliance, extraction, response length, and blinded Lucid evaluation procedures.

  • 2,107 Pigeonholing Partisans participants supplied complete partisan identification, ethnicity, gender, income, and age information for analysis.
  • The researchers generated 2,107 matched silicon individuals by conditioning GPT-3 on ordered first-person demographic backstories.
  • Human and GPT-3 subjects each described Republicans and Democrats with word lists, although some responses were incomplete or non-compliant.
  • 3,592 human texts and 4,083 GPT-3 texts remained after non-compliant responses were culled, producing 7,675 unique lists.
  • The mean human response length was 4.54 words, compared with 7.78 words for GPT-3; the modal response was four words for both.

B.3 Lucid results analysis

The Lucid analysis compares human and GPT-3-generated partisan word lists using controlled regression models, evaluator fixed effects, and clustered standard errors. Full model tables report list-content evaluations, partisan-identification accuracy, and ideology-subset predictions.

  • Inference: Standard errors are clustered two ways by evaluators and lists, with evaluator fixed effects included in the reported specifications.The regression outputs display two-way standard errors in parentheses.
  • Model specification: The models control for gender, ethnicity, income, age, and party identification, with word length added for partisan-identification accuracy.Predicted values use fixed reference levels for included covariates.
  • Model specification: OLS models predict the percent of texts receiving each of five content evaluations from source and list-writer characteristics.The five characteristics are positivity, extremity, and mentions of traits, issues, and groups.
  • Partisanship prediction: Table 3 reports models predicting whether Lucid participants correctly guessed the list writer’s partisanship.These results correspond to the top-left bars in Panel B of Figure 4.
  • Ideology-subset results: Tables 4–7 provide ideology-subset predictions for positivity and extremity when describing Republicans or Democrats.The tables use the seven-point ideology scale from Extremely Conservative to Extremely Liberal.

C.2 Data analysis

The vote analysis converts GPT-3 candidate probabilities into binary predictions and compares them with ANES votes using four correspondence statistics. It also tests how backstory elements and model family affect prediction performance.

  • Vote correspondence: GPT-3 probabilities above 0.50 are coded as Republican votes and probabilities below 0.50 as Democratic votes before comparison with ANES responses.No predicted probability equals exactly 0.50.
  • Vote correspondence: Four statistics assess correspondence between ANES and GPT-3 vote variables: tetrachoric correlation, Cohen’s Kappa, ICC, and proportion agreement.The analysis reports the full metric set beyond the two measures shown in the main text.
  • Metric interpretation: Near-complete lack of vote variation can make correlations and Kappa unreliable even when ANES and GPT-3 estimates almost completely agree.Proportion agreement is retained as a descriptive measure in such cases.
  • Ablation analysis: Ablation removes one or two of ten backstory elements or tests single-element backstories to assess their contribution to vote prediction.The experiment compares the full template with reduced templates.
  • Ablation analysis: No single element accounts for all predictive power; Party is more predictive than Ideology, while some additions mildly reduce performance.The remaining eight demographic elements outperform any single element even after removing Party and Ideology.
  • Limitation: The template and ten selected elements were not optimized, leaving room for future improvement through template and backstory-element optimization.The authors describe the template as their first try.
  • Model comparison: The largest GPT-Neo model, with 6B parameters, rivals GPT-3’s performance at 175B parameters on vote prediction.Performance is shown as a function of parameter count, while training-corpus differences also matter.

D.1 Data generation

Study 3 constructs virtual ANES interviews by conditioning GPT-3 on demographic, attitudinal, and behavioral variables, then predicts each target response from the others. The resulting synthetic data are compared with human relationships across multiple generation settings.

  • Study 3 design: Study 3 constructs virtual ANES participants from twelve demographic, attitudinal, and behavioral variables, with vote turnout and choice combined into one metric.The study therefore analyzes eleven unique items.
  • Template construction: ANES variables are mapped to first-person template fragments and concatenated into conditioning contexts using interview wording whenever possible.Missing ANES variables cause the corresponding question to be omitted from the context.
  • Response generation: For each target factor, GPT-3 receives the other factors as context and generates five tokens to elicit the target response.The target fragment is placed at the end and excluded from the conditioning context; sampling uses temperature 0.7.
  • Sample construction: The 12 variables and 4,270 respondents produce more than 50,000 conditioning texts, with 1,782 complete ANES–GPT-3 observations retained.Complete-case filtering avoids variation in the comparison set adding statistical noise.
  • Data quality: GPT-3 produces valid compliant answers in more than three-quarters of cases for every question item, with three items exceeding 99% compliance.The authors note that question-specific training could reduce non-compliance.
  • Synthetic-data analysis: Synthetic vectors can be compared with one another using Cramer’s V, and their relationship pattern is highly similar to comparisons using ANES inputs.The all-synthetic comparison introduces additional estimation noise.
  • Temperature variation: Mid-range temperature settings appear to balance valid completions against the greater invalid-response rate produced by higher temperatures.Lower temperatures are more deterministic, whereas higher temperatures sample more invalid tokens.
Loading 2209.06899v1…