Research ethics and integrity · Lesson 3 of 3

Using AI in research, responsibly

Use AI for finding, drafting and structuring. Verify everything factual, never let it be your only analyst, and disclose what you used.

9 min read · Free — no account needed to read

AI tools are now part of ordinary research practice, including in this product. The useful frame is not whether to use them but what each use does to the integrity of the result.

Search and discovery is low risk and high value — finding papers, suggesting terms, surfacing adjacent literature. The caution is that you still do not know what the tool did not show you, so it supplements a documented search rather than replacing one.

Summarising is useful and lossy. A summary of a paper is not the paper, and a summary of a summary is considerably further away. Never cite a source you have only read a summary of — the summariser cannot tell you what the methods section says about the sample.

Hallucination is the specific risk that matters most. Language models produce plausible citations that do not exist, attribute real findings to the wrong paper, and state numbers with complete confidence. Every citation, quote and figure must be verified against the source. This is not optional and it is not paranoia; it is the single most common way AI-assisted work fails publicly.

Analysis and coding. AI can propose codes for qualitative data or draft analysis scripts. It must not be the only coder, and it must not be treated as agreeing with you independently — it has no independent view, so "the AI agreed with my interpretation" is not corroboration of anything.

Writing. Drafting, restructuring and improving clarity are widely accepted. Generating substantive content presented as your own reasoning is not. The line most fields draw is that the ideas, the analysis and the responsibility must be yours.

Disclosure. Say what you used and for what. Many journals, funders and institutions now require it, conventions are moving quickly, and the safe position is specific and matter-of-fact: which tool, for which task.

Confidentiality. Participant data, unpublished work and anything commercially sensitive should not be pasted into a tool without knowing where it goes and whether it is retained. This is a data protection obligation, not a preference — and it is one participants consented to on your account.

Everything in this product is built on the same principle: the library is the truth, AI reads it and does not replace it, and no AI claim appears without the sources behind it.

See it happen

You ask an AI assistant for sources on a topic and it returns five citations with titles, authors and years.

The read that arrives on its own

Five relevant, specific, correctly formatted citations — add them to the reference list.

The same situation, with this idea

You look up each one. Three exist. One has the right title and the wrong authors. One does not exist at all — the title is a plausible combination of words in the field. The three real ones are genuinely useful, so the tool did real work, and the two fabricated ones would have been in your bibliography if you had not checked. You check every time, because there is no signal in the output distinguishing the real from the invented.

After this, you can

verify every AI-supplied citation against the actual source, every time.

The takeaway

Use AI for finding, drafting and structuring. Verify everything factual, never let it be your only analyst, and disclose what you used.

Easy to take the wrong way

That careful researchers avoid these tools. Used with verification they are straightforwardly useful; used without it they manufacture confident errors faster than anything else available.

Try this

Take any AI-assisted output in your own work and verify every factual claim in it against a source. Note how long it took, and what you found.