90% of PubMed papers are now written by LLMs, and we are just fine with it

A new arXiv preprint suggests that the landscape of scientific publishing has shifted dramatically, with nearly 90% of biomedical articles published in December 2025 showing clear markers of LLM assistance. Compared to the 52% recorded in 2024, the speed at which researchers have adopted these tools is nothing short of a digital gold rush.

Previous studies that tracked only abstracts likely missed the true scope of the automation, as those methods relied on specific vocabulary patterns rather than a deep statistical analysis of the entire document structure. By scanning the broader PMC archive, the authors found that these models are not just being used to spice up the language in the discussion sections—they have infiltrated the results section in 58% of the analyzed papers.

While the discussion section is essentially the place where scientists try to make sense of their findings, the results section is supposed to be the bedrock of empirical truth. The discovery that LLMs are potentially hallucinating data points in the most critical parts of clinical research is a spectacular testament to the trust we place in black-box software. Since this preprint has not yet faced peer review, the academic community remains in a state of polite, nervous silence regarding whether this constitutes helpful editing or mass-scale data fabrication.

The era of authentic human-written science is effectively becoming a luxury good, akin to handmade artisanal pottery in a world of 3D-printed plastic. Whether this leads to a new golden age of discovery or a complete collapse of institutional credibility depends entirely on how soon the industry mandates disclosure over denial.

Source: arXiv