LLM use in biology preprints has nearly caught up with specialized AI tools

Epoch AI checked every bioRxiv biology preprint from January 2020 through September 2026 for acknowledgments of AI use, and found that mentions of large language models have almost caught up with mentions of specialized AI tools. By September 2026, 17% of preprints acknowledged using an LLM for some purpose, close to the 19% that cited specialized tools such as AlphaFold, which has been available since 2021. LLM acknowledgments were near zero before 2023 and about 7.5% as recently as January 2026, so the climb has been steep.

The breakdown matters more than the headline. Only about 4% of preprints credit an LLM for research or analysis, meaning work like experimental design, data labeling, or statistics. Most acknowledgments are for writing, coding, and literature review. So the fast growth is driven mainly by general-purpose help around the edges of research, not by models doing the science itself. AlphaFold, by contrast, appears in roughly half of all preprints that cite a specialized tool, because it does a specific scientific job.

Epoch is careful about what the data can and cannot show. The numbers come from voluntary acknowledgments, which reflect disclosure norms as much as actual use, and both the detection and the use-case sorting were done by AI classifiers that can make mistakes. True usage is probably higher than reported.

Why it matters

If you track how AI reaches real scientific work, the gap between 17% disclosing any LLM use and 4% using one for research is the number to watch: adoption is wide but still shallow where it would change results.