LimitationsLimitations
AI acknowledgments/use case classification
The AI acknowledgments and use case classifications are performed by an AI model, GPT-5.6 Sol, and therefore may contain errors. We estimate the extent of errors in classification in the AI acknowledgment and use case classification section. Judging the extent to which AI contributed to research solely from AI use disclosures is difficult and imprecise, even for human judges. Every assigned tag is publicly auditable through the github repository associated with this data explorer. Classification errors can be reported at math@epoch.ai.
Voluntary disclosure
We track AI acknowledgments that were voluntarily disclosed in the main text of arXiv preprints. Since we measure voluntary disclosures only, the trends could reflect changes in disclosure behavior in addition to changes in AI use. We do not use any other measure of AI use, such as indirect signals or text patterns that correlate with AI use. Some signals of AI use, such as AI prompts and AI use disclosures, are present only in source file comments and not in the main text. We exclude such signals from our data.
Author identification
Author identification may be inaccurate or incomplete. In particular, authors whose names are common, who are earlier in their careers, and/or who lack an arXiv ID or ORCID are more difficult to uniquely identify using our methodology. Author errors can be reported at math@epoch.ai.