Where AI Actually Saves Time in Research, and Where It Costs You

Most advice about using AI in research is either breathless or dismissive. Neither is much help when you have a dissertation to finish. Here is where it has actually saved me time, and where it has cost me time by looking as though it was saving me time.

Where it genuinely helps

Reading you were never going to do

There is a tier of papers adjacent to your work that you should probably read and realistically will not. Getting a summary of the method and the claimed result is not a substitute for reading the ones that matter, but it is far better than the alternative, which is ignoring them.

The rule I use: if the summary suggests the paper matters, read the paper properly. The summary is a filter, never a citation.

Getting unstuck on code

Research code is written once, by one person, under time pressure. Debugging it alone at eleven at night is slow. Describing the problem and getting three plausible causes is genuinely faster than staring, and roughly the same value as asking a colleague who is not available.

The first draft of anything administrative

Ethics applications, grant boilerplate, conference bios, reviewer responses. This writing has to exist and nobody’s career advances because of its quality. Drafting and then editing is meaningfully faster than starting from a blank page.

Rubber-ducking an argument

Explaining your reasoning and being asked where it is weak is useful even when the questions are obvious — sometimes especially then, because the obvious question is the one you stopped asking.

Where it costs time while appearing to save it

Anything requiring a citation to be correct

Fabricated references remain a real failure, and they are dangerous precisely because they look right: plausible authors, plausible venue, plausible year. Every reference must be verified against the actual source. If you were going to verify anyway, the time saved is close to zero.

Technical claims in your own specialism

Output in your field is fluent and sometimes subtly wrong, and subtle wrongness is expensive. You will catch it — you are the expert — but the catching costs more attention than writing the passage yourself.

The inversion worth noticing: AI is most reliable where you are least able to check it, and least reliable where you can.

Anything that becomes your voice

Accepting generated prose into a paper produces text that reads as competent and anonymous. Reviewers notice, supervisors notice, and it slowly erodes the thing that makes your writing recognisably yours.

The rule I have settled on

Use it for the parts of the work where being wrong is cheap and detectable. Do not use it for the parts where being wrong is expensive or invisible.

That single test resolves most cases. Debugging: cheap and detectable, because the code either runs or it does not. Citations: expensive and invisible, because a fake reference survives until a reviewer checks. Administrative drafting: cheap. Novel technical argument: expensive.

The disclosure question

Norms are still forming and vary by venue, so read your journal’s and your institution’s policy rather than guessing. The general shape is that using AI for language editing is broadly accepted with disclosure, using it to generate substantive content is not, and it can never be an author because it cannot take responsibility.

When uncertain, disclose. The cost of over-disclosing is a sentence in a methods note. The cost of under-disclosing, discovered later, is your reputation.