Ninety percent of biomedical papers published in December and archived in PubMed show signs of AI assistance, according to a study published today in Nature. That's higher than previous estimates of large language model use, and it underscores how quickly AI has moved into the core of scientific writing.
What the study found
The study, which appeared in Nature on August 20, looked at biomedical papers from December that are stored in PubMed, the U.S. National Library of Medicine's database. Researchers found that 90% of them carried some mark of AI help — whether that's a particular turn of phrase, a structural pattern, or another stylistic fingerprint. The exact year for December isn't specified, leaving a question mark over whether this reflects a sudden recent surge or a slower, longer build-up.
đź“Š Market Data Snapshot
The number is notable because earlier estimates of large language model use in scientific literature were lower. The study doesn't name those earlier figures, but the contrast is the point: AI is now pervasive in biomedical publishing.
The trust problem
For scientists and readers, the finding raises a practical issue: distinguishing human work from AI-generated text is no longer straightforward. The study itself cautions that AI-detection tools are imperfect. Linguistic pattern matching can produce false positives, so the true rate could be lower — or higher — than 90%. But even with that caveat, the scale of AI involvement is a signal that the peer-review process and the broader scientific record are now dealing with a different kind of authorship.
That's not just an academic concern. If most papers are AI-assisted, then verifying the origin of data, methods, and conclusions becomes harder. The study doesn't offer a solution, but it highlights a gap that some technologists think blockchain could fill.
A crypto angle
For the crypto sector, the study is a fresh data point in the argument that AI needs a verification layer. If nine out of ten papers rely on AI, then proving provenance and human authorship becomes valuable. Blockchain-based tools for timestamping, decentralized identity, and content verification are often floated as answers to exactly this kind of problem. The study doesn't mention crypto, but it's the type of finding that projects focused on research integrity — sometimes grouped under the "DeSci" label — are likely to cite.
The immediate market impact is probably neutral, since the study isn't crypto-specific. But it reinforces a narrative that's been building: as AI spreads into critical fields, the demand for tamper-proof records and decentralized trust grows. That's a long-term story for investors, not a short-term catalyst.
What's still unknown
The biggest open question is whether the 90% figure is accurate. The study doesn't publish the sensitivity or specificity of its detection method, and the unspecified year for December makes it hard to judge how fast the trend is moving. If the number is right and recent, it's a strong argument for verification infrastructure. If it's inflated, the urgency is overstated. Either way, the study is a reminder that AI's footprint in science is bigger than previously estimated — and that the tools to keep that record trustworthy are still catching up.
