When Can Using AI Change the Ethical Risk of a Research Study?

Can Adding AI to an Existing Research Procedure Make the Study Ethically Different?

You have designed a qualitative study involving interviews with university instructors. The approved protocol describes how participants will be recruited, how interviews will be conducted, and how the resulting transcripts will be analyzed.

During data analysis, you discover that a generative AI tool could summarize the transcripts, identify possible patterns, and organize preliminary codes much faster than your original workflow.

The research questions remain unchanged. You are still analyzing the same interviews, and the intended findings have not changed.

But the proposed AI workflow may involve transferring participant information to an external provider, introducing automated interpretations, or generating additional classifications that were not part of the original plan.

Does the study remain ethically equivalent simply because its research objectives have not changed? Or can the introduction of AI alter the risks even when the underlying research question stays the same?

AI Can Change Research Risk Without Changing the Research Question

How Generative AI Can Alter the Ethical Risk Profile of Research

Ethical Risk Is Not Determined Only by the Research Topic

Researchers sometimes assess risk primarily by examining the topic of a study. An interview about ordinary teaching practices may appear less sensitive than one about personal trauma or professional misconduct.

However, ethical risk also depends on how information is collected, processed, stored, interpreted, and reported.

For example, interviews about instructional practices may contain identifiable information about instructors, departments, or institutional disputes. Processing those transcripts through a new external AI service could introduce confidentiality concerns even though the interview questions themselves remain unchanged.

The relevant assessment should therefore consider both the research activity and the conditions under which it is conducted.

International research ethics frameworks, including the Declaration of Helsinki for medical research involving human participants, emphasize evaluating foreseeable risks and burdens in relation to the importance of the research objectives. Its specific scope should not be automatically extended to all educational or social research, although the general principle of prospective risk assessment is broadly informative.

Distinguish a Change in Procedure From a Change in Risk

Not every procedural modification necessarily changes the ethical risk of a study in a meaningful way.

For example, using an authorized AI tool to improve the grammar of a nonconfidential methodological description may have little effect on participant-related risks.

By contrast, replacing an approved internal data analysis workflow with an external AI service that receives identifiable transcripts may introduce additional processing and disclosure concerns.

A procedural change may increase, reduce, or leave risk largely unchanged. Researchers should evaluate its actual implications rather than assume that introducing AI is inherently harmful or automatically beneficial.

What AI-Related Changes May Affect Research Risk?

Several features of an AI-assisted workflow may change the ethical characteristics of a study.

These risks are possibilities requiring assessment, not inevitable consequences of every AI system. Different service configurations may provide substantially different safeguards.

When Does an AI Provider Change the Confidentiality Risk?

Using an external AI service may introduce an additional organization or technical environment into the research data workflow.

For example, a researcher who previously analyzed transcripts on an institutionally managed computer may propose submitting them to a cloud-based generative AI platform.

The relevant considerations include the provider's contractual role, data retention, access arrangements, security measures, permitted secondary uses, and possible international transfers.

However, an external provider does not automatically create an unacceptable risk. An appropriately authorized service with suitable contractual and technical safeguards may be compatible with the research requirements.

The researcher should verify the actual arrangements rather than assume that all cloud-based AI services handle information identically.

Determining whether the provider becomes a new recipient of research data is an important part of this assessment.

AI Can Change Risk by Generating Information That Was Never Collected

Traditional data collection procedures usually specify the information researchers intend to obtain from participants.

Generative AI may expand the analysis by producing inferences beyond those original variables or questions.

For example, a study may collect interviews about instructors' decisions concerning AI use. A model might also infer their emotional stability, professional competence, or likely willingness to comply with institutional policies.

Such inferences may be unsupported, unnecessary, or sensitive. They could also introduce new risks if reported or used to characterize identifiable individuals.

The fact that an inference is generated from existing data does not necessarily mean it falls within the original purpose or authorization.

Researchers should examine sensitive AI-generated inferences and determine whether their creation and use are justified.

Can AI Increase the Risk of Discrimination or Stigmatization?

AI-generated interpretations may reproduce stereotypes or assign labels that are not adequately supported by participant evidence.

For example, a model might classify instructors who do not use generative AI as resistant to innovation or lacking technological competence.

Such classifications could misrepresent participants whose decisions reflect pedagogical preferences, limited institutional access, or concerns about the suitability of AI for particular teaching tasks.

The risk may become more consequential when results concern small professional groups or identifiable institutions.

Researchers should examine the theoretical justification, accuracy, and possible effects of AI-generated categories before incorporating them into findings.

Potential discrimination and stigmatization risks should be considered alongside privacy and confidentiality concerns rather than treated as purely technical model-performance issues.

AI-Mediated Participant Interaction May Introduce New Risks

When an AI system interacts directly with participants, the technology may influence the participant's research experience rather than merely assist the researcher behind the scenes.

For example, a study might use an AI chatbot to conduct interviews or provide personalized feedback during an educational intervention.

The system could generate inaccurate explanations, unexpected follow-up questions, or responses that are inappropriate for particular participants.

Researchers may also need to consider whether participants understand that they are interacting with AI, how their responses are stored, and whether the system behaves consistently with the approved procedures.

Human oversight, system constraints, participant information, and procedures for responding to unexpected events may be relevant safeguards.

The significance of these concerns depends on the purpose and consequences of the interaction.

Can AI Reduce Research Risks?

Potentially. AI may support safer or less intrusive research procedures when used appropriately.

For example, an authorized tool might help identify direct identifiers that should be removed from transcripts before analysis or assist researchers in screening draft interview questions for unnecessarily intrusive wording.

However, these benefits are not guaranteed. An AI system may overlook identifiers, introduce errors, or create new processing risks of its own.

Researchers should evaluate the proposed benefits alongside any new risks and compare the AI-assisted workflow with feasible alternatives.

The relevant question is whether the specific arrangement improves the overall ethical conditions of the study, not whether AI is generally more efficient.

Assess Likelihood, Severity, and Distribution of Potential Harm

Ethical risk assessment should consider more than whether a harmful outcome is theoretically possible.

Researchers should examine how likely the event may be under the proposed conditions, how serious its consequences could be, and who would bear those consequences.

For example, disclosure of an identifiable transcript involving a sensitive workplace dispute could have different consequences from disclosure of a nonconfidential methodological outline.

Likewise, an inaccurate AI classification may have limited consequences when used only as a provisional analytical suggestion but greater consequences when used to characterize identifiable participants in a published report.

Researchers should consider whether particular individuals or groups face disproportionate risks.

Risk matrices may help organize the assessment, but numerical ratings should not be treated as precise probabilities unless supported by appropriate evidence.

How Should You Prompt AI to Identify Possible Changes in Ethical Risk?

AI can assist researchers in comparing an existing procedure with a proposed AI-assisted workflow.

The prompt should provide the relevant procedures and ask for potential changes rather than assuming that AI necessarily increases risk.

This prompt can help prepare a structured assessment, but the researcher must verify provider conditions and institutional requirements independently.

AI-generated risk ratings should not be treated as authoritative determinations of ethical acceptability.

When Does a Risk Change Require Protocol Review?

A change in research risk may require institutional notification, amendment, or additional review depending on the approved protocol and applicable institutional procedures.

For example, transferring identifiable interview data to a previously unapproved external service may constitute a consequential modification.

However, a minor change in a nonconfidential drafting workflow may not require the same process.

Researchers should follow the institution's procedures for determining whether modifications require prior approval.

Where the proposed activity could materially affect participant information or welfare, it is prudent to seek a determination before implementation.

The requirements for introducing AI after ethics approval should be considered separately from the broader question of whether AI is permitted in research.

Do Not Treat Risk Assessment as a One-Time Exercise

The ethical characteristics of an AI-assisted workflow may change when the system, service configuration, data types, or research purpose changes.

For example, a researcher may initially use AI only to review nonconfidential interview questions. Later, the same tool may be used to process actual participant transcripts.

The second activity raises different considerations even though the software name remains unchanged.

Changes in provider terms, data retention settings, or institutional requirements may also affect the suitability of a processing arrangement.

Researchers should therefore reassess consequential modifications rather than assuming that one earlier approval or risk review covers all future uses.

When Might Avoiding AI Be the Safer Option?

Sometimes the proposed AI-assisted procedure may offer limited research benefit while introducing risks that are difficult to justify or manage.

For example, uploading complete identifiable transcripts to an external service may be unnecessary if the researcher only needs help organizing a coding framework that can be developed from nonconfidential descriptions.

Alternatively, an authorized institutional system or manual analytical procedure may achieve the same purpose with fewer confidentiality concerns.

Researchers should evaluate whether the expected benefit justifies the additional processing and residual risk.

Where suitable safeguards or authorization are unavailable, not using the proposed AI tool may be the more appropriate decision.

How AI Can Change the Risk Profile of an Approved Qualitative Study

The study's research questions have not changed, but its data processing arrangements and possible analytical consequences may have changed.

This illustrates why ethical risk should be evaluated according to the actual research procedures rather than assuming that an unchanged research objective means an unchanged risk profile.

Common Misconceptions About AI-Related Changes in Research Risk

Compare the Research Workflow Before and After Introducing AI

When considering generative AI, begin by documenting the original research procedure and the proposed AI-assisted alternative.

Identify what changes in the information processed, who can access it, how outputs are generated, and what consequences may follow from errors or disclosures.

Document the assessment and identify safeguards proportionate to the risks. Where an alternative approach offers similar benefits with fewer risks, consider whether it is preferable.

The objective is to understand whether AI changes the ethical conditions of the study, not merely whether it makes the research process faster.

Before Introducing AI Into an Existing Research Workflow

Frequently Asked Questions About AI and Changes in Research Ethical Risk

AI Can Change Research Risk Even When the Study's Purpose Remains Unchanged

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.