The situation
A research coordinator at a hospital research institute is three weeks behind on a qualitative study of patients living with chronic pain. There are forty interview transcripts to code. She pastes the first ten into a personal ChatGPT account and asks for recurring themes. The output is useful, and the backlog starts to shrink.
The transcripts include first names, the clinics participants attend, a spouse's job, a child's diagnosis. The protocol the research ethics board (REB) approved said data would be stored on an encrypted institutional server and accessed only by the study team. The consent form told participants the same thing. Nobody on the team thinks of the chatbot as a disclosure; it feels like software on her desk. The principal investigator's question, when she finds out, is whether the study is still within its approval.
What the rules actually say
The Tri-Council Policy Statement, TCPS 2 (2022), sets the ethics framework for research involving humans in Canada. Compliance is a condition of funding from CIHR, NSERC and SSHRC for researchers and their institutions, and other organizations may adopt it. It does not mention generative AI. Its privacy and review rules apply to any tool that handles participant information.
What researchers promise the REB
Chapter 5 sets out the duties that matter here:
- Article 5.1: researchers must safeguard information entrusted to them and not misuse or wrongfully disclose it; institutions must support them.
- Article 5.2: researchers must describe their confidentiality measures in their REB application and in the consent process.
- Article 5.3: researchers must give the REB details of how information will be safeguarded across its full life cycle (collection, use, dissemination, retention and disposal) and assess privacy risks at every stage. The REB weighs limits on use and disclosure and the risk of re-identification.
- Article 5.4: institutions where data are held must provide appropriate security safeguards.
What participants were told
Under Article 3.2, the information given to prospective participants generally includes what will be collected, for what purpose, and who will have access to information about their identity. If the consent form named the study team and an institutional server, an outside AI provider was not part of that picture.
Changing how data are handled
Article 6.16 requires researchers to submit substantive changes to approved research to the REB in a timely way. Its guidance names changes to measures that protect privacy and confidentiality as examples, and says substantive changes should not be implemented without documented REB approval, except to eliminate an immediate risk to participants. Article 6.15 asks researchers to report unanticipated issues that may increase risk to participants.
Is there AI-specific guidance?
Not at the national level yet. As of its July 2026 update, the Panel on Research Ethics' list of guidance documents did not include one on generative AI. Some institutions have filled the gap. A 2024 research ethics guide from Island Health, a BC health authority, says generative AI tools should not be used on personally identifiable research data without consent, treats participants' own words as potentially identifying, and recommends naming the specific tool and where its data are held in the consent form. That is one institution's policy, not a national rule, but it shows the direction REBs are taking. Provincial health privacy laws add their own research rules; see our guide on Quebec's health information act.
Why policies and bans fall short
Most research teams already have a data management plan and signed confidentiality agreements. Neither was written with a chatbot in mind, and "the protocol does not mention AI" is easy to misread as permission. Coordinators and graduate students under deadline pressure are exactly the people who reach for the fastest tool.
A blanket ban tends to push that work to personal laptops and phones, which is worse for participants and invisible to the institution. LayerX reported in 2025 that organizations have no visibility into about 89% of AI usage. The REB cannot review what it never hears about, and a team cannot report an issue it did not notice. Our article on keeping health information out of AI prompts covers why that content cannot be pulled back once submitted.
What a practical control looks like
These steps keep AI use inside the approval rather than around it. Your REB and institutional privacy office have the final word.
- Map where AI would help. Transcription, coding, translation, literature review, analysis code. Note which uses touch participant data and which do not.
- Amend before you use. For any use involving participant data, submit a change under Article 6.16 describing the tool, the data, where it is processed and held, and retention. Ask the REB whether consent materials need updating for new participants or existing ones.
- Prefer institution-approved tools and de-identified data. Ask your privacy or IT office which AI tools have been assessed, and agree with the REB on what de-identification means for your data type.
- Train the whole team, including students and trainees, with examples of indirect identifiers in transcripts and notes. Our guide for technology transfer offices covers the parallel problem for invention data.
- Set the incident path. If identifiable data reach an unapproved tool, report to the REB and the privacy office promptly and document what happened.
- Keep records of amendments, approvals and training so the team can show the safeguards it described are the ones it uses.
Sanitized Ai is a browser extension for Chrome, Edge and Firefox that supports steps 3 to 6. It detects personal information, health information and identifiers in prompts and file uploads to the major AI assistants, and redacts or blocks them before submission. The researcher sees a plain-language explanation of what was flagged and why, which reinforces the protocol at the moment someone is tempted to step outside it.
Once participant data are submitted to a public AI tool they cannot be recalled and become subject to the provider's terms. Administrators see a dashboard of flagged-event metadata (which tool, what type of data, which policy, when) and never the prompt content, giving the institution audit-ready records without reading anyone's prompts. Those records can help show an REB that the described safeguards operate in practice, though no tool guarantees an ethics or regulatory outcome. See our healthtech page.