Artificial intelligence is rapidly transforming not only journalism, education, and business, but also the very way modern science is written.
A new study published in August 2026 on arXiv has revealed a striking trend: by the end of 2025, signs of large language model usage were found in nearly nine out of ten biomedical papers.
But it is important to understand the main point here: this does not mean that 90% of scientific papers are written entirely by AI.
It means that artificial intelligence is increasingly helping authors formulate, edit, and rewrite individual parts of the text.
The scale turned out to be much higher than previous estimates
The researchers analyzed the full texts of biomedical papers from PubMed Central.
They estimate that the share of publications with signs of LLM involvement was about 52% in 2024 and grew to approximately 77% in 2025.
And by December 2025, the figure approached 89%.
This is much higher than previous estimates because, in the past, scientists more often analyzed only abstracts.
When the team looked at the full papers, it became clear: AI usage may be significantly more widespread than assumed.
Where AI is found most often
Researchers found the most noticeable traces of LLMs in those parts of an article where an author needs to explain, interpret, and formulate thoughts.
Signs of AI were found especially often in introductions and discussions.
Much less often — in the methods and results sections.
And this is quite logical.
It is precisely the introduction and discussion that require a large amount of coherent text, phrasing, comparisons with previous studies, and explanations of the significance of the obtained data.
In other words, AI is increasingly being used as a kind of scientific editor or style assistant.
But a new risk arises
The problem begins when a language model is used not only for editing text but also for describing actual results.
LLMs are capable of creating very convincing phrasing, even when the information is erroneous.
That is precisely why researchers are particularly concerned about the use of AI in the Results sections.
If a model starts to "hallucinate" data, this is no longer a question of style — it is a potential threat to the reliability of science.
There is another problem as well.
If thousands of scientists use the same models, scientific texts may gradually begin to resemble one another.
Identical speech patterns, identical ways of explaining things, and identical logical templates may spread through the literature significantly faster than before.
Scientists themselves admit to using AI
The scale of the trend is also confirmed by surveys.
In 2025, about 71% of researchers admitted to using artificial intelligence to assist in writing scientific texts.
And the real share may be even higher.
For many scientists, especially those for whom English is not their native language, LLMs have become a convenient tool: they help correct grammar, make text more understandable, and prepare manuscripts for publication faster.
Therefore, it is unlikely that the scientific community will be able to completely abandon AI.
“You can’t put the toothpaste back in the tube”
One researcher, commenting on the situation, described what is happening very accurately:
“The toothpaste is already out of the tube, and you can’t put it back.”
And this is, perhaps, the main thing.
The question is no longer whether scientists will use AI.
They are already using it en masse.
Now something else is more important: where is the line between help with writing text and interference with the content of scientific work.
Is it okay to ask a model to improve English?
Is it okay to rewrite a complex paragraph?
Is it okay to allow it to interpret results?
And to generate part of the discussion?
The scientific community has yet to formulate answers to these questions.
Science is entering a new era
There is one more important caveat: the study has so far been published as a preprint and has not undergone full scientific peer review.
Therefore, the figures may be refined.
But even if the final estimate turns out to be slightly lower, the general trend is already obvious.
AI is becoming a part of scientific writing—just as computers, search engines, and statistical software once became a part of the work.
And now the main question is no longer:
“Is science being written with the help of AI?”
But rather:
“Will scientific literature be able to maintain originality, transparency, and trust if almost every text passes through the same language models?”
