Private AI for researchers: interview transcripts, IRBs, and the Common Rule

The Morse teamPublished October 26, 20269 min read

As of October 2026, private AI for researchers means keeping participants' words, from interview transcripts and survey answers to field notes and coded datasets, out of AI tools your IRB protocol, consent form, and university data rules don't cover. The university rules we read let identifiable research data go only into tools approved for its risk level, and for NIH-funded studies a federal statute bars disclosing identifiable, sensitive information to "any other person not connected with the research."

We read the Common Rule (45 CFR 46), NIH's Certificate of Confidentiality policy and the statute behind it, NIH's notice on controlled-access data, the HHS advisory committee's recommendations on AI, the AI guidance of the University of Michigan, Harvard Medical School, Stanford, and the University of Wisconsin-La Crosse, and OpenAI's and Otter's own pages, all on October 26, 2026. This isn't legal advice; your IRB and your institution's data office decide what your study may do.

Can I use ChatGPT for qualitative research?

Yes, for work your protocol allows, and usually not with identifiable transcripts in a personal account. The University of Michigan's IRB guidance says, "Do not submit personally identifiable or sensitive information to third-party or public AI tools," and such data "may only be used with AI tools that are approved for the applicable data classification level and have appropriate data protection agreements in place." It adds that study teams "may be required to use U-M–supported AI tools," and since May 2026 its IRB application asks whether and how a study uses AI (University of Michigan, read in a July 22, 2026 archive copy because the live page blocked us).

University licenses matter because OpenAI says that by default it doesn't train on ChatGPT Edu or on "ChatGPT for Academic Researchers workspaces," unlike its Free, Plus, and Pro plans (OpenAI, an archived copy from October 19, 2026, because help.openai.com blocked us); the consumer settings are in does ChatGPT train on your data.

Does putting participant data into an AI tool count as a disclosure?

Not in so many words, but the Common Rule's tests are written in terms of disclosure. The exemption most interview and survey studies rely on, 45 CFR 46.104(d)(2), turns on what would happen if responses left the study. One route requires that "any disclosure of the human subjects' responses outside the research would not reasonably place the subjects at risk of criminal or civil liability or be damaging to the subjects' financial standing, employability, educational advancement, or reputation." Studies that record identities under the third route get a limited IRB review of privacy instead (eCFR).

Two more lines in the same rule apply to every AI upload. To approve research, an IRB must find "adequate provisions to protect the privacy of subjects and to maintain the confidentiality of data" (46.111(a)(7)). The consent form must include "a statement describing the extent, if any, to which confidentiality of records identifying the subject will be maintained" (46.116(b)(5)). A transcript sent to an outside AI service changes that answer, which is why IRBs ask about it.

Does a Certificate of Confidentiality cover AI tools?

It covers the data, and it restricts who you may give it to. Since October 1, 2017, NIH has issued Certificates of Confidentiality automatically to NIH-funded research that collects or uses identifiable, sensitive information, including most exempt human subjects research, except studies that record no identities (NIH). The statute says the holder "shall not disclose or provide to any other person not connected with the research the name of such an individual or any information, document, or biospecimen that contains identifiable, sensitive information" (42 U.S.C. 241(d)).

The exceptions are narrow: disclosure required by law (outside legal proceedings), for the person's medical treatment with their consent, with the person's consent, or for other research that follows the federal human subjects rules. NIH requires recipients to make sure any investigator or institution that gets a copy understands it is bound too (NIH, page updated August 10, 2024).

Neither the statute nor NIH's policy mentions AI tools. On its face, though, an AI company that receives an identifiable transcript is a person outside the research. The IRB guides that ranked highest for this topic on October 26, 2026, from Lehigh, UW-La Crosse, Columbia's Teachers College, and Michigan, don't mention Certificates. If your study is NIH-funded and identifiable, ask your IRB how the Certificate applies before any transcript goes into an AI tool.

Is de-identified data safe to put into AI?

Safer, and not settled. The Common Rule calls information identifiable when the subject's identity "is or may readily be ascertained by the investigator or associated with the information," and it requires agencies to reexamine that meaning, with experts "in data matching and re-identification," at least every four years (eCFR). The HHS Secretary's Advisory Committee on Human Research Protections said in recommendations approved October 19, 2022 that AI and big data "expose the limits of the traditional concept of identifiability," because combined datasets let researchers "learn or infer information about individuals that they may not have knowingly disclosed" (SACHRP, read in a January 2, 2026 archive copy because hhs.gov blocked us). The recommendations are advisory, not HHS policy. Michigan's IRB warns that AI systems can "re-identify data previously considered de-identified."

Data from NIH's controlled-access repositories has a firmer rule. NIH's NOT-OD-25-081 says sharing it "with public generative AI tools (e.g., third party tools) via prompts" breaks the Data Use Certification (NIH). We cover that notice in can you use AI for peer review.

Which AI tools does a university approve for research data?

The ones it licenses, for the data levels it names. Here is what three universities' own pages said on October 26, 2026.

Sources: Harvard Medical School, Stanford's approved AI services and its responsible AI page, and Michigan (an archived copy from July 22, 2026, because the live page blocked us). Signing up for ChatGPT with your university email doesn't put you on the approved list at Stanford, and a tool cleared for your data's level can still be off limits under the agreement you signed to get the data.

Can you use AI transcription for IRB interviews?

Only a service your IRB has approved for the recording's risk level. In the table, Harvard Medical School clears meeting transcription up to Level 2 and Stanford up to Moderate Risk. Otter's privacy policy, effective June 16, 2026, lists "training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)" among its uses, and says data labeling providers "use the data we share to create training and evaluation data" (Otter). A recorded voice can identify a speaker on its own, so ask your IRB which level your recordings fall under before a meeting assistant transcribes them. Reporters face a version of the same problem, which we cover in private AI for journalists.

What should a consent form say about AI?

What the tool gets, what it does with it, and whether it can be taken back. The University of Wisconsin-La Crosse's IRB tip sheet gives sample wording for a study that sends de-identified survey answers to ChatGPT: "Once your deidentified survey responses are submitted to the AI tool, they can't be individually removed and may be used to train or improve the tool." The same sheet warns that "even if this default setting is turned off, there also can be content monitoring that happens by employees" (UW-La Crosse). Michigan asks consent forms to say "whether participant data will be used to train or improve AI systems" and whether data will be shared with "third-party or external AI tools."

How Morse keeps participants' words off its servers

Every rule above turns on a copy of participants' words leaving the research team for a company that keeps it, may train on it, or both. Morse is built not to keep that copy. Your chat history is stored only in your browser, encrypted. Morse's servers don't keep messages or log message content, and Morse doesn't train AI on your content, sell your data, or show ads.

That doesn't replace your IRB. None of the university lists above names Morse, so if your protocol or consent form doesn't cover an outside AI service, it doesn't cover Morse, and identifiable data under a Certificate of Confidentiality waits for your IRB's answer. Where your protocol allows an outside tool, pick the mode per chat:

• E2EE, for passages of transcripts or field notes no one else should read. Your browser checks the secure enclave's proof that it's genuine, then encrypts each message so only that enclave can read it, and Morse's servers relay text they can't read. The enclaves are run by third-party providers, not by Morse, and the guarantee rests on the enclave hardware and attestation working as designed. E2EE runs open-weight models and has no web search. It comes with Pro, at $19 a month.
• Private, for coding de-identified excerpts or drafting a codebook. It's the default for a new chat. Open-weight models like DeepSeek, Kimi, and GLM run under the zero data retention agreements Morse has with its inference, search, and other providers, so they keep nothing; where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't (what zero data retention means).
• Anonymous, for methods questions to GPT, Claude, or Gemini. Morse's proxy strips your account, email, and IP address, so the provider sees the message without knowing it's you. It may keep it for a limited time to check for abuse, under its own policy, so ask how to run an inter-rater check in general terms and leave the data out.

In every mode, your history locks when you sign out, and unlocking it takes a key that never leaves your device plus one our server gives you at sign-in. Because we don't store chats, we usually have very little to give authorities, as our privacy policy explains. Clearing your browser or deleting your account deletes your chats, and Morse can't recover them, so keep transcripts and analysis files where your data management plan says they live.

So where can your participant data go?

• Identifiable transcripts or recordings: only tools approved for their risk level and named in your protocol, and for NIH-funded work, only after your IRB says how the Certificate of Confidentiality applies.
• De-identified excerpts: strip names, places, dates, and file metadata, then use the tool your IRB approved. In Morse, a Private chat runs on providers that keep nothing, and every chat's history stays in your browser instead of on a company's servers.
• Controlled-access genomic data: not into public or third-party AI tools, Morse included.
• Methods questions with no data in them: any tool, and in Morse, Anonymous mode sends the question without the provider knowing it's you.

You can try a chat on the Morse home page without an account, or start free and move to Pro when your transcripts need E2EE.