03 · What You Need to Know
Why LLM-Generated Answers and Human Survey Responses Have Different Evidentiary Status
What counts as human survey data?
Human survey data consist of responses obtained from actual people through a documented survey procedure.
Depending on the research design, respondents may be recruited through probability sampling, purposive sampling, convenience sampling, or other approaches.
The resulting data reflect answers provided by those individuals under the conditions of the survey.
However, human survey data are not automatically representative or unbiased. Their interpretation depends on sampling, measurement quality, nonresponse, and other methodological considerations.
The defining feature is that the responses originated from genuine human respondents rather than being generated to resemble them.
What are LLM-generated survey responses?
LLM-generated survey responses are outputs produced by a language model when it is asked to answer questionnaire items.
Researchers may provide demographic profiles, hypothetical circumstances, or instructions to adopt a particular perspective.
For example, a model might be instructed to respond as a first-year engineering student attending online classes.
The resulting answers may appear coherent and contextually appropriate.
However, the model has not recruited a student, observed the student's circumstances, or obtained a genuine self-report.
Its output reflects a computational generation process rather than an independent human response.
How do LLM responses differ from human survey data?
| Dimension |
Human survey data |
LLM-generated responses |
| Origin |
Answers supplied by actual respondents |
Outputs generated by a computational model |
| Sampling |
Participants recruited through a specified procedure |
Responses produced through prompts and generation settings |
| Personal experience |
May reflect respondents' actual circumstances and perceptions |
Reflect learned patterns and supplied context |
| Response variability |
Arises from human differences, measurement conditions, and other factors |
Arises from model behavior, prompts, and sampling procedures |
| Sample size |
Number of genuine respondents or observations |
Number of generated records or model calls |
| Population inference |
Depends on sampling and measurement assumptions |
Requires a separately validated predictive or simulation framework |
| Scientific interpretation |
May support claims about respondents or populations |
Directly supports claims about model-generated responses |
The differences remain even when both datasets contain identical variable names and numerical scales.
Can LLM responses be analyzed using statistical software?
Yes. Statistical software generally operates on numerical or categorical values without independently verifying how those values originated.
Researchers can calculate means, correlations, regression coefficients, reliability estimates, and other statistics from generated responses.
However, computational feasibility does not establish scientific validity.
For example, a statistically significant relationship between instructor support and student engagement in an LLM-generated dataset may reflect assumptions embedded in the prompt or model.
It does not independently establish that the same relationship exists among actual university students.
The analysis must be interpreted according to the data-generating process.
Can LLM-generated responses be counted as research participants?
No. Generating 500 artificial questionnaire records does not mean that 500 people participated in a study.
The records may be described as simulated respondents, synthetic observations, or model-generated responses, depending on the design.
However, they should not be included in the count of recruited human participants.
For example, a study involving 150 actual respondents and 350 generated records does not have 500 human participants.
Researchers must distinguish empirical sample sizes from synthetic record counts throughout analysis and reporting.
The distinction between synthetic data and genuine participant data is central to accurate research documentation.
Can LLM responses establish questionnaire reliability?
Researchers can calculate reliability coefficients from LLM-generated responses, but their interpretation is limited.
For example, Cronbach's alpha calculated from synthetic Likert-scale responses describes the internal consistency of those generated values under the relevant assumptions.
It does not independently establish the reliability of scores obtained from actual human respondents.
A model may generate highly consistent answers because it associates questionnaire items with the same underlying concept.
Such consistency may reflect the generation procedure rather than authentic human response processes.
Evidence of score reliability in a target population generally requires appropriate empirical evaluation.
Can LLM-generated data validate a questionnaire?
Not as a substitute for the relevant human evidence.
Questionnaire validation may involve examining content, response processes, internal structure, relationships with other variables, and consequences of score interpretation.
LLM-generated responses may help test scoring procedures or explore whether items are likely to produce particular computational patterns.
However, they cannot independently establish how actual respondents interpret the questions.
For example, a model may interpret an ambiguous questionnaire item consistently even when real students interpret it differently.
Human cognitive interviews, pilot testing, and empirical psychometric analyses may therefore remain necessary.
The appropriate evidence depends on the intended interpretation and use of the questionnaire scores.
Can LLM responses produce valid factor analysis results?
Factor analysis can be performed on artificial response matrices, but the resulting factor structure describes the generated data.
For example, an LLM instructed to produce responses reflecting three theoretical constructs may generate a dataset exhibiting approximately three factors.
That result does not establish that the questionnaire has the same structure among actual respondents.
The factor pattern may have been introduced by the generation instructions or learned associations within the model.
Researchers should distinguish testing whether factor-analysis code works from obtaining empirical evidence of a questionnaire's internal structure.
Statistical convergence and acceptable fit indices do not resolve the provenance problem.
Can LLM responses reproduce human survey patterns?
Under some conditions, yes.
Argyle and colleagues examined whether language models could reproduce selected patterns in human survey responses when conditioned on relevant demographic and political information.
Such research suggests that models may approximate particular empirical distributions.
However, successful reproduction of selected patterns does not establish general equivalence with human respondents.
Performance may vary across topics, populations, languages, cultural contexts, and model versions.
Researchers should therefore evaluate correspondence with empirical data rather than infer representativeness from plausible outputs.
Does demographic prompting make LLM survey responses representative?
No. A hypothetical demographic profile is not equivalent to recruiting a person from the corresponding population.
For example, prompting a model to answer as a working student may produce responses associated with generalized assumptions about employment, workload, and academic motivation.
Those assumptions may not adequately represent the diversity of actual working students.
Models may also reproduce stereotypes or distortions present in training material.
Researchers should evaluate demographic predictions against appropriate empirical benchmarks and examine subgroup-specific errors.
Demographic conditioning may be a component of a simulation method, but it does not independently establish population representativeness.
Are repeated LLM responses independent observations?
Not necessarily.
Repeated outputs may share model-specific patterns, common prompts, training information, and generation assumptions.
Even when responses differ, their statistical dependence may not resemble that of independently sampled human respondents.
For example, generating 2,000 questionnaire records from the same language model does not automatically justify treating the records as 2,000 independent observations from a human population.
Researchers should specify the relevant unit of analysis and account for the generation design when estimating uncertainty.
Conventional standard errors calculated under independent human-sampling assumptions may be inappropriate for model-generated records.
Can LLM-generated responses support population estimates?
They may be used to construct model-based predictions, but those predictions require independent validation.
For example, a researcher may use a language model to estimate how many students would prefer online learning over face-to-face instruction.
The resulting proportion is a model-generated estimate, not a direct measurement of the population.
Its accuracy depends on the model, prompt design, calibration, and the extent to which the simulation represents the target population.
The issue of predicting human survey responses with generative AI requires evaluating these factors against appropriate empirical evidence.
Researchers should not attach a conventional probability-sampling margin of error to artificial response counts without a justified statistical framework.
Can LLM responses be used for survey pilot testing?
Yes, for certain preliminary methodological activities.
Researchers may generate artificial responses to examine questionnaire branching, scoring rules, missing-data handling, and analytical code.
They may also ask a model to identify potentially ambiguous wording or generate hypothetical interpretations.
However, these activities do not establish how genuine respondents understand or experience the questionnaire.
Human pilot testing may still be necessary to evaluate response processes, burden, accessibility, and contextual interpretation.
The use of AI-generated participants for research pilot testing should therefore be framed as methodological preparation.
Can LLM responses be used as empirical data in a study about AI?
Yes. When the research question concerns the behavior of the language model itself, its outputs are genuine observations of that computational system.
For example, a study may examine how different models answer the same questionnaire or how demographic prompts influence generated responses.
The outputs are empirical data about model behavior because the researchers actually observed the system producing them.
However, they remain distinct from empirical observations of human respondents.
LLM responses as AI research data
Actual observations of how a specified model responds under documented testing conditions.
LLM responses as human survey data
An inappropriate classification unless the records genuinely originated from human respondents.
The same output can be legitimate evidence about AI behavior without being valid evidence about human attitudes.
What happens if LLM responses are presented as genuine human survey data?
The research record misrepresents the origin of the observations.
The United States Office of Research Integrity defines fabrication as making up data or results and recording or reporting them.
For example, a manuscript claiming that 400 students completed a questionnaire when all responses were generated by an AI model falsely represents participant recruitment.
Such conduct raises a serious research-integrity concern and may constitute fabrication under applicable misconduct standards.
Formal misconduct findings require the relevant evidentiary and culpability requirements to be satisfied.
Researchers should distinguish accidental analytical errors from knowingly misrepresenting artificial records as genuine evidence.
Does disclosing AI use make LLM survey responses acceptable?
Disclosure supports transparency, but it does not establish that artificial responses are suitable for every research question.
For example, a researcher might state that 1,000 questionnaire responses were generated by AI and then conclude that 70 percent of actual university students support a particular policy.
The disclosure identifies the data's origin, but it does not independently validate the population estimate.
Researchers must distinguish truthful reporting from methodological validity.
A simulation may be transparently described while still providing insufficient evidence for a claim about real people.
How should studies using LLM-generated survey responses be reported?
Researchers should identify the records as model-generated and describe the generation procedure.
Relevant information may include the model, available version details, prompts, hypothetical profiles, generation settings, testing dates, and number of generated outputs.
Where empirical human data are used for comparison, the manuscript should explain how those data were obtained and whether they were independent of model development.
Researchers should report relevant validation results, prompt sensitivity, subgroup limitations, and uncertainty.
Artificial response counts must remain distinct from human sample sizes.
The conclusions should identify whether the study establishes findings about model behavior, simulation performance, or empirically observed human responses.
