What happened
Epoch AI’s October 30, 2026 data insight, by Tara Abrishami, compares acknowledgments of LLM use with acknowledgments of specialized AI tools in bioRxiv preprints. It examines preprints posted from January 2020 through September 2026. In September 2026, 17.19% acknowledged LLM use for any purpose, while 18.94% acknowledged use of a specialized AI tool. The rounded headline figures are 17% and 19%.
The specialized-tool category includes systems for protein structure and design, image analysis, single-cell sequencing, and genomics. AlphaFold, a protein-structure prediction tool that became publicly available in July 2021, is the most commonly acknowledged specialized tool. Epoch says it appears in about half of preprints that acknowledge specialized AI-tool use. LLM uses are more varied: literature review, writing, and coding assistance are among the uses described.
Epoch separately classifies LLM acknowledgments into research, code and computation, writing, and literature review. It defines research use to include conceptual contributions such as experimental design, methodological contributions such as data labeling, and analysis such as statistical analysis. In September 2026, about 4% of biology preprints acknowledged LLM use for research or analysis, up from about 1% at the start of that year. The monthly table gives 3.86% for September.
To produce the analysis, Epoch says it used GPT-5.6 Sol to check each bioRxiv preprint in the period for AI-use acknowledgments, counting LLM and specialized-tool mentions separately. It used the OpenAI Decisions API, based on GPT-6 Luna, to classify LLM-use acknowledgments into the four use-case categories. Epoch describes the work as free to use, distribute, and reproduce with source and author credit under a Creative Commons Attribution license. The insight links to a data explorer and a full methodology.
What the archive adds
New Runtime’s earlier coverage described thousands of distributed AI agents used by OpenAI, PerturbAI, and Scripps to find research leads in data from millions of brain cells (Using thousands of AI agents to uncover new leads in biological data, October 21, 2026). That was an example of agent-assisted biological discovery; Epoch’s analysis adds a broader, disclosure-based view of LLM and specialized-tool uptake across biology preprints.
Trend signal
This is adoption evidence, not a direct measure of capability or research productivity. It supports the broader idea that AI is entering scientific work in multiple roles, but does not establish how much labor it saves or whether findings improve. This does not yet fit a pattern New Runtime tracks.
Why now
Epoch’s series runs through September 2026 and shows the acknowledged share for LLM research or analysis rising to about 4% from about 1% at the start of the year. The current comparison matters because the LLM acknowledgment rate is now close to the rate for specialized AI tools in the same corpus.
Connections to other events
Epoch’s descriptive trend complements New Runtime’s earlier coverage of agent-assisted biological research: that earlier note described a project using thousands of agents to generate leads from brain-cell data, while this analysis counts acknowledged AI use across bioRxiv preprints. The two are different kinds of evidence—one project example and one corpus-level disclosure analysis—and neither alone establishes the overall effect of AI on research outcomes.
Why it matters
For teams building AI for scientific users, the headline comparison is a reminder that general-purpose LLMs and domain-specific tools coexist in real workflows. A product should not assume that a single model replaces specialized systems: literature review, coding, and writing assistance are distinct from tasks such as protein-structure prediction, image analysis, or genomic work.
The 4% figure for acknowledged research or analysis use is also useful as a boundary against overclaiming. It suggests that acknowledged research-facing use is growing, but it remains a smaller category than broad LLM use. Teams evaluating scientific agents should track what the system actually contributes—such as a literature search, analysis step, or experimental design suggestion—and preserve enough provenance for a researcher to review it. Disclosure rates can inform questions to investigate, not answer whether a tool is reliable or valuable.
Limits and uncertainty
Epoch measures voluntary acknowledgments, which reflect both usage and disclosure norms and may underestimate actual use. Its corpus is bioRxiv preprints, which may not represent peer-reviewed papers or biology research as a whole. The detection and use-case classification steps rely on AI models and can be wrong. Epoch also says its list of 150 specialized AI tools may not cover every tool used in biology. The analysis does not establish whether acknowledged use improved research quality, speed, or reproducibility.
Skill to build now
For a research workflow that uses LLMs, add a simple usage record this week: note the task (for example, literature review, coding, or analysis), the model or tool, and what a human checked before relying on the result. Keep writing assistance separate from research or analysis contributions. That makes internal evaluation and eventual disclosure more precise without treating a paper acknowledgment as proof of quality.
Reader takeaway
LLM use is approaching specialized-tool use in bioRxiv acknowledgments, but the numbers describe what authors disclose, not everything researchers do with AI. For builders, the useful next step is to distinguish workflow roles and evaluate each one separately rather than treating “AI use” as a single category.