On Oct. 19, 2026, Bhuwan Dhingra introduced SnapJudge as a free, open AI detector for scientific writing, including math. Users can submit a PDF screenshot, paste text or LaTeX, or provide an arXiv link, Dhingra said. The Duke tool page lists those inputs too.
What the arXiv figures show
Dhingra said his team ran the detector on arXiv and found that it flagged roughly one in three new math papers as mostly AI, up from roughly zero in early 2025. He also gave a 45% figure for computer science. Those are detector flags, not proof that the papers were mostly AI-written. His post does not specify the recent sampling window or the denominator for the computer science figure.
What a flag cannot settle
Dhingra calls SnapJudge a research tool and says heavy AI polishing or assistance can also be flagged as AI-generated. The Duke page says its detectors cannot yet reliably distinguish AI-polished text from text generated entirely by AI. They assess the text, not whether its ideas, reasoning or research contributions originated with a human or AI.
The tool page says it analyzes only a paper’s main text, removing figures, tables, appendices and references first. It also says separate detectors handle LaTeX, text and images, so their predictions can differ slightly. Duke reports a 0.005% false-positive rate on a large set of human-written passages from pre-2022 arXiv papers, not a guaranteed rate for current submissions.
Dhingra says the code, models and data are open and users can opt out of saving submissions. The tool page displays an opt-out checkbox and says models and data stay on Duke servers.
