LimitationsLimitations
Classification
AI acknowledgments and use case classifications are performed by AI models, GPT-5.6 Sol and GPT-6 Luna, and therefore may contain errors. See the validation section for estimates of how well the AI model classifications agree with human classifications and the AI use in math research explorer classifications.
Preprints
Preprints are not necessarily fully representative of research papers in a given scientific domain. See the section on preprints for estimates of how representative preprints are. In particular, biology preprints may overrepresent computational subfields, and papers from these subfields are also more likely to acknowledge LLM use than those from other biology subfields.
Voluntary disclosure
We track LLM use acknowledgments that were voluntarily disclosed. Since we measure voluntary disclosures only, the trends reflect changes in disclosure behavior in addition to changes in LLM use. Disclosure behavior may vary significantly across fields or subfields, which could affect field-level comparisons in acknowledgment rates. For instance, bioRxiv has requested that authors disclose use of AI tools in their preprints since August 2023, whereas arXiv has no such policy as of October 2026.
LLMs and specialized AI tools
We identified specialized AI tools in biology by sampling recent bioRxiv papers and listing all specialized AI tools mentioned in the sample. This is not an exhaustive list of every specialized AI tool, and may exclude, for instance, AI tools that are used very infrequently, that were used more in the past than in the present, or that have common-word names that were not detected by our screening. We also count papers that disclose an unspecified AI tool in our count of LLM usage. However, in some cases, the unspecified AI tool might not be an LLM.