There’s a particular feeling a lot of researchers have right now, and it goes something like this: everyone else seems to be using AI, they all seem to be writing faster, funders keep mentioning it in their calls, and somewhere underneath it all is a nagging little voice asking – am I being left behind?

And then, almost in the same breath, a second voice pipes up: but research is supposed to be my original thinking. Attribution matters. There are policies. What exactly am I responsible for here?

That tension – the FOMO versus the integrity – is real, and I don’t think it helps anyone to pretend it isn’t. I’ve been delivering sessions on generative AI for researchers at GCU, and rather than handing people a list of “do this, don’t do that” rules (which date badly and never quite fit your actual situation anyway), I wanted to give people a way to think about the decision. So I developed a framework I call TASK. This post walks through it.

A quick note before I dive in: TASK helps you decide whether to use AI for something. My TASK framework works well with the AIR: AI in Research framework that helps you document what you did – the AIR framework has been designed by Jo Young at Electv Training, shared under a CC BY 4.0 licence.

Why bother with a framework at all?

Because the alternative is reaching for a tool on autopilot, and autopilot is exactly where things go wrong. The regulatory backdrop has caught up too. The EU AI Act‘s AI literacy obligation has been in force since February 2025, the UK’s cross-sector principles – safety, transparency, fairness, accountability, contestability – already apply, and Scotland’s AI Strategy names AI literacy in the research community as a national priority.

But honestly, the regulation is the least interesting reason. The best reason is that a few seconds of thinking before you paste something into a chatbot will save you from the kind of mistake that’s genuinely hard to walk back.

The TASK framework

Before you reach for an AI tool, ask four questions. They spell out TASK: Task type, Accountability, Sensitivity, Knowledge.

T – Task type: what kind of task is this?

Not all tasks are equal. I find it helps to picture a spectrum. At one end sits generative work – drafting a blog post, summarising your own notes. This is often perfectly fine, with your oversight and edits. Next comes analytical work – coding qualitative data, spotting themes in the literature. Appropriate with caution: AI can assist, but you interpret. Then evaluative work – peer reviewing, judging the quality of evidence. Tread carefully, because this is core researcher judgement. And at the far end is creative or original work – generating your actual research argument or theory. This is rarely appropriate, because this is the thing. It’s your intellectual contribution.

A – Accountability: who is responsible for this output?

The answer is always you. Always. AI tools have no academic accountability – you can’t cite one as a responsible researcher, and it can’t be held to account when something is wrong. The question I’d really sit with here is this: if this AI output turns out to be wrong, do I know enough to catch it and correct it? If the honest answer is no, that’s a flag. Verify against your own knowledge and primary sources, disclose your use in line with policy and journal guidelines, and never pass off AI-generated content as your own without genuinely reviewing and reworking it.

S – Sensitivity: what data or knowledge is involved?

This is the one with the sharpest edges. Published papers, public documents, your own anonymised low-sensitivity notes, generic writing with nothing confidential in it – broadly safe to pop into public tools. Unpublished manuscripts, commercially sensitive research, institutional data – proceed with caution, check the relevant policies and contracts, and use your institution’s properly credentialed tool (for us, that’s M365 Copilot with a GCU login). And then there’s the absolute no: personal data about research participants, anything under ethics or funder confidentiality, unpublished IP with commercial value. These do not go into public AI tools. A PhD student uploading interview transcripts to a public tool is, very possibly, a GDPR breach – not a productivity hack.

K – Knowledge: does AI doing this diminish your expertise?

This is the one people skip, and it might be the most important. Expertise is built by doing hard things. If AI does all the hard things, what exactly are you developing? Three questions help: Is this a skill you’re still building? If you’re newer to research and learning to critically appraise literature, doing your own first pass really matters. Is your expert judgement the whole point? If the value of your analysis comes from your situated, disciplinary knowledge, AI can’t replicate that anyway. Or is this a well-established, repetitive task – formatting references, transcribing audio, a first-draft summary of your own content? Then AI assistance is reasonable, and frankly a relief.

Let’s actually use it: two scenarios

Frameworks are easy to nod along to and harder to apply, so here are three quick scenarios I use in the sessions. Try running each through TASK yourself before you read my verdict.

Scenario 1: qualitative coding

A PhD student uploads interview transcripts to a public tool, uses it to identify themes, and presents those themes as their own thematic analysis. Task type? Analytical – and interpretation is the core intellectual contribution. Accountability? They’re claiming it as their own work without disclosure, which is misleading. Sensitivity? Interview data is very likely GDPR personal data and should never go near that tool. Knowledge? Thematic analysis is a core doctoral skill, and outsourcing it hollows out their development. Verdict: red, on multiple counts. This one needs a complete rethink.

Scenario 2: a research blog post

A researcher asks an AI tool to write a blog post about their published paper for a general audience, then posts it on their institutional blog without review or edits. Task type? Generative – adapting published work for the public. Accountability? No review means that if the tool misrepresents the research, the reputational risk lands squarely on them. Sensitivity? Published content only, though it’s worth checking the tool’s data terms. Knowledge? Public engagement writing is a genuine skill, and your voice is the point. Verdict: amber. AI can help draft, but it must be reviewed, edited, and made to sound like an actual human – you. (Yes, I’m aware of the irony of writing that in a post that AI helped me draft. More on that in a second.)

The point isn’t a single right answer

When I run these scenarios live, the most useful moment is always the disagreement. Reasonable researchers reach different conclusions, and there are real disciplinary differences in how people weigh each dimension. That’s not a bug in the framework – it’s the whole point. TASK isn’t a vending machine that dispenses verdicts. It’s a way of making sure you’ve actually looked at the decision from four angles before you commit to it.

And once you’ve decided: document it

TASK answers “should I use AI here?” The follow-up question is “how do I describe and disclose what I did?” That’s where AIR comes in – a framework for describing AI use transparently, stage by stage, with bands running from A0 (no AI) through to A4 (substantial, output-shaping use). A higher band isn’t worse; it just means more responsibility for transparency and checking. A simple disclosure might read: “AI was used to draft the methodology section from researcher notes. The final text was reviewed, edited and approved by the author. AI was not used in data collection or analysis.” Used together, TASK before and AIR after, you’ve covered the whole decision-and-accountability loop. The AIR framework is AI in Research by Electv Training, used here under its CC BY 4.0 licence.

So, should AI do this?

Sometimes yes, sometimes absolutely not, and a lot of the time it depends – which is exactly why a quick mental checklist beats a rigid rulebook. Before you reach for the tool, run through TASK. Is the task type appropriate? Will you review and own the output? Is the data safe to put in? And does this build your expertise, or quietly bypass it?

Full disclosure, in the spirit of the thing: I drafted this post from my own session slides with AI help, then rewrote it in my own voice and checked the lot. Call it Outreach, A3.

As ever, I’d love to hear how you’re navigating this. What’s your trickiest TASK dimension to apply? Leave a comment.