PaperlayerView exact revision on arXiv

Source-linked AI summary

Consequential Behaviour and Representational Fairness in the Validation of Synthetic Research

Florian Kutzner, Celina Kacperski, Laura de Molière, Edoardo Chidichimo, Min Jun Jung, Felix P. S. Wallis, James K. He

arXiv:2609.27690v2cs.CLcs.CY

TL;DR

Synthetic research is often validated against human survey responses, although those comparisons do not establish whether synthetic samples predict consequential behaviour. The paper proposes a framework for behavioural and subgroup validity, with diagnostic levels, justice measures, and within-persona counterfactual experiments. It concludes that dispersion and structure can be misrepresented despite aggregate agreement, and notes that prompt changes can confound counterfactual contrasts by changing the inferred respondent.

  • Problem

    Comparisons with human survey responses do not establish whether synthetic samples predict consequential behaviour or preserve the human relationship between intention and behaviour.

    Source: Introduction · p. 2, §1.2 · p. 3

  • Method

    The paper proposes validity claims that specify behavioural criteria, diagnostic levels, subgroup measures, and within-persona counterfactual experiments.

    Source: Introduction · p. 2, §2 · p. 4

  • Results

    Synthetic samples can compress response variance and distort variable relationships even when aggregate distributions look acceptable.

    Source: Dispersion (L1) · p. 6, Structure (L3) · p. 6

  • Takeaways & Limitations

    Researchers should report validity claims for subgroups and state the target level of correspondence, including whether validation compares experimental effects.

    Source: Introduction · p. 2, §2 · p. 4

    Changing a prompt to implement a treatment can alter the model’s inference about the respondent, confounding the contrast between conditions.

    Source: §2.3 · p. 7

Abstract

from arXiv · show

Researchers in industry and academia use synthetic survey respondents powered by large language models as substitutes for human samples. These synthetic populations require validation against real-world data, so researchers often address them using ad hoc comparisons with human surveys. Inspired by the intention-behaviour gap in behavioural science, we argue that these validations test the wrong thing for most applied cases where decision makers commission synthetic research to anticipate consequential behaviour. To address this problem, we propose a validation framework with two requirements. First, every validity claim must state its level of correspondence with human data: does the sample predict what the represented people do, which of four diagnostics (location, dispersion, response process and structure) does the validation address, and does the validation compare against experimental effects? Second, researchers must report validity claims for subgroups, since these groups are often the most affected by consequential decisions and aggregate accuracy hides their misrepresentation. Our validation framework operationalises three justice dimensions (distributional, procedural, and recognition) as measurable quantities and defines within-persona counterfactual experiments as a validation requirement. We then apply the framework to electric vehicle charging tariffs, before closing with a reporting checklist that researchers can use to make convincing validity claims.

Introduction

Synthetic survey respondents can make research faster and easier to conduct, but their familiar persona format can also make inaccuracies hard to detect. The paper argues that validation should address consequential behaviour and subgroup representation, not rely on aggregate agreement with survey answers.

Introduction

  • Synthetic respondents can reduce months of fieldwork to days, but population-scale simulation still consumes substantial compute and energy.Their persona format also makes them easy to access and use in familiar practitioner workflows.Source: Introduction · p. 2
  • Synthetic personas are easy to use but difficult to validate, and can be duplicated as parallel counterfactuals in ways that are impossible for humans.No convention currently exists to evaluate their potential for estimating individual-level treatment-effect heterogeneity.Source: Introduction · p. 2
  • Agreement with human survey responses cannot establish whether synthetic samples predict consequential behaviour or preserve humans’ intention–behaviour relationship.Stated preferences can overstate willingness to pay relative to revealed choices, and self-reports of pro-environmental behaviour correlate only moderately with actual behaviour.Source: Introduction · p. 2
  • The framework requires explicit validity levels and subgroup claims, operationalising distributional, procedural, and recognition justice through subgroup error, documentation, and preserved variance.It also describes within-persona counterfactual experiments and remains agnostic to how synthetic respondents are constructed.Source: Introduction · p. 2

1 What to validate

The paper treats a persona as a synthetic respondent specified by a profile and distinguishes that representation from a named living person. It argues that validation for applied research should target consequential behaviour, because matching what people say does not establish matching what they do.

1.1 Personas and degrees of specificity

  • In machine learning, a persona usually means the text conditioning a model, ranging from a demographic attribute to an interview transcript.More specific profiles can improve representation of individuals, but whether greater specificity improves behavioural prediction remains open.Source: §1.1 · pp. 2–3
  • The paper restricts persona to a synthetic respondent representing a hypothetical archetype, and reports claims about sets of personas by stratum.Constructing a persona to represent a named living person from material about that person can raise data-protection and profiling issues under the GDPR and EU AI Act.Source: §1.1 · p. 3

1.2 Behaviour as the target of prediction

  • Validation should target consequential behaviour in a defined population and conditions, since reports and actions are different measurable quantities.Cost, opportunity, and social desirability affect reports and behaviour differently, and the intention–behaviour gap widens when behaviour is costly.Source: §1.2 · p. 3
  • Prior work moved from survey-item comparisons and questionnaire-bias studies toward laboratory tasks and economic games, which have their own lab-to-field gap.The paper argues that such validation must still distinguish stated responses from consequential behaviour.Source: §1.2 · p. 3
  • The paper calls for validation to state whether it targets verbal or behavioural correspondence, noting that behavioural studies in flagship journals have declined.Some counts put the share of studies measuring actual behaviour at 6 to 18% (Doli´nski, 2018).Source: §1.2 · p. 3

2 Levels of correspondence between synthetic and human responses

The framework treats validity as dependent on a study’s interpretation and use, so claims must specify the correspondence level they address. It separates the behavioural criterion from diagnostic levels and calls for uncertainty, acceptable discrepancies, and comparison with experimental effects to be reported.

2 Levels of correspondence between synthetic and human responses

  • Validity claims depend on whether a study is exploratory, substantive, or predictive, because representativeness means different things across applications.The framework uses levels of correspondence to license different kinds of validity claim.Source: §2 · pp. 3–4

2.1 The criterion: Behaviour

  • The behavioural criterion asks whether personas predict the represented group’s observed behaviour; Table 1 separates this criterion from diagnostics L0–L3.The behavioural evidence may be expressive, physical or economic, or a high-stakes decision, and should use held-out outcomes unavailable during persona construction or tuning.Source: §2 · p. 4, §2.1 · p. 4, Table 1 · p. 5
  • Validation should report uncertainty, specify an acceptable discrepancy before comparison, and assess experimental effects when judging suitability for resource allocation.When self-report is available, evaluators should compare synthetic and human intention–behaviour gaps; baselines and within-stratum calibration are also required.Source: §2 · p. 4, §2.1 · p. 5

2.2 Diagnostic levels

The four diagnostic levels locate different forms of representation failure, from aggregate location to relationships among variables. Evidence cited by the paper highlights compressed variation and distorted structure, while response-process validation concerns whether synthetic respondents react to context as humans do.

Location (L0)

  • At L0, evaluators compare a synthetic mean or proportion with the human population statistic relative to its sampling error.The claim is limited to approximating that aggregate for the item, population, model, and prompt; it says nothing about variation, associations, or intervention effects.Source: Location (L0) · pp. 5–6

Dispersion (L1)

  • At L1, synthetic samples often compress variance and underrepresent distribution tails despite matching the human mean.The paper calls this distribution collapse or flattening: reducing a group’s diversity to a single representative position.Source: Dispersion (L1) · p. 6

Response process (L2)

  • At L2, validation tests whether synthetic respondents react to question wording, item order, and response format as humans do.The cited evidence says LLMs fail to reproduce human response biases and are mainly dominated by answer-position and label effects.Source: Response process (L2) · p. 6

Structure (L3)

  • At L3, evaluators test whether relationships among variables match human data, and a structure cleaner than the human one counts as failure.Studies cited report distorted covariance structures, inflated belief–intention correlations, and associations reproduced for the wrong reasons.Source: Structure (L3) · p. 6

2.3 Experimental correspondence

Experimental correspondence tests whether manipulations affect synthetic samples as they do human respondents. Within-persona contrasts can probe profile-level effects, but their interpretation and validation have important limits.

2.3 Experimental correspondence

  • Experimental correspondence (E) measures whether a manipulation affects a synthetic sample as it would real respondents.Effects may be assessed through outcomes such as stated support or uptake.Source: §2.3 · p. 6
  • Duplicating a persona across conditions enables within-persona contrasts that estimate effects for a profile, while correspondence to human effects requires separate validation.These contrasts can reveal heterogeneity within the simulation.Source: §2.3 · p. 7
  • Prompt-induced user drift can confound contrasts, stochastic outputs require a same-condition null distribution, and profile-level estimates are not observed individual treatment effects.Human estimates are also noisy and subject to publication bias, so the authors recommend limiting claims to rough treatment-effect magnitudes.Source: §2.3 · p. 7
  • A proposed validation compares within-persona effects with effects from human samples matched to the same persona definitions, including average effects and subgroup-effect ordering.Systematic divergence becomes the object of study.Source: §2.3 · p. 7

2.4 What each use requires

  • Exploratory use requires L0–L2 correspondence within human bands for each relevant subgroup; substitutive use additionally requires L3 invariance, benchmark E correspondence, and a human anchor.Predictive use and resource allocation require C at or above declared baselines for relevant subgroups; groups lacking behavioural evidence or invariance cannot inherit aggregate conclusions.Source: §2.4 · p. 7

3 Target group validity

Aggregate accuracy can conceal subgroup misrepresentation, so the authors make subgroup-level validity claims part of validation. They connect subgroup error, process documentation, and preserved within-group variation to measurable justice dimensions.

3.1 Aggregate accuracy hides subgroup failure

  • A population-level average can hide errors that offset across subgroups, leaving a synthetic sample wrong for every subgroup despite reproducing the overall mean.Models may learn population statistics without group structure, have weaker coverage of specific populations, or fail to represent incapacity that contributes to non-uptake.Source: §3.1 · pp. 7–8
  • Researchers should define decision-relevant subgroups before validation, examine important intersections when sample sizes allow, and report worst-group performance alongside averages.The subgroup definitions should reflect the decision context rather than be selected from the synthetic data.Source: §3.1 · p. 8
  • Subgroup checks map to validity levels: L0 assesses point estimates, L1 distributional distance, L3 measurement invariance, and E subgroup-specific treatment effects.L1 can reveal when a group mean is reproduced but within-group disagreement is flattened.Source: §3.1 · p. 8
  • Subgroup validity claims are bounded by human benchmark size; with scarce human data, the authors suggest manual rectification, while guarantees apply to corrected subgroup estimates.A subgroup of thirty human respondents cannot support a strong claim in either direction.Source: §3.1 · p. 8

3.2 Justice dimensions as measurable quantities

  • Distributional justice maps to subgroup error, procedural justice to documentation, and recognition justice to preserved within-group variance.Documentation includes model identity and version, prompts, sampling settings, persona construction, and the human benchmark; these indicators do not establish that deployment is necessarily fair.Source: §3.2 · pp. 8–9

4 Application: Charging behaviour in the energy transition

The charging-tariff example applies the validation framework to observed charging behaviour rather than stated willingness. It emphasizes subgroup-specific constraints and treats within-persona effects as hypotheses for meter-data testing.

4 Application: Charging behaviour in the energy transition

  • For a time-of-use tariff, the consequential outcome is household charging start time, assessed against meter data rather than stated willingness to shift.The framework checks prediction, baseline shares, spread of start times, characteristic associations, and tariff effects at distinct validity levels.Source: §4 · p. 9
  • Subgroups should track behavioural constraints such as access to private charging, because aggregate shifts may overstate benefits for low-income and renting households.Such overstatement could support a tariff whose benefits accrue to households with flexibility capital.Source: §4 · p. 9
  • Within-persona tariff-versus-control contrasts can estimate shifts by household type as hypotheses that meter data can test in a pre-test.A framework-structured report makes missing comparisons across validation levels and subgroups visible.Source: §4 · p. 9

5 Reporting checklist

The paper closes with a reporting checklist for studies making claims about human behaviour using synthetic samples. It presents inconsistent reporting, rather than disagreement about standards, as the main obstacle to comparing validation studies.

5 Reporting checklist

  • The paper provides a reporting checklist for studies using synthetic samples to make claims about human behaviour.The checklist is presented in Table 2.Source: §5 · p. 9, Table 2 · p. 10
  • The authors identify inconsistent reporting as the main obstacle to comparing validation studies, rather than disagreement about standards.They therefore organize the reporting items as a checklist.Source: §5 · p. 9
  • Studies that use synthetic samples to make claims about human behaviour should report the items listed in Table 2.The passage identifies the checklist's intended users and purpose but does not enumerate its items.Source: Table 2 · p. 10

6 Limitations

The framework faces limits from training-data contamination, sparse subgroup benchmarks, imperfect behavioural criteria, and changing model versions. The authors also identify theoretical and practical questions for future validation.

6 Limitations

  • Training-data contamination can make a model reproduce classic experimental results without evidencing the underlying process, while subgroup analysis depends on human benchmark availability.Held-out experiments or models trained before the event are suggested as likely mitigations; the authors also point to human review when systems operate at the edges of training data or outside their competence.Source: §6 · p. 10
  • Behavioural criteria are imperfect, and model updates can quickly age validation evidence, making model versions and access dates important to report.Public posting is performed behaviour, administrative records reflect the systems producing them, and incentivised laboratory tasks are externally mandated and observed.Source: §6 · p. 10
  • Attitude–behaviour theories may help identify when verbal correspondence suffices, while situated agents and interaction patterns remain questions for later work.The authors propose examining agents in situations and comparing coordination patterns with those observed between people.Source: §6 · pp. 10–11

7 Conclusion

The conclusion argues that validation should target consequential behaviour and report results separately for groups defined before analysis. It proposes a framework and four testable predictions to make such claims more comparable and transparent.

7 Conclusion

  • The authors argue that consequential-behaviour claims should be checked against observed behaviour and reported separately within groups defined before analysis.They frame this as a response to evidence that often tests questionnaire-answer correspondence despite questions about consequential behavioural decisions.Source: §7 · p. 11
  • The framework predicts that aggregate fidelity will overstate worst-subgroup fidelity, with a larger gap as training-data coverage falls.It also predicts that synthetic samples will reproduce human treatment-effect direction more reliably than effect size, and that stated-outcome correspondence will not transfer to behavioural outcomes without loss.Source: §7 · p. 11
  • The framework combines predictive-validity claims, subgroup reporting, measurable justice dimensions, and within-persona counterfactual design that itself requires validation.Its claims are intended to be comparable and interpretations transparent.Source: §7 · p. 11

Declarations

The declarations describe the project’s funding and competing interests, state that no new participant data or analysed data were involved, and list author contributions.

Funding.

  • Artificial Societies Ltd. funded the working time of the decision-context authors, but the organisation had no role in specifying validation criteria or decision rules and no authority over publication.The framework is described as population- and vendor-agnostic, with final authority over content and submission held by the decision-context authors.Source: Funding · p. 11

Competing interests.

  • Several co-authors are founders or employees of Artificial Societies Ltd., which develops commercial synthetic-audience simulations evaluated by the framework.The framework requires disclosure of validator independence.Source: Competing interests · p. 11

Ethics approval.

  • No ethics approval was required because the article proposes a framework and reports no new human-participant data.Source: Ethics approval · p. 11

Availability of data and methods.

Authors’ contributions.

  • All authors contributed to conceptualisation and review and editing, while methodology, original drafting, and funding acquisition were assigned to named subsets.Methodology: L.d.M., F.K., and C.K.; original draft: F.K., C.K., and L.d.M.; funding acquisition: F.K., L.d.M., and J.K.H.Source: Authors’ contributions · pp. 11–12
Loading 2609.27690v2…