35% of scientists use AI to write papers
But most still do not admit it
In a recent Nature survey of over 5,000 researchers, a quiet contradiction emerged. The majority of scientists — 65% — believe it’s ethically acceptable to use generative AI tools like ChatGPT to help write research papers. And yet, only 8% say they’ve actually done it. Of those, many admit they never disclosed it.
This gap between approval and admission isn’t just about technology. It’s about discomfort. We’re watching a fundamental norm — how scientific writing works, and who gets to do it — start to unravel in real time.
The survey, published on May 15, asked researchers around the world what kinds of AI use they considered appropriate. Editing and translating papers with AI received overwhelming support: more than 90% of respondents were fine with it. Even writing a first draft with AI was seen as acceptable by a clear majority.
But the pattern changed when behavior was measured instead of opinion. Only a small fraction of scientists have used AI to write papers, and even fewer have disclosed doing so. The dissonance is telling.
Disclosure itself is a battleground. Some journals require it, some don’t. Some ask for a basic note; others demand details about the prompts used. Many researchers don’t know where their institution stands. Others simply assume — maybe hope — that it’s nobody’s business.
There are practical reasons for the silence. Academic publishing is built on trust — or at least the appearance of it. To admit to using AI is to risk being seen as lazy, unoriginal, or worse: fraudulent. Even if the tool was used ethically, even if the content is accurate and the conclusions are sound, the stigma lingers.
This isn’t new. Every new writing technology — from typewriters to spell check to Grammarly — has carried its own period of suspicion. But generative AI feels categorically different. It doesn’t just correct grammar; it generates ideas, structures arguments, imitates style. It moves from support to substance. And that shift unsettles people.
One researcher put it bluntly: “Using AI feels like taking a shortcut through a forest we’re meant to walk.”
Usage patterns reveal another tension. Younger researchers, and those for whom English is not a first language, are significantly more likely to use AI tools for writing and editing. For them, these tools are not an unfair advantage — they’re a necessary equalizer.
In this light, AI doesn’t just disrupt existing norms — it exposes their inequalities. For decades, global academia has privileged a particular kind of polished, Anglophone prose. Now, AI offers a workaround. But it also raises the stakes. If everyone suddenly has access to the same level of fluency, what distinguishes original voice from engineered clarity?
And does that even matter?
That question leads to the deeper issue: authorship itself. What does it mean to be the “author” of a paper if large sections were drafted, translated, or refined by a machine? If an AI helped organize your argument or smooth out your sentences, is your voice still yours?
Science, as a culture, has always prized transparency. Methods must be replicable. Data must be cited. But when it comes to writing — the medium through which scientific ideas are shared, shaped, and immortalized — the norms are murkier.
Should prompt disclosure be standard? Should AI tools be footnoted like software packages or thanked like research assistants? Some journals say yes. Others say nothing. Many researchers are left to guess.
It’s tempting to reduce all this to policy — a matter of disclosure forms and editorial guidelines. But something deeper is in play. Scientific writing is not just a medium of communication. It’s a ritual of cognition. It’s how researchers clarify their own thinking. To write is, in some sense, to know.
If that process is delegated — even partially — to a system trained on the statistical exhaust of other people’s writing, does it change the nature of the knowing?
Some would argue no. Others, like one survey respondent from the UK, aren’t so sure:
“By using it, we rob ourselves of the opportunity to learn through the labor.”
This isn’t nostalgia for suffering. It’s an acknowledgement that the friction of writing — the hesitations, the false starts, the careful stitching of claim and evidence — is part of how science refines itself. Take that away, and what’s left might be cleaner, but not necessarily truer.
Despite all the ethical noise, the technological shift is already underway. Quietly. Incrementally. Sometimes invisibly.
Researchers aren’t staging protests or rewriting manifestos. They’re opening new tabs. Asking GPT to rewrite an awkward sentence. Feeding it a messy paragraph and asking for clarity. It starts small. Then it becomes normal.
In a few years, it may seem quaint to worry about this. Or we may realize we’ve crossed a line without noticing.
What’s clear now is that the AI revolution in scientific writing isn’t being led by policies or proclamations. It’s being shaped by private choices, by quiet habits, by researchers working late, looking for help, and hoping not to get caught.