AI for scientists and researchers: what works, what doesn’t

TL;DR: AI can help a scientist with expression, structure, and completeness: drafting section openers for a grant, turning one well-understood paper into summaries at three depths, translating a mechanism for a general audience. It cannot help with interpretation, mechanism, or scientific judgment – what it produces is plausible, and plausible is not correct. I’m Julie Kaiser, a working science writer, and this is the map I use to decide which side of that line a task falls on.

Who this is for

Scientists and researchers whose work has to leave their own head: grants, papers, explainers, conference materials, anything where the science is sound and the writing is the bottleneck. If academic writing is a real part of your job, this post is for you. If your question is whether AI can do the science, the answer is shorter: no (at least not yet).

Is AI accurate enough for scientific work?

For some of the work, yes. Here’s the rule my practice is built on: AI can help with expression, structure, and completeness. It can help explain things to me that I don’t understand. It cannot help with interpretation, mechanism, or scientific judgment.

The reason is mechanical, not moral. If you don’t already know whether the mechanism you’re describing is correct, no AI is going to correctly tell you. It will tell you something plausible, and plausible is not correct. In scientific writing, the gap between those two words is where careers get damaged.

The test I apply is simple: if the AI gets this wrong, would I catch it? If the answer is yes, that’s expression work, and I proceed. If I have to trust the output of the AI, it’s judgment work, and I stop. Being able to answer that question sorts nearly all of my working scenarios.

How I use AI as a working science writer

I’ve been a science writer since 2011. My clients are life science and pharmaceutical companies. My job is turning what they know into what everyone else needs to know – technical explainers, application notes, marketing materials, conference materials.

AI has changed how I do that job, but it has not changed what “good” looks like. The sections below are the tasks where it earns its place in my workflow, and the ones where I keep it out.

Can AI do a literature review?

Not the part that matters, and it fails in a specific, dangerous way. AI will happily summarize papers it hasn’t read and cannot even access, using titles and abstracts to generate plausible-sounding syntheses. Sometimes those are correct. Sometimes they invent findings that don’t appear in the paper. For a research paper you will defend, the verification cost of AI-drafted synthesis is much higher than doing the review by hand.

Inside the review itself, the safe job is expression: turning one paper you have actually read into three summaries at different depths – two sentences for a grant intro, a paragraph for a review section, a page for a background chapter. That’s expression work, not interpretation. Fine.

Two places it has earned its keep for me, both a step short of the review itself. The first was a literature survey – I was mapping what was out there on a topic, and alongside my own database searches I asked AI tools to hunt for specific papers and publications. Databases want exact queries, and when a topic sits tangential to everything around it, the right paper can hide from every search string you try. The AI turned up publications my own searches had missed. The second was screening: a big pile of papers in hand, and the question of which ones actually dealt with certain terms. AI tools helped me screen them – the reading stayed mine.

Explaining a mechanism to a non‑specialist audience

This is where I find AI very helpful. If I’ve explained something correctly and thoroughly to the AI, it can produce a general-audience version faster and better than I can from a cold start.

My job changes shape at that point. I review whether the analogy holds, whether the simplification lost anything load-bearing, whether the register fits the intended reader. The generation was the slow part. Now it’s the review that’s the work – and the review needs exactly the expertise the AI doesn’t have, which is why this division of labor holds.

Should scientists use ChatGPT for grant writing?

For a couple of specific jobs, yes. If you’re stuck on which framing will land with your reader, AI can generate multiple versions of a section opener – having real options to choose from beats staring at the one you have. And before submission, you can use it to list every objection a reviewer might raise against a specific claim, so you anticipate them in your draft instead of in a rejection letter.

What AI can’t do is know whether the science underneath the claim is sound. That’s still a working scientist reading a working draft, and no grant should skip that step because a fluent paragraph looked finished.

Where AI should not touch scientific writing

There are three places where AI should not touch the work at all:

1. Any claim about mechanism or interpretation that hasn’t been independently confirmed by a working scientist.
2. Any citation you haven’t verified against the source paper’s actual claims.
3. Any conclusion the reader will be asked to trust based on a data interpretation.

These aren’t ethical rules. They’re rules for not producing science writing that gets you flagged, called out, or unpublished. So remember: the stakes in this field aren’t engagement metrics; they’re your name on a permanent record.

The meta use: fact-check your own draft

One of the most underrated uses of AI in scientific writing points backward at your own text. Before anything ships, ask it to list every claim in the draft a hostile reviewer would want verified. Then verify those by hand, against sources the AI never touched. The AI finds the claims; you’re the one who decides whether they’re true.

The verification reflex behind that habit has its own post: AI hallucinations: why your AI lies to you and how I catch it.

The one rule for scientists

Stay the expert in the loop. AI can help you write about the science faster and clearer than you would alone, but it will never be the scientist. The moment your draft contains something you cannot defend without the AI, you have crossed the line. And plausible is not correct.

And if you’d rather build that disciplined setup with someone on a call with you, on your own machine, that’s what I do with people.