01 · The Question
You Already Have Ethics Approval. Can You Simply Add AI?
Your study was approved before you planned to use generative AI. Data collection is underway, or perhaps it is already complete. Now you want AI to transcribe interviews, code qualitative data, classify survey responses, analyze images, extract information from clinical records, or perform another research task.
Does that require an ethics amendment?
Not every change in software necessarily changes the ethical character of a study. But introducing AI can alter where participant data go, who processes them, what information is inferred, how confidentiality is protected, or what analytical procedures are actually being performed.
The safest approach is therefore not to decide for yourself that AI is “just another tool.” Determine whether the proposed change falls within your existing approval and follow your institution's procedure for modifications to approved research.
03 · What You Need to Know
The Relevant Question Is What AI Changes About the Approved Study
Ethics Approval Applies to the Research That Was Actually Approved
An ethics committee or institutional review board reviews a defined research plan. Depending on the study and institution, that plan may specify recruitment, consent, interventions, data collection, storage, access, confidentiality safeguards, analytical methods, external collaborators, and data-sharing arrangements.
If the actual study changes, the approved protocol and the research being conducted can begin to diverge.
Under current U.S. HHS guidance, investigators who want to modify an ongoing IRB-approved study must submit the proposed modification and obtain IRB approval before implementing it, unless the change is necessary to eliminate an apparent immediate hazard to participants. HHS also recognizes expedited review procedures for qualifying minor changes. These requirements apply within that regulatory framework and should not be presented as a universal rule for every country or institution.
The Presence of AI Alone Does Not Tell You Whether the Change Is Material
Imagine an approved quantitative study in which researchers use standard statistical software to calculate descriptive statistics. The research team later uses an institutionally approved AI coding assistant to help write R syntax using synthetic example data. No participant information enters the AI system, and the approved analysis remains unchanged.
Now compare that with an interview study in which identifiable recordings previously stored only on an institutional server are uploaded to an external AI provider for transcription.
Both involve AI. Only one obviously changes the participant-data flow.
The ethical significance lies in what changes about the research, not in the acronym attached to the software.
External Processing Is a Strong Reason to Check the Protocol
If participant data were previously accessible only to the research team but will now be transmitted to an AI provider, the change may affect confidentiality and data recipients.
The provider may function as an external processor or occupy another role under the applicable legal framework. Subprocessors and international processing may also become relevant.
Before implementing the change, determine whether the approved protocol already permits external AI processing of participant data. If it does not, the change may need ethics, privacy, security, contractual, or other institutional review.
A New AI Workflow Can Change Confidentiality Without Changing the Research Question
A common mistake is to ask only whether AI changes the scientific aims.
Suppose the original study asked the same research question and planned exactly the same thematic analysis. The only change is that interview transcripts will now be uploaded to a cloud AI system to suggest codes.
The research question has not changed. The confidentiality arrangement may have.
Ethics review is concerned with participant protections as well as scientific aims. A modification can therefore matter even when the hypotheses, research questions, and planned publication remain identical.
AI Can Also Change the Nature of the Analysis
Some AI uses go beyond replacing one software package with another. A system might infer participant characteristics that were not explicitly measured, generate predictions, classify people into risk categories, perform facial recognition, or combine variables in ways not contemplated in the approved study.
SACHRP has specifically highlighted AI-related concerns involving identifiability, privacy, bias, discrimination, new inferences, group harms, and informed consent. Its recommendations emphasize that AI research can create ethical issues not adequately captured by older assumptions about data and identifiability.
If the AI use changes what researchers are attempting to learn about participants, the modification may be more consequential than simply automating an already approved analysis.
Consent Materials May Need to Change Too
If AI introduces a material change in data handling or risk, the question is not limited to the protocol. Participant-facing materials may also need revision.
For example, a consent form might state that recordings will be transcribed by the research team and remain on institutional systems. An amendment proposing external AI transcription could make that description inaccurate.
Whether informed consent should mention the AI processing, and whether already enrolled participants need additional information or renewed consent, are determinations that should be made under the applicable ethics framework rather than improvised by the investigator.
An Amendment Does Not Automatically Mean Re-Consenting Everyone
These are separate questions.
An ethics committee may determine that a protocol modification requires review but does not materially affect information participants need to know. Alternatively, it may require revised consent for future participants, notification of existing participants, renewed consent, or another action.
Researchers should therefore avoid reasoning backward: “Re-consenting everyone would be difficult, so this cannot be an amendment.” Administrative inconvenience does not determine the ethical classification of the change.
De-Identifying the Data Can Reduce the Significance of the Change Without Automatically Eliminating It
If AI will receive genuinely anonymous information rather than identifiable participant records, privacy risks may be substantially lower. Depending on the applicable framework, some data-protection or human-subjects requirements may also change.
But the approved protocol may still contain restrictions on external services, data reuse, analysis, or sharing. Furthermore, de-identification does not by itself answer every ethical question about AI use.
Do not therefore use de-identification as a shortcut around the modification process. Describe the proposed workflow accurately and allow the appropriate authority to determine its significance.
Exempt Research Can Still Require an Institutional Determination When It Changes
Researchers sometimes assume that an exempt determination means subsequent modifications never need review.
Institutional procedures vary. HHS advises investigators to consult the appropriate institutional authority when changes to an exempt study might make it nonexempt human-subjects research.
The broader lesson applies well beyond U.S. regulation: the researcher should not unilaterally decide that a changed study remains within its previous ethics category when the change may affect the basis for that determination.
“Minor Change” Is Usually an Institutional Classification
Some frameworks permit streamlined or expedited review of minor modifications. That does not necessarily mean investigators can simply label their own AI addition “minor” and implement it.
Current HHS guidance recommends that institutions have procedures describing which minor changes can receive expedited review and how investigators submit modifications.
Follow the local process. The paperwork may be delightfully unexciting, but that is still preferable to discovering during publication, audit, or an incident investigation that the study you conducted was not the study that was approved.
Watch Out
Do not upload participant data first and ask the ethics committee afterward whether an amendment was necessary. If prior approval is required, the relevant event is the disclosure or processing itself. A later amendment cannot retroactively make an unauthorized upload part of the originally approved protocol.
Ethics Approval May Be Only One of Several Approvals Needed
An ethics committee may approve a modification while another institutional unit still prohibits the proposed AI environment. Privacy offices, information-security teams, research offices, legal counsel, data custodians, sponsors, or funders may have separate authority over particular aspects of the processing.
Ethics approval should therefore not be treated as a universal technology authorization.
For sensitive participant data, the appropriate workflow may require several aligned approvals before AI processing begins.
07 · A Quick Checklist
Before Introducing AI Into an Ongoing Study
Compare the proposed AI workflow with the approved protocol:
Identify exactly what AI will do and which participant data it will receive.
Check whether the AI changes data recipients, storage, access, processing location, retention, or confidentiality arrangements.
Determine whether the AI introduces new predictions, classifications, inferences, or analytical procedures involving participants.
Compare the new workflow with the approved protocol, consent materials, data-management plan, and confidentiality commitments.
Check institutional AI, privacy, information-security, contractual, sponsor, and data-custodian requirements.
Ask the responsible ethics authority whether the change requires amendment, expedited review, another determination, or no further review.
Obtain required approval before implementing the change with participant data.
Document the determination and any revised procedures before the AI workflow begins.
