AI in scientific writing — how much is happening, and what the journals decided

What this review covers

Scientific publishing ran the experiment the rest of the writing world is still arguing about. Within months of ChatGPT's release, the major journals and ethics bodies had to answer two questions at scale: how much AI-assisted writing is actually entering the literature? and what should the rules be?

The answers are now on the record, and they are instructive far beyond academia. On measurement, two large corpus studies — one across nearly a million arXiv, bioRxiv, and Nature-portfolio papers, one across more than fifteen million PubMed abstracts — put numbers on how fast LLM-assisted prose spread. On policy, Nature, the Committee on Publication Ethics (COPE), and the International Committee of Medical Journal Editors (ICMJE) converged, independently and quickly, on the same two-part rule: an AI tool can never be an author, and any use of one must be disclosed.

Notice what the rule is not. It is not a ban, and it is not detection. The institutions with the most at stake in textual integrity chose a disclosure-and-accountability regime over a catch-them regime — a decision this review examines closely, because the reasoning is spelled out in the primary documents. Every quotation below is verbatim from the linked source, checked on the asOf date.

The scale of the question

By late 2023 the question was no longer hypothetical. Nature's own survey of the field opened with the finding, as the news feature put it:

"Artificial-intelligence (AI) tools are becoming increasingly common in science, and many scientists anticipate that they will soon be central to the practice of research, suggests a Nature survey of more than 1,600 researchers around the world."

That survey (Van Noorden & Perkel, Nature 621, 672–675) captured a community that was already using the tools and already worried about them — "concerned, as well as excited," in the feature's framing. What it could not do is measure actual prevalence in published text. Two corpus studies did.

Measurement one: Liang et al. map the increase

The Stanford-led study "Mapping the Increase of Use of LLMs in Scientific Papers" (arXiv:2404.01268) applied a population-level statistical estimator to a very large corpus. The scope, from the abstract:

"950,965 papers published between January 2020 and February 2024 on the arXiv, bioRxiv, and Nature portfolio journals"

The method matters as much as the number. Liang et al. do not classify individual papers — they estimate, at the population level, what fraction of the text shows LLM modification. That design choice sidesteps the false-positive problem that wrecks per-document AI detectors: an estimator that is honest about being unable to name any specific paper can still be accurate about the aggregate.

The headline findings:

• Computer science showed the largest and fastest growth in LLM-modified content — "up to 17.5%" of papers' content by February 2024.
• Mathematics papers and the Nature portfolio "showed the least LLM modification" — up to 6.3%.

Two things follow. First, the spread is real, large, and fast: from effectively zero before ChatGPT to a sixth of computer-science abstracts in about fourteen months. Second, it is uneven — the fields closest to the technology adopted it fastest, and the most tightly edited venues (the Nature portfolio) show the least.

Measurement two: Kobak et al. and the excess-vocabulary method

The second study attacks the same question with a completely different instrument, which is what makes the agreement between them persuasive. Kobak et al. (arXiv:2406.07016) analyzed:

"over 15 million biomedical abstracts from 2010–2024 indexed by PubMed"

Their method borrows from epidemiology: just as excess-mortality analysis detects an event by comparing observed deaths to the pre-event trend, excess vocabulary detects LLM involvement by an "abrupt increase in the frequency of certain style words" after ChatGPT's release — words the models reach for far more often than human biomedical authors historically did. No detector, no classifier, no per-paper accusation; just corpus statistics against a fourteen-year baseline.

The result:

"at least 13.5% of 2024 abstracts were processed with LLMs"

And the distribution is again uneven — the estimate "differed across disciplines, countries, and journals, reaching 40% for some subcorpora."

Read the two studies together and the triangulation is strong: two independent teams, two unrelated methods, two different corpora, both landing on roughly one in seven to one in six recent papers showing LLM involvement in the prose, with hotspots far above that.

What the measurements do not say

Both studies measure LLM involvement in text, not misconduct. An author who drafted every idea, ran the prose through a model for grammar and flow, and disclosed it in the acknowledgements contributes to these percentages exactly as much as an author who pasted a prompt. The corpus numbers cannot tell those cases apart — and the policy bodies, as we are about to see, decided the difference between them is disclosure, not detection.

This is the same lesson the detection literature teaches from the other direction: statistical methods are informative about populations and unreliable about individuals. The benchmark literature and the false-positive record show what happens when population-grade statistics are used to accuse individual writers.

The policy layer: Nature's ground rules

Nature moved first among the major journals. Its editorial of 24 January 2023 — "Tools such as ChatGPT threaten transparent science; here are our ground rules for their use" — set out two principles adopted across all Springer Nature journals:

"First, no LLM tool will be accepted as a credited author on a research paper. That is because any attribution of authorship carries with it accountability for the work, and AI tools cannot take such responsibility."

"Second, researchers using LLM tools should document this use in the methods or acknowledgements sections. If a paper does not include these sections, the introduction or another appropriate section can be used to document the use of the LLM."

Note the architecture of the rule: accountability attaches to humans; transparency attaches to the tool. Nothing in it bans the tool.

The same editorial is strikingly candid about why the rule is disclosure rather than detection. Asking whether editors can detect LLM text, it answers:

"Can editors and publishers detect text generated by LLMs? Right now, the answer is 'perhaps'. ChatGPT's raw output is detectable on careful inspection, particularly when more than a few paragraphs are involved and the subject relates to scientific work."

"Perhaps," for raw output, on careful inspection, at length — written in January 2023, before the paraphrase-attack literature showed how fragile even that is. (The subsequent evidence is collected in paraphrase attacks on AI detectors.) A publishing system that could not rely on detection built its integrity policy on disclosure and human accountability instead.

The policy layer: COPE

The Committee on Publication Ethics — the standards body whose guidance most journals follow in misconduct cases — issued a formal position statement, "Authorship and AI tools." Its core, verbatim:

"COPE joins organisations, such as WAME and the JAMA Network among others, to state that AI tools cannot be listed as an author of a paper. AI tools cannot meet the requirements for authorship as they cannot take responsibility for the submitted work. As non-legal entities, they cannot assert the presence or absence of conflicts of interest nor manage copyright and license agreements."

And the disclosure half:

"Authors who use AI tools in the writing of a manuscript, production of images or graphical elements of the paper, or in the collection and analysis of data, must be transparent in disclosing in the Materials and Methods (or similar section) of the paper how the AI tool was used and which tool was used."

The reasoning is legal-philosophical rather than technical: authorship is a bundle of responsibilities — accountability, conflict-of-interest assertion, copyright — that only a legal person can carry. The tool is not banned; it is placed permanently on the instrument side of the instrument/author line, with its use reportable like any other method.

The policy layer: ICMJE

The International Committee of Medical Journal Editors — whose "Uniform Requirements" govern submissions to thousands of medical journals — wrote the most operational version of the same rule. From the AI-assisted-technology section of its Recommendations:

"At submission, the journal should require authors to disclose whether they used artificial intelligence (AI)-assisted technologies (such as Large Language Models [LLMs], chatbots, or image creators) in the production of submitted work."

On authorship:

"Chatbots (such as ChatGPT) should not be listed as authors because they cannot be responsible for the accuracy, integrity, and originality of the work, and these responsibilities are required for authorship."

And on where the residual responsibility lands:

"Authors should carefully review and edit the result because AI can generate authoritative-sounding output that can be incorrect, incomplete, or biased."

"Authors should be able to assert that there is no plagiarism in their paper, including in text and images produced by the AI."

The ICMJE text adds the operational detail the others imply: disclosure in the cover letter and the manuscript, writing assistance in the acknowledgements, data/figure use in the methods. The human signs for everything, including the machine's mistakes.

Summary table

Source	Type	Corpus / scope	Headline finding or rule
Liang et al. 2024 (arXiv:2404.01268)	Corpus measurement	950,965 papers, Jan 2020–Feb 2024	LLM-modified content "up to 17.5%" in CS; least (up to 6.3%) in Mathematics and Nature portfolio
Kobak et al. 2024 (arXiv:2406.07016)	Corpus measurement	15M+ PubMed abstracts, 2010–2024	"at least 13.5% of 2024 abstracts were processed with LLMs"; up to 40% in some subcorpora
Nature editorial (Jan 2023)	Journal policy	All Springer Nature journals	No LLM as credited author; document use in methods/acknowledgements
COPE position statement	Ethics-body policy	COPE member journals	AI tools "cannot be listed as an author"; transparent disclosure of how and which tool
ICMJE Recommendations	Ethics-body policy	Medical journals following ICMJE	Disclosure required at submission; chatbots "should not be listed as authors"; humans responsible for AI output
Van Noorden & Perkel (Nature 621)	Survey	1,600+ researchers worldwide	Community "concerned, as well as excited"; AI expected to become central to research practice

What the literature does and does not establish

Established:

• LLM-assisted prose entered the scientific literature at large scale within roughly a year of ChatGPT's release — two independent methods agree on double-digit percentages by 2024, with strong field and venue variation.
• Population-level measurement of AI involvement is tractable and credible precisely because it declines to accuse individual documents.
• The major publishing-ethics bodies (Nature/Springer Nature, COPE, WAME, JAMA Network, ICMJE) converged on the same rule: AI cannot hold authorship, because authorship is accountability; AI use must be disclosed, because methods must be transparent.
• The institutions that wrote these rules did not build them on detection. Nature's own editorial rated detection of raw output at "perhaps" — and the rules bind regardless of whether any detector fires.

Not established:

• That any specific percentage of misconduct exists in the literature. Both corpus studies measure text modification, which includes fully disclosed, policy-compliant editing assistance.
• That disclosure rates match usage rates. The gap between Kobak's 13.5% measured involvement and the much smaller fraction of papers carrying AI acknowledgements is real but unquantified in these sources.
• That per-document detection works in scientific prose. No policy body endorses it as evidence; the independent accuracy record is reviewed in how accurate are AI text detectors?.

Why this matters for provenance

Scientific publishing is the canary for every other writing institution, because it faced the problem first, at scale, with the strongest incentives to get it right — and it chose disclosure plus human accountability over detection. The rule that survived contact with reality is procedural: say what you used, sign for the result.

But a disclosure regime has an obvious open edge: it runs on trust, and the corpus studies suggest usage runs well ahead of acknowledgement. What a disclosure regime lacks is evidence — a way for the honest discloser to show their account of the process is true, and for an editor to check it without becoming a forensic statistician. That is precisely the gap writing provenance fills: a sealed, tamper-evident record of how a manuscript was actually produced, created with the author's consent while the writing happens, and verifiable by anyone afterwards. A journal that asks for disclosure can accept a process record as its substantiation; an author accused of undisclosed AI use can answer with evidence rather than assertion — the same shift from guessing to records argued across this silo, from detector accuracy to watermarking to keystroke-level process research.

Related reading: student AI-use surveys — what the numbers say · vendor accuracy claims vs independent measurements · what is writing provenance?