Overview

Epoch’s AI use in research explorer uses metadata and source text from arXiv and bioRxiv to track AI use acknowledgments in physics, statistics, mathematics, and biology. We use an AI model to read relevant passages in papers with AI keywords and decide whether the paper contains an AI use acknowledgment. We focus on use of Large Language Models (LLMs), rather than specialized AI tools, but we also track specialized AI tool use in biology.

The data is available on the GitHub repository for this project. Any questions can be directed to impacts@epoch.ai.

Data

This explorer uses data from the following sources.

1. arXiv metadata from the arXiv OAI-PMH interface, and arXiv source text from the arXiv Amazon S3 bulk data.
2. bioRxiv metadata from the bioRxiv API, and bioRxiv full text from bioRxiv’s text and data mining bucket.
3. OpenAlex authorship information for each paper sourced from the OpenAlex API.

We thank arXiv and bioRxiv for their open-access interoperability. This project was not reviewed or approved by arXiv or bioRxiv, nor does it necessarily express or reflect their policies or opinions.

Papers

An arXiv paper belongs to a field according to its primary category.

We categorize physics papers into seven subfields, based on arXiv subfield categories: astrophysics (astro-ph), condensed matter (cond-mat), quantum physics (quant-ph), high-energy physics (hep-ex, hep-lat, hep-ph, hep-th), gravitation and cosmology (gr-qc), nuclear physics (nucl-ex, nucl-th), and other physics (the physics.* categories and math-ph). The downloadable data keeps each arXiv category.

Mathematics and statistics papers are categorized according to their arXiv subfield categories. The statistics theory (stat.TH) subfield is an alias of the math.ST subfield, so theoretical statistics papers count under mathematics, not statistics.

We categorize biology papers into eight subfields, based on bioRxiv subfield categories: neuroscience and behavior; molecular and cell biology (cell, molecular and developmental biology, biochemistry, biophysics); computational biology (bioinformatics, systems biology); bioengineering and synthetic biology; microbiology and immunology; genetics, genomics and evolution; ecology and organismal biology (ecology, plant biology, zoology, paleontology); and biomedical and physiology (cancer biology, physiology, pharmacology and toxicology, pathology). The downloadable data keeps each bioRxiv category.

We sourced the full text of the most recent available version of each paper as of October 13, 2026. LLM use acknowledgments added in revisions are counted under the date of the first version of the paper.

LLM acknowledgments

We say that a paper acknowledges LLM use if it acknowledges the use of LLMs or unspecified AI tools for any purpose. We exclude the use of specialized AI tools such as AlphaFold from this count. We used the following process to identify acknowledgments of LLM use.

1. We identified each paper with at least one AI keyword using a regular expression search with the same keyword set as the AI use in math research explorer.
2. For each paper with at least one keyword match, GPT-5.6 Sol read the passages around every keyword match, the paper’s acknowledgment section, and any declaration sections, and decided whether the authors acknowledge LLM use.
3. Papers that only study LLMs, mention them in related work, or describe other people’s use do not count, and neither do denials (“no AI tools were used”). We do count papers that use an LLM as part of their research method, for example to label or extract data.

Use case classification

We use the OpenAI Decisions API to classify each LLM use acknowledgment into five use cases: research, code and computation, writing, literature review, and unspecified. A single LLM use acknowledgment can contain multiple use cases. The Decisions API uses GPT-6 Luna to read text and answer a decision question about the text. In our case, GPT-6 Luna reads passages containing an AI use acknowledgment and, for each use case, answers yes or no: was the LLM used for this purpose? The prompt for each use case is available on the GitHub repository.

Validation

We validated the acknowledgment identification and use case classification in two ways: against our AI use in math research explorer methodology and against a human classification.

Specialized AI tools in biology

Large language models are not the only useful AI tools for science. AlphaFold, for example, is a notable specialized AI tool designed for scientific research. To compare LLM use and the use of specialized AI tools, we also track acknowledgments in bioRxiv preprints that acknowledge use of specialized AI tools. We used the following process.

1. We searched the full text of every bioRxiv paper for the names of more than 150 specialized deep-learning tools across four categories.
   1. Protein structure and design: structure predictors, namely AlphaFold (including the AlphaFold Protein Structure Database, ColabFold, AlphaPulldown and AFsample), RoseTTAFold and trRosetta, ESMFold, OmegaFold, OpenFold, Uni-Fold, HelixFold, RaptorX, Protenix, Boltz and Chai-1, together with AlphaFold’s confidence scores (pLDDT, pTM and ipTM); and protein language models (ESM, ProtT5, SaProt), protein design (ProteinMPNN, RFdiffusion, BindCraft, EvoDiff, ProGen), conformational ensembles (BioEmu, AlphaFlow), function, interface and antibody models (DeepLoc, DeepTMHMM, DeepFRI, MaSIF, IgFold), and docking and molecular property models (DiffDock, GNINA, Chemprop, ADMET-AI).
   2. Image analysis: cell segmentation (Cellpose, StarDist, DeepCell, InstanSeg, U-Net), image restoration (Noise2Void, CSBDeep), pose tracking (DeepLabCut, SLEAP), cryo-EM (ModelAngelo, crYOLO, cryoDRGN), histology foundation models (UNI, Virchow, GigaPath), brain imaging (SynthSeg, FastSurfer), and detection in camera-trap, video and audio recordings (MegaDetector, YOLO, BirdNET, DeepSqueak).
   3. Single-cell sequencing: scVI and its relatives, single-cell foundation models (scGPT, Geneformer, scFoundation), spatial models (GraphST, SpaGCN, STAGATE) and CellBender.
   4. Genomics: variant effect and splicing (AlphaMissense, PrimateAI, SpliceAI, DeepVariant), sequence-to-function models (Enformer, Borzoi, BPNet, AlphaGenome), DNA and RNA language models (DNABERT, Nucleotide Transformer, Evo 2, RNA-FM), and deep-learning genome annotation (geNomad, DeepARG).
   5. Mass spectrometry models (Prosit, MS2PIP, AlphaPeptDeep, Casanovo, InstaNovo and MS2DeepScore) count toward any specialized AI tool but are not a category of their own.
2. For each paper with a keyword match, GPT-5.6 Sol read the passages around each keyword match and decided whether the authors acknowledged using a specialized deep-learning model. Two Claude Opus 5.5 agents checked a random sample of 200 papers and agreed with 98% of GPT-5.6 Sol decisions.
3. AlphaFold accounts for the vast majority, about 95%, of all protein structure model acknowledgments.

Preprints as a sample

This explorer uses preprints as a sample of the research papers in different scientific domains. We estimate how representative the preprints on arXiv and bioRxiv are for each domain. To do this, we identified top-tier journals in physics, mathematics, statistics, and biology, identified every research paper published in one of these journals from 2021 to 2025, and checked how many also appear on arXiv or bioRxiv.

We also include an estimated breakdown of coverage by subfield. For the sake of this estimated breakdown, the arXiv subfield for papers published in journals was determined by GPT-5.6 Sol using the title, journal, OpenAlex topic, and, where available, the abstract of the paper.

Physics (subfields from a random sample of 6,281 papers)

Mathematics (592 papers)

Statistics (2,911 papers).

Biology (17,579 papers).

The analysis of coverage on preprint servers suggests that, especially in biology, varying coverage of preprints by subfield may bias overall results. For instance, computational subfields of biology are better represented among biology preprints than experimental subfields, and are also more likely to acknowledge LLM use; this could lead us to overestimate LLM use in biology preprints overall.