AI for researchers: keeping unpublished data unpublished

AI is genuinely useful in research — summarising a literature corpus, drafting a methods section, suggesting a first-pass coding frame, turning a messy result into a readable paragraph. The catch is that a lot of research data is confidential by obligation: participant interviews under consent, unpublished results under embargo, data held under a sharing agreement. Pasting any of it into a consumer chatbot is a data-governance decision — and often one your ethics approval and data-management plan never sanctioned.

You don't have to avoid AI to stay clean. You have to be able to answer three questions about a tool before confidential data goes near it.

1. Where does the data go, and who operates it?

Funders and ethics committees increasingly ask exactly this in the data-management plan: where will the data be processed and stored, and by whom. "A cloud AI tool" is not an answer that survives review. Know the operator and the country — and remember that a "UK region" of a US-owned platform names a location, not an owner.

2. Who could be compelled to disclose it?

For data held under participant consent or a third-party agreement, this is the question that matters most. Under the US CLOUD Act, a US-owned provider can be required to disclose data in its possession regardless of where it's stored — so a US AI tool sits within reach of a foreign legal order even on a UK setting. For most projects that's a low practical risk, but your data-management plan should state the position honestly rather than imply a guarantee it can't make. Our free CLOUD Act exposure checker works it out for your current tools in two minutes, with no email and no tracking.

3. Does it train on your inputs?

If unpublished data or transcripts are used to improve a model, that is a disclosure and a use you almost certainly didn't obtain consent for — and it can compromise both confidentiality and, for some journals, prior-publication rules. Consumer tools often train by default with an opt-out; confirm the behaviour for the exact product and tier, and record it.

Integrity matters as much as privacy

One more honest point specific to research: AI outputs can contain fabricated citations and confident errors. The tool can draft and summarise, but you remain the author and remain responsible for accuracy and for checking every reference against the primary source. Used that way — as a drafting assistant under your scholarship, on infrastructure you can name in your data-management plan — AI is an asset rather than a governance gap.

Hush AI (hush-ai.uk) is a UK-sovereign AI assistant for regulated and confidential work. It drafts and summarises under your review — outputs may contain errors and fabricated citations, and you remain responsible for accuracy and authorship. This article is general information, not research-governance advice; follow your institution's ethics and data-management requirements.