Did AI speed up science?What ten years of publishing data say about AI and research

Start with the peer-reviewed record. The US National Science Foundation publishes a world series built on Scopus, a curated index of journals.²

If ChatGPT had broken the trend, growth would have climbed after it launched in November 2022. It fell. Output compounded at 5.6% a year from 2015 to 2021 and 4.1% a year from 2021 to 2024. Start the clock at 2022 and it is 4.3%.

Two fair objections. Journals lag: a paper drafted in 2023 may appear in 2025, and the 2025 counts are still filling in. And the answer shifts with the database. OpenAlex, a broader open index, shows growth rising from 6.8% a year (2016–2022) to 8.3% (2022–2025), but users warn that part of the recent jump is expanded indexing.³ A 2026 field-level study of Scopus concluded that total output “did not display a generalized discontinuity after 2022.”⁴

So the honest reading is that no AI break has shown up in the journals yet. The next section explains what is really moving that line.

Between 2021 and 2024 the world added 399,195 articles. China and India together added 402,310. The rest of the planet, taken together, went slightly backwards.

This shift began long before anyone had an LLM, or large language model, the technology behind ChatGPT. China passed the US in 2017. It crossed one million articles in 2024, about 31% of the world total. India doubled its output between 2019 and 2024. NSF credits the two with “most of the growth in annual publications since 2014.” WIPO’s separate count tells the same story.⁵

Volume has not closed the impact gap. In NSF’s data, 1.7% of 2022 US papers landed in the world’s top 1% most cited, against 1.3% for China and 1.1% for India.

Anyone who credits global paper growth to AI is mostly describing national research systems.

Now move closer to where writing happens. A preprint is a paper posted before peer review, so it has no journal queue to hide the change.

Computer science went from 34.8% of arXiv submissions in 2022 to 47.6% in 2026.⁶ AI conferences moved even faster. ICML went from 12,107 submissions in 2025 to 23,918 in 2026.⁷ ICLR went from about 7,250 in 2024 to about 19,000 in 2026.⁸ Part of that is the AI boom itself: more people are working on AI, and they need somewhere to publish.

Journals feel it at the front door. Organization Science, a leading management journal, saw submissions rise 42% after ChatGPT, and a study of the journal attributes nearly all of the rise to AI-written papers.⁹

Biology is calmer. bioRxiv grew about 13% in 2025 and medRxiv about 16%, steady growth with no spike.¹⁰

You would expect computer science to lead, and for 2024 it did. Up to 22% of its abstracts showed LLM editing, against about 9% in mathematics.¹¹

Then, in late 2025, agentic tools arrived, meaning AI that can carry out a multi-step task on its own. A new preprint by Carlo Contaldi finds a clear break in arXiv from December 2025.¹² Outside computer science, submissions ran 17% above trend. Mathematics shows the largest break: 26.5% above trend, and by his abstract’s count submissions “doubled within a year.”

That detail about the first draft is a neat summary of the moment. It also helps explain the pattern. Contaldi notes that the gradient does not line up with surveyed rates of AI adoption. It lines up with what the work requires. A theorem needs a person and a keyboard. A particle detector needs a decade.

In mathematics, about two-thirds of the rise comes from more distinct authors, and solo papers rose about 70%. New people are publishing, which is a different thing from a few prolific authors publishing more.

The most quoted number comes from Kusumegi and colleagues in Science in December 2025.¹⁴ Across 2.1 million preprints, researchers who adopted LLMs published 36.2% more on arXiv, 52.9% more on bioRxiv and 59.8% more on SSRN.

Then came the challenge. Renault, Bergeaud and Bosquet pointed out a flaw in dating adoption from a researcher’s first flagged abstract.¹⁵ You are more likely to be flagged in a month you submit a lot, so the “adoption date” tends to land on a busy month. Random flags reproduce the same apparent gain.

The original authors replied in August 2026 with tighter designs.¹⁶ A link survives. Its size shrinks.

Other studies sit across that range. One matched panel of social scientists finds a 67.4% rise in publications, in a preprint.¹⁷ A Nature study of 41.3 million papers finds that scientists using AI methods publish about three times as many papers, though that is a correlation and likely selection: stronger scientists adopt faster.¹⁸

If you are budgeting on AI doubling your research output, plan on a modest gain per person instead.

The largest gains in the Science paper went to authors for whom English is the hardest barrier.

Writing is converging too. Papers from non-English-speaking countries have moved closer to US scientific English since 2022.¹⁹ For researchers who were judged on their prose as much as their ideas, that is a real gain.

The gatekeepers have not moved with them. At Organization Science, non-native-English teams and newcomers are the most likely to use AI, yet they “do not benefit in the review process and, in fact, may be hurt.” Kusumegi and colleagues warn that as polished prose stops signalling quality, reviewers may lean on pedigree and institutional affiliation instead. AI detectors add to the problem, because they have been shown to be biased against non-native writers.²⁰

AI gave more people a voice. It may have made the room more likely to check their badge first.

For decades, complex writing was a rough proxy for strong work. In the Kusumegi data, that holds for human-written papers. For LLM-assisted papers it inverts: the more complex the writing, the less likely the paper is to be published. The cheap signal has stopped working.

This is the chart I would show anyone arguing for a ban. Papers with up to 30% AI text did as well as the rest. The penalty starts when AI does most of the writing. The tool is fine. The substitution is the problem.

The strain shows up in several places:

• Reviews. Pangram, which sells AI detectors, flagged 21% of ICLR 2026 reviews, or 15,899, as fully AI-generated.⁸ Treat the exact number with care: a rival method tested on a sample of about 600 papers put the rate at 5%, against 24% for Pangram’s model.²¹ Over 30% of Organization Science reviews use some AI.
• Citations. GPTZero, another detector vendor, found 100 invented citations across 51 of the 4,841 NeurIPS 2025 papers it scanned.²² NeurIPS replied that incorrect references do not necessarily invalidate a paper. Most papers were clean, and these ones passed review.
• Fraud. Suspected paper-mill output doubles every 1.5 years, against 3.3 years for retractions and 15 years for all publications. Only 28.7% of suspected mill papers found in OpenAlex have been retracted.²³ That trend predates LLMs. They lower its cost.

Put these together and you get a strange loop: AI writes more papers, tired volunteers lean on AI to review them, and the system that certifies knowledge runs on the labour that AI did not make cheaper. It is like buying a faster printing press and keeping the same few proofreaders.

The Nature study has the sharpest finding I saw. AI use expands the individual and contracts the field. Scientists using AI get about 4.8 times the citations and lead projects 1.4 years earlier. Yet adoption goes with a 4.6% narrower range of topics and about 22% less engagement between scientists, as work clusters in data-rich areas.¹⁸

That sits on top of an older worry. A 2023 Nature study of 45 million papers found science becoming less disruptive over decades, though its measure has been challenged.²⁴ So AI did not start this, and we do not yet know whether it will make it worse.

I found no evidence that the rate of real breakthroughs has risen with paper counts. Tools built for discovery, such as AlphaFold, may be a different story from tools built for writing. I could not find a solid causal estimate for them.

If you run a lab, a university or a research budget, four things follow.

• Retire paper counts as a measure. Organization Science found the AI-writing surge was stronger at schools that respond to journal-list incentives. AI amplifies whatever you reward. Assess a few chosen papers in depth. I made a similar argument about hiring in Is the coding interview dead? When AI can produce the artefact, the artefact stops proving the ability.
• Pay for verification. Reviewers are volunteers carrying a load that has doubled in places. Fund review time, and require open code and data so that claims can be checked.
• Set a rule for disclosure, and allow moderate use. The data does not support a ban. Language help is a real gain for many researchers. Punish fabrication and thin work, whatever tool produced them.
• Protect early-career and non-native researchers. Blanket caps fall hardest on small groups, and reviewers falling back on prestige hurts newcomers most. The same pressure on junior people runs through Are tech workers cooked?

If you build AI, as I do, read this as a market signal. I asked what stays valuable when machines do the producing in Did AI kill the modern tech stack? Research gives the same answer. Writing tools are abundant. The scarce product checks a claim against its data, its code and its sources, then shows its working.

AI sped up the paper. Whether it speeds up science depends on whether we rebuild the part that decides what is true.

If your institution dropped paper counts tomorrow, what would you measure instead? Tell me on LinkedIn or X.