When Is the Ethical Decision Simply Not to Use AI in a Research Study?

What if the Most Responsible Way to Use AI in Research Is Not to Use It?

You are conducting a qualitative study involving university instructors. You have collected interviews about their professional experiences, including sensitive descriptions of institutional disagreements and decisions concerning generative AI.

You discover an AI tool that could summarize the transcripts, generate preliminary codes, and help organize your analysis. The potential time savings are substantial.

However, the available service has not been authorized to process identifiable participant information. Removing names would not eliminate all identification risks because several instructors occupy distinctive institutional positions.

You could attempt to remove more information, but doing so might eliminate the contextual details necessary to interpret their experiences. Alternatively, you could upload the transcripts and hope that the provider's general privacy assurances are sufficient.

Neither option seems methodologically or ethically satisfactory.

At what point should a researcher stop searching for ways to justify AI use and recognize that the more appropriate decision may be to conduct the task without it?

Do Not Use AI When the Proposed Activity Cannot Be Adequately Justified or Safeguarded

Recognizing When AI Should Not Be Used in Research

Responsible AI Use Includes the Decision Not to Use AI

Discussions of responsible AI often emphasize safeguards, transparency, human oversight, and technical controls. These measures can be valuable, but they do not establish that every proposed AI application should proceed.

Sometimes the intended benefit is limited, the necessary information is too sensitive for the available processing arrangement, or the system cannot produce outputs suitable for the research purpose.

For example, a researcher may want AI to classify participants according to psychological characteristics inferred from interview transcripts. Even if the system can generate plausible labels, the study may lack the necessary measurement validity evidence and ethical authorization.

In that situation, adding a disclaimer that AI was used would not resolve the underlying problem.

The NIST Artificial Intelligence Risk Management Framework (2023) recognizes that responses to AI risks may include avoiding a particular application when risks are unacceptable or cannot be adequately managed.

Risk avoidance is therefore a legitimate governance decision, not necessarily evidence that the researcher has failed to adopt an available technology.

When Is Avoiding AI Different From Rejecting AI Entirely?

Researchers should distinguish rejecting a particular AI-assisted procedure from adopting a general prohibition against generative AI.

For example, an external AI service may be inappropriate for processing identifiable interviews but suitable for helping draft nonconfidential methodological explanations.

Similarly, a general-purpose chatbot may be unsuitable for generating validated psychological scores while remaining useful for suggesting alternative questionnaire wording.

The first decision can follow a careful assessment of evidence and risk. The second requires a broader justification and should not be assumed to be the only responsible position.

What Circumstances May Justify Not Using AI?

Several circumstances may make a proposed AI application inappropriate or require researchers to defer its use until relevant concerns are resolved.

These examples do not establish universal prohibitions. They identify circumstances in which researchers should question whether the proposed application is appropriate and whether the necessary conditions can realistically be satisfied.

Do Not Use AI When Necessary Authorization Is Missing

A researcher should not submit protected research information to an AI service when the processing requires authorization that has not been obtained.

For example, a study may have ethics approval for researcher-led analysis using institutionally managed systems. The researcher subsequently considers uploading complete transcripts to a consumer-facing AI platform.

The original approval may not cover the new processing arrangement.

Even if the researcher intends to delete the conversation afterward, the initial transfer may already involve processing by an additional provider.

Researchers should follow the applicable institutional procedures before implementing the change.

If an appropriately authorized system is unavailable, the researcher may need to continue with the original analytical method.

The requirements for introducing AI after ethics approval should be resolved before protected data are transferred.

When Confidentiality Cannot Be Adequately Protected

Some research datasets contain information that is difficult to de-identify without substantially changing their meaning.

For example, an interview with a senior university administrator may describe a distinctive institutional event, a unique professional responsibility, or a disagreement involving a small number of identifiable colleagues.

Removing the participant's name may not prevent deductive identification.

Removing every contextual detail may make the transcript unsuitable for the intended qualitative analysis.

In such circumstances, the researcher should not assume that the need for context justifies submitting the complete transcript to an unauthorized service.

A protected institutional processing environment may provide an alternative. If no suitable arrangement is available, researcher-led analysis may be preferable.

The ethical problem is not that the data are too complicated for AI. It is that the proposed processing arrangement may be unsuitable for the information the research legitimately needs.

Do Not Use AI to Circumvent Participant Choices

Participants may decline to disclose particular characteristics or agree to provide information under specified conditions.

For example, an interview participant may choose not to answer a question about a sensitive personal circumstance.

Using AI to infer that characteristic from other responses could undermine the practical significance of the participant's decision, particularly when the inference is unnecessary or outside the approved research purpose.

However, researchers should not assume that every derived variable is prohibited merely because it was not directly reported. Some studies legitimately involve estimation or inference under appropriately authorized designs.

The important distinction concerns necessity, participant information, methodological justification, and the applicable consent and data-protection arrangements.

Researchers should not use AI to reconstruct unreported participant characteristics simply because the technology can produce a prediction.

When AI Outputs Are Not Valid for the Intended Research Purpose

Generative AI may produce convincing outputs without sufficient evidence that they are suitable for the intended measurement or analytical task.

For example, a researcher might ask a model to assign academic self-efficacy scores based on students' written reflections.

The model may generate numerical ratings and explanations, but this does not establish that the scores measure academic self-efficacy accurately.

Measurement requires an appropriate construct definition, operationalization, and validity evidence supporting the intended interpretation and use.

If these requirements cannot be satisfied, the researcher should not use the AI-generated scores as though they came from a validated instrument.

AI may still assist with developing candidate items or examining wording, provided the researcher follows appropriate instrument development and evaluation procedures.

The decision not to use AI for measurement does not necessarily prohibit its use for less consequential preparatory tasks.

Why Human Oversight Does Not Automatically Solve Every Problem

Researchers sometimes justify AI use by stating that a human will review all outputs.

Human oversight can reduce certain risks, but its effectiveness depends on what the reviewer can actually verify.

For example, a researcher may be able to compare an AI-generated interview summary with the original transcript.

However, verifying an AI-generated psychological diagnosis from ordinary teaching interviews may be impossible because the data were not collected to establish that diagnosis.

Similarly, human review does not retroactively authorize an improper transfer of confidential participant information.

Researchers should distinguish problems that can be addressed through output verification from those requiring different data, additional authorization, or rejection of the proposed procedure.

Human oversight is a safeguard, not a universal remedy.

When the Expected Benefit Is Too Small to Justify Additional Risk

Generative AI may offer practical advantages that are insufficient to justify the risks introduced by a particular workflow.

For example, a researcher may save several hours by uploading identifiable transcripts to an external service. However, the same analytical task may be completed through an approved local procedure without exposing participant information to additional processing arrangements.

The researcher should distinguish operational convenience from direct participant benefit or scientific value.

Under frameworks such as the Belmont Report and the United States Common Rule, ethical evaluation considers the relationship between foreseeable risks and anticipated benefits or the value of the knowledge expected from the research.

These frameworks do not imply that any amount of efficiency can compensate for otherwise unacceptable processing.

Researchers should examine how AI changes the risk-benefit balance of the protocol before deciding whether the proposed advantage justifies the additional exposure.

When AI May Compromise the Meaning of Qualitative Data

Qualitative research may depend on linguistic nuance, contextual interpretation, and the relationship between participants' statements and their circumstances.

AI-generated summaries can be useful, but they may overlook ambiguity, normalize distinctive expressions, or introduce categories that are not supported by participants' accounts.

For example, a model may interpret an instructor's decision not to use AI as resistance to innovation, even when the participant describes a considered pedagogical judgment.

Such errors may be manageable when AI suggestions remain provisional and the researcher can examine the source material.

However, if the intended workflow requires the model to replace essential interpretive decisions that cannot be adequately reviewed, the procedure may be methodologically inappropriate.

Researchers should not use AI merely to produce the appearance of systematic analysis when the resulting interpretations cannot be justified.

Do Not Use AI When It Creates Unnecessary Sensitive Inferences

AI systems may generate personal characteristics beyond those explicitly collected for the study.

For example, a model asked to analyze instructors' experiences with generative AI may infer their psychological states, personal beliefs, or professional competence.

These inferences may be unsupported and unrelated to the research questions.

If the information is unnecessary, researchers should not generate it merely because it may appear analytically interesting.

Where the inference is sensitive or linked to identifiable individuals, additional privacy and fairness concerns may arise.

The researcher should consider whether sensitive AI-generated inferences fall outside the approved purpose and whether the workflow should be restricted or abandoned.

When AI Creates Unacceptable Discrimination or Stigmatization Risks

AI-generated classifications may reproduce stereotypes or attach unsupported judgments to participants and groups.

For example, a system may label instructors who rarely use AI as technologically incompetent or resistant to change.

Such categories may be inaccurate and could contribute to stigmatization, particularly when findings concern small or identifiable professional communities.

Researchers should first determine whether the categories can be revised and supported through defensible analysis.

However, if the proposed AI system systematically generates unreliable or unfair classifications that cannot be adequately managed, its use for that purpose should be reconsidered.

The objective is not to remove every unfavorable finding. It is to avoid unsupported discrimination or stigmatization risks while preserving evidence-supported interpretations.

Why More Safeguards Are Not Always the Best Answer

Researchers may respond to AI-related risks by adding more controls, additional reviewers, or increasingly complicated data preparation procedures.

These measures may be appropriate when they address identifiable risks without undermining the research purpose.

However, there may be a point at which the proposed AI workflow becomes unnecessarily complex or still leaves significant unresolved concerns.

For example, removing all potentially identifying context from a qualitative transcript may make the material analytically unusable. Adding multiple layers of AI-generated reconstruction would not necessarily restore the original meaning.

Researchers should compare the proposed safeguards with the simpler alternative of conducting the analysis through the originally approved procedure.

Responsible research does not require preserving an AI workflow at any cost.

How Can Researchers Compare AI With Non-AI Alternatives?

A practical assessment should examine whether the intended research purpose can be achieved through another method with fewer or more manageable risks.

The alternatives are illustrative. Their suitability depends on the research design, available resources, institutional requirements, and the nature of the information.

A non-AI method is not automatically free of risk, but it may avoid particular risks introduced by the proposed AI processing arrangement.

How Should You Prompt AI to Evaluate Whether It Should Be Used?

AI can help organize a preliminary comparison of possible workflows. However, asking an AI system to judge its own suitability should not replace independent methodological or institutional evaluation.

This prompt can support preliminary reflection, but the final decision should be based on verified information and the applicable institutional requirements.

Researchers should not submit protected participant data to an unauthorized AI system merely to ask whether that system is safe to use.

Who Should Decide That AI Will Not Be Used?

Researchers should evaluate the methodological necessity and suitability of AI within their study, but some decisions require institutional authority.

For example, a data protection officer may determine that the proposed processing arrangement does not satisfy applicable institutional requirements, while an ethics committee may conclude that the participant-related risks are unacceptable under the submitted protocol.

Methodological specialists may identify that an AI-generated measure lacks sufficient validity evidence for its intended use.

These assessments address different concerns and should be coordinated according to institutional procedures.

Researchers should not attempt to override an unresolved requirement simply because another office has provided a favorable assessment.

The question of who should evaluate AI-specific ethical risks depends on the expertise and authority required for the proposed activity.

Can Researchers Reconsider AI Use Later?

Yes. A decision not to use a particular AI system does not necessarily prohibit reconsideration when circumstances change.

For example, an institution may later provide an authorized processing environment with suitable contractual and technical safeguards.

Additional methodological evidence may also become available concerning the suitability of an AI-assisted analytical procedure.

However, researchers should reassess the proposed workflow under the conditions that actually exist at the time rather than assuming that an earlier rejection or approval applies indefinitely.

Where the study has already received ethics approval, consequential changes should follow the applicable post-approval procedures.

Ethical decisions can be revisited when relevant evidence or safeguards change, but they should not be reversed merely because a tool becomes more popular or convenient.

How Should Researchers Document a Decision Not to Use AI?

Not every decision to avoid an AI tool requires extensive documentation.

However, when the decision concerns a consequential methodological or ethical issue, a concise record may be useful.

The researcher can document the proposed task, anticipated benefit, identified concerns, alternatives considered, and reason for retaining a different procedure.

For example, a research team may record that an external AI service was not used for transcript analysis because the available arrangement was not authorized for identifiable participant data.

Such documentation can support methodological transparency and explain why the study used a particular analytical approach.

Researchers should avoid presenting a rejected AI workflow as though it had been implemented or implying that the decision establishes a universal prohibition against similar tools.

Choosing Researcher-Led Analysis Instead of an Unapproved AI Service

The decision does not establish that AI-assisted qualitative analysis is inherently unethical. It reflects the circumstances of the particular study, including the sensitivity of its data, the absence of an authorized processing arrangement, and the availability of a defensible alternative.

In this situation, choosing not to use AI preserves both the confidentiality conditions and the methodological context necessary for the research.

Common Misconceptions About Deciding Not to Use AI in Research

Keep the Option Not to Use AI Available Throughout the Research Process

When considering generative AI, begin with the research purpose rather than the tool's capabilities.

Determine what the proposed application would contribute, what information it requires, and whether its risks can be adequately managed within the approved research arrangements.

Where uncertainty concerns institutional authorization, seek the appropriate determination before implementing the activity. Where the problem concerns methodological validity, obtain suitable evidence rather than relying on the model's apparent confidence.

The objective is not to avoid AI unnecessarily. It is to avoid allowing the availability of a tool to determine what the research should do.

Before Deciding Whether to Use Generative AI in a Research Study

Frequently Asked Questions About When Not to Use AI in Research

Responsible Research Does Not Require AI to Be Used Everywhere

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.