Should Researchers Use AI to Infer Characteristics That Participants Did Not Provide?

Should You Let AI Fill in Participant Characteristics That Your Study Never Collected?

You are analyzing interviews with university instructors about their experiences with generative AI. Your demographic questionnaire collected age, teaching experience, and academic discipline, but it did not ask participants about their political beliefs, psychological characteristics, or socioeconomic circumstances.

During analysis, you notice that some interview responses contain contextual clues about participants' backgrounds. You wonder whether generative AI could use those clues to predict characteristics that were not included in the original data collection.

You ask the system to infer participants' likely personality traits, levels of technological confidence, or socioeconomic circumstances from their statements.

The model produces plausible descriptions. Some are accompanied by explanations that make the predictions appear reasonable.

But these characteristics were never measured, and participants did not necessarily agree to have them inferred. Can researchers legitimately treat AI-generated predictions as additional participant information simply because the existing data make those predictions technically possible?

Generally, Do Not Infer Unreported Characteristics Without a Clear Research Justification

When AI Inference of Unreported Participant Characteristics Becomes Problematic

What Counts as an Unreported Participant Characteristic?

An unreported participant characteristic is an attribute that was not explicitly supplied by the participant or otherwise established through an authorized research procedure.

It might concern demographic background, personal beliefs, psychological characteristics, health-related circumstances, or professional attributes.

For example, a participant may explain that they rarely use generative AI because their teaching responsibilities do not require it. An AI model might interpret this as evidence of low digital competence.

The participant has reported a practice and a reason for that practice. The model has introduced an additional characteristic that was not measured.

Researchers should distinguish the original statement from the inferred attribute rather than treating both as equivalent forms of evidence.

Distinguish Participant Reports From AI Predictions

One of the most important methodological distinctions concerns the source of the information.

For example, a participant who states "I have taught for 15 years" has provided information about teaching experience.

By contrast, a model that predicts the participant's age from teaching experience has generated an inference. The prediction may be plausible, but it is not equivalent to the participant reporting their age.

The distinction remains important even when the AI system assigns a confidence score. A model-generated confidence value does not, by itself, establish that the prediction is well calibrated or valid for the intended population.

Why Might Researchers Want to Infer Missing Characteristics?

Researchers may consider AI inference when demographic information is missing, when they want to compare participant groups, or when they believe additional characteristics could explain patterns in the data.

For example, a researcher may want to examine whether instructors' AI-related practices differ by professional experience or socioeconomic circumstances.

However, the absence of a variable from the dataset does not automatically justify reconstructing it from indirect clues.

Missing information may reflect the study's original scope, participant choice, or deliberate data minimization.

Researchers should first ask whether the characteristic is necessary for answering the research questions and whether its estimation is methodologically defensible.

The desire to conduct an additional comparison is not, by itself, sufficient justification for generating potentially sensitive personal information.

What Problems Can AI Inference Introduce?

AI-generated participant characteristics may create several distinct methodological and ethical problems.

These risks should not be treated as interchangeable. An inaccurate analytical variable may undermine validity, while the unnecessary generation of sensitive personal information may raise an additional ethical concern even if the prediction happens to be correct.

Can AI Reliably Infer Demographic Characteristics From Writing?

Language and writing patterns may contain statistical information associated with demographic or social characteristics. However, this does not mean a general-purpose generative AI system can reliably determine an individual's identity or background from a short research response.

For example, writing style may vary with education, language proficiency, professional context, and the circumstances in which a response was produced.

A model might incorrectly attribute these differences to age, gender, socioeconomic position, or another characteristic.

Research by Kosinski, Stillwell, and Graepel (2013) demonstrated that certain personal attributes could be predicted from digital behavioral records. However, that finding does not establish that generative AI can accurately infer those same attributes from qualitative interview transcripts.

Researchers should not generalize predictive performance across different data types, populations, models, or analytical purposes without appropriate validation.

Why AI-Inferred Characteristics Are Not Automatically Valid Measures

A researcher may be tempted to treat an AI-generated classification as a new variable in a dataset.

For example, a model might classify participants as having low, moderate, or high technological confidence based on interview excerpts.

However, technological confidence is a construct that requires an appropriate conceptual definition and evidence supporting the intended interpretation.

Assigning categories does not establish that the system measures the construct accurately.

Researchers would need to consider the operational definition, reference measures or criteria, classification performance, uncertainty, and potential differences across relevant participant groups.

The necessary evidence depends on the intended use. A classification used for exploratory methodological research may require different evaluation from one used to make decisions about individuals.

In either case, model fluency should not be mistaken for measurement validity.

Does AI Inference Solve Missing Data Problems?

Not automatically. Researchers should distinguish AI prediction from established statistical approaches to missing data.

For example, multiple imputation uses statistical models to generate plausible values for missing observations under specified assumptions. Its appropriateness depends on the missing-data mechanism, model specification, and intended analysis.

Asking a chatbot to guess a participant's unreported characteristic from an interview transcript is not equivalent to implementing a validated multiple-imputation procedure.

Moreover, some characteristics may be absent because participants deliberately declined to disclose them. Predicting those attributes could raise ethical concerns that go beyond ordinary missing-data treatment.

Researchers should not use AI inference as an informal substitute for a defensible missing-data strategy.

What If Participants Deliberately Chose Not to Disclose a Characteristic?

A participant's decision not to answer a question may be ethically significant.

For example, a participant may decline to provide information about a personal belief or sensitive circumstance.

Using AI to infer the same characteristic from other responses could undermine the practical meaning of that choice, particularly when the research team had no independent justification or authorization for the inference.

However, the ethical implications depend on the study and applicable requirements. A declined response does not automatically prohibit every statistically justified analysis involving related information.

The researcher should examine whether the proposed inference respects the approved research purpose and the participant's reasonable expectations.

Where the characteristic is unnecessary, not generating it may be the most appropriate decision.

Can Inferred Characteristics Be Personal or Sensitive Data?

Potentially. Under data-protection frameworks such as the GDPR, information relating to an identified or identifiable person may constitute personal data regardless of whether it was directly reported or generated through inference.

Additional restrictions may apply to legally defined special categories of personal data.

In the Philippines, the Data Privacy Act of 2012 establishes definitions and processing requirements for personal and sensitive personal information.

The applicable classification depends on the nature of the inferred information, its connection to identifiable individuals, and the governing law.

Researchers should not assume that an AI-generated characteristic falls outside data-protection requirements simply because the participant never supplied it.

The related issue of AI-generated sensitive inferences concerns both the generation of such information and the consequences of retaining or using it.

Can Researchers Ever Justifiably Infer Unreported Characteristics?

Yes, in appropriately designed and authorized research. Some studies explicitly investigate whether particular characteristics can be estimated from observed information.

For example, a methodological study may evaluate whether an AI model can estimate a defined attribute from specified data and compare its predictions with independently established reference information.

Such research requires an explicit purpose, appropriate ethical authorization, and a defensible evaluation design.

Researchers should establish whether the attribute can be meaningfully operationalized, whether the available data support the prediction, and whether the proposed analysis introduces disproportionate privacy or fairness risks.

The distinction is between studying inference as a justified research objective and casually generating additional participant characteristics because the model can do so.

Even an explicitly designed inference study may be ethically inappropriate if the potential harms cannot be adequately justified or managed.

Why Accuracy Alone Does Not Establish Ethical Acceptability

Suppose a model could predict a particular personal characteristic with relatively high accuracy under a validated research design.

That performance would not automatically establish that generating the characteristic is ethically appropriate.

Researchers must still consider the purpose of processing, participant information, applicable authorization, data minimization, and possible consequences.

For example, a highly accurate inference about a sensitive personal characteristic may remain unnecessary for a study investigating teaching practices.

Technical capability and ethical entitlement are different questions.

Researchers should therefore evaluate both whether an inference can be supported and whether it should be generated for the particular research purpose.

How Can AI Inference Distort Group Comparisons?

Suppose a researcher uses AI to infer socioeconomic circumstances from interview transcripts and then compares participants' AI-use practices across the inferred categories.

If the classifications are inaccurate, the group comparisons may be misleading.

The problem may become more serious if classification errors differ across language groups, academic disciplines, or other participant characteristics.

Researchers should also consider whether the inference process uses information closely related to the outcome being analyzed, potentially introducing circularity or other forms of analytical bias.

For example, classifying participants' technological confidence from their statements about AI use and then concluding that technological confidence predicts AI use may be methodologically circular.

The inferred predictor may partly reproduce the outcome rather than provide an independent explanatory variable.

Researchers should evaluate these dependencies before treating AI-derived attributes as suitable for statistical comparisons.

How Should You Prompt AI to Review Proposed Participant Inferences?

AI can help identify methodological and ethical questions associated with a proposed inference, but it should not be asked to generate sensitive participant profiles merely to demonstrate that such profiling is possible.

This prompt supports preliminary methodological reflection without generating new personal information about actual participants.

The researcher should verify the relevant literature and obtain institutional guidance where necessary.

What Should Researchers Do With AI-Generated Characteristics They Did Not Request?

Sometimes an AI system may introduce personal interpretations during an otherwise legitimate analysis.

For example, a model asked to summarize teaching experiences may describe a participant as psychologically insecure or professionally incapable.

The researcher should distinguish the unsupported inference from the participant's actual statements.

If the characteristic is irrelevant or unjustified, it should not be incorporated into the findings or treated as verified data.

Researchers should also examine whether the output needs to be handled under applicable data management procedures, particularly if it is sensitive and linked to an identifiable individual.

Repeated unwanted inferences may indicate that the analytical workflow requires revision.

When Should the Researcher Consult the Ethics Committee?

Institutional review may be necessary when the proposed inference introduces new sensitive information, changes the approved analytical purpose, or creates additional participant risks.

For example, a study approved to investigate instructors' experiences with generative AI may later propose inferring their psychological characteristics from interview transcripts.

That activity may differ substantially from the original research purpose.

Researchers should determine whether the proposed analysis requires amendment, notification, or additional authorization before implementing it.

They should not assume that previously collected information can be used to infer any additional characteristic without further consideration.

The broader concern is whether AI changes the ethical risks of the approved study.

When AI Attempts to Infer Technological Competence From Interview Responses

The participant's limited AI use does not establish low technological competence. A person may be capable of using a technology while choosing not to use it for a particular purpose.

The example illustrates why researchers should not transform unmeasured personal characteristics into analytical variables merely because AI can assign labels to them.

Common Misconceptions About AI-Inferred Participant Characteristics

Do Not Turn Unreported Characteristics Into Research Variables Without Justification

When considering AI-generated participant characteristics, begin by asking whether the variable is necessary for the study and whether the available data can support a defensible inference.

Do not use model-generated labels merely to complete a dataset or create additional comparisons.

Where a characteristic is genuinely necessary, use a defensible research design rather than informal model guessing. Where it is unnecessary, leave it uncollected.

The researcher's responsibility is to justify why the information is needed and how it can be established, not simply demonstrate that an AI system can produce a prediction.

Before Using AI to Infer Unreported Participant Characteristics

Frequently Asked Questions About AI-Inferred Participant Characteristics

AI Can Predict an Unreported Characteristic, but That Does Not Make It a Measured Fact

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.