Consent, confidentiality, governance

• Recital 33 - broad consent. It recognises that research purposes may not be fully specifiable at collection, and allows consent to "certain areas of scientific research". Useful, and not unlimited: "areas" is still a scope, so a consent taken for one area does not travel to another.
• Recital 156 - the feasibility duty. It points at checking whether the research purpose can be met without identification. This is the gate most AI workflows fail without noticing, because pasting the raw transcript is simply easier than de-identifying first. If the summary you need could have been produced from de-identified text, the recital expects you to have asked.
• Recital 159 - scope. "Scientific research purposes" is read broadly, including privately funded research. That cuts both ways for a commercial insights team: you may be inside the research framing, which means you are inside its safeguards too, not exempt from them.

The PDPA 2012 Second Schedule research exception is narrower than most people assume. Use of personal data without consent for research requires all of the following:

• the purpose cannot reasonably be accomplished without individually identifiable data;
• there is a clear public benefit;
• the results will not be used to make decisions affecting the individual;
• and the published results do not identify the individual.

Disclosure without consent adds a further condition: that it is impracticable to seek consent.

Read the third condition against a commercial reality. If your research output will drive an offer, a price, a segment assignment or a retention action aimed at that person, you are outside the research exception on its own terms. Which is why the route commercial insights teams actually rely on is usually the narrower First Schedule Part 5 Business Improvement Purposes, with its own conditions. Naming the right route in your documentation is half of surviving a review.

An honest gap, recorded on purpose rather than filled. In building this course we did not find US institutional review board guidance addressing generative AI tools specifically. That absence is worth knowing for two reasons.

First, do not let anyone tell you the ethics review question is settled by existing IRB practice, because we could not verify guidance that addresses it. Second, and more usefully: an absence of specific guidance is not permission. The general duties - consent scope, confidentiality, data minimisation, participant expectations - apply whether or not anyone has published an AI-specific page about them. Where the map is blank, you navigate by the principles and you write down what you decided and why.

A client's confidential material is somebody's manuscript. The peer-review rules exist because a reviewer holds material that was given to them in confidence, for one narrow purpose, by someone who has no relationship with the reviewer's tool vendor. That describes almost every commercial research engagement.

So the policy your team needs is not new, it is transposed:

• Other people's confidential material: prohibited in non-approved tools. Not discouraged. Prohibited, with an approved list that names the tools and the tiers.
• Your own draft writing: permitted with the Science conditions - your text, declared where required, no training on inputs.
• The vendor question that decides both: where does the data go, how long is it kept, and is it used for training. If the answer is not in a contract, treat it as unanswered.

And one thing to watch for that a5 covers in detail: some AI features in reference and research tools are routers. Paperpile's "Ask AI" sends your full-text PDF onward to a general assistant, so the confidentiality question moves downstream to a vendor you did not evaluate. Your PDFs leave by design, not by accident.

You are not publishing in Science, so translate rather than copy. A disclosure clause that works for a commercial insights function has four parts, and each is borrowed from a body above:

• What was used - the tool, the version if it matters, and the stage: search, screening, extraction, drafting, figures. The stage distinction is ICMJE's, and it is the useful one, because writing help and analysis help are different risks.
• Where it is recorded - name the section, per output type. Board paper, client report, published white paper, and journal submission may each have a different location. Decide once.
• Who verified the output - COPE's logic: because the tool cannot take responsibility, a named person must. Put the name in the document.
• The images rule - Springer Nature and Elsevier both restrict generative images in primary research. For a commercial report the equivalent line is that no chart, diagram or data visual may be generated rather than computed from the underlying numbers.

• Article 9(b) - disclose the role. Researchers must disclose whether AI, synthetic data or other emerging techniques and technologies "played a significant role in sampling, deployment, analysis, or interpretation of the data, and to what extent human oversight was involved". Read the list of four stages. Most teams would only think to mention analysis.
• Article 7(e) - tell the client. The client must be told when AI or synthetic data and personas are used, and the extent of human oversight must be stated. Oversight is not implied by your professionalism; it is a disclosed quantity.
• Article 4(a)(ii) - tell the participant, at the start. Use of a synthetic persona for data collection must be clearly notified to the data subject at the beginning of the research. Not in a debrief afterwards.
• Article 7(f) - keep it contained. Research data used with AI "must remain confidential... Access... strictly limited to a secure and controlled environment". This is the article that decides whether a consumer tool tier is usable at all.
• Article 6(c) - re-identification by inference. Guard against it. This is the one AI changes most, because a model that reads across your open-text fields can reconstruct an identity that no single field revealed.

Fundamental Principle 5 is the sentence that closes every argument about tooling. Researchers hold overall responsibility and oversight "irrespective of the method, technique and technology applied". There is no configuration of tools that moves accountability off your desk, and no vendor claim that does it either. That is the fourth rail, and it is why the other three are yours to enforce.

Then the Insights Association Code (September 2025) adds the provenance duty in one clean line: AI-generated data "must be clearly distinguished from data directly derived from human research participants". Combine it with a3's Bisbee finding and you get a concrete operating rule: synthetic material must be labelled at the row level, not mentioned in a footnote, because a reader who cannot tell which rows were human cannot judge the driver analysis at all.