Skip to content

Field test · Detector v10 · November 2026

AI detection in scientific papers

NeurIPS checked its 2026 position papers for AI-written text with Pangram. We ran the same check with ParaTrace.

Before screening its position papers, NeurIPS tested Pangram on papers accepted at an AI ethics conference, some from 2022, before ChatGPT, and some from 2025. We ran the same check with ParaTrace v10. On 106 papers from 2022 it flags none, at any level, exactly like Pangram. On 123 papers from 2025 it flags 25 as at least half AI-written, 20.3 % against 1.0 % for Pangram. No record shows which of those papers used AI, so the gap shows where the two detectors disagree, not which one is right.

Papers from before ChatGPT

0of 106 papers flagged by ParaTrace, at any AI score

The same result as Pangram, which flagged 0 of 159.

What it shows

Clean on human research, stricter on 2025 papers

No false alarms

0.0 %

of 106 papers from before ChatGPT flagged, at any level, the same as Pangram. The highest AI score among them is 2.7 %, and passage by passage only 0.2 % lean AI.

More papers flagged in 2025

20.3 %

of papers from 2025 reach an AI score of 50 % or more, against 1.0 % for Pangram.

Part of it AI-assisted

35 %

of the text it labels AI in the papers from 2025 is called AI-assisted, human writing that AI edited or paraphrased. The rest is called AI-generated.

The results

Share of papers by AI score

Papers from 2022Written before ChatGPT was released, so any paper flagged here is a false alarm.
DetectorPapers≥ 50 %≥ 90 %100 %
Pangram 3.3.21590.0 %0.0 %0.0 %
ParaTrace v101060.0 %0.0 %0.0 %
Papers from 2025Written when AI writing tools were in wide use, with no record of which papers used them.
DetectorPapers≥ 50 %≥ 90 %100 %
Pangram 3.3.22041.0 %1.0 %0.0 %
ParaTrace v1012320.3 %4.1 %2.4 %

The AI score is the share of a paper’s passages scored as AI. Pangram’s rows are from the NeurIPS post, ours from scoring the papers with the production service.

How we measured

The same check

Papers.
NeurIPS tested Pangram on papers accepted at an AI ethics conference in 2022 and 2025, close in style to its own position papers. We scored 106 papers from 2022 and 123 from 2025 from the same conference, with no errors. Pangram’s figures cover 159 and 204, so compare the shares rather than the counts.
AI score.
Each detector splits a paper into passages of a few hundred words and scores each one. The AI score is the share of passages flagged as AI. A score of 100 % means AI use across many parts of a paper, not that every word came from AI.
No answer key for 2025.
Which papers from 2025 used AI is not recorded, so the 2025 figures can't be checked for either detector. ParaTrace's extra flags could be AI-assisted writing that Pangram doesn't flag. They could be human writing reworded by AI, which ParaTrace counts as AI by design (it flags 52.2 % of AI-paraphrased human text), although NeurIPS allows AI for copy-editing. Or they could be false alarms that papers from 2022 can't reveal, for example on papers about generative AI or on wording that became common with AI tools. The clean 2022 result makes that less likely but doesn't rule it out.
Setting.
ParaTrace v10, scored by the production service in November 2026 at the setting this site uses. Passages from 2022 scored as AI were 9 of 4,577, or 0.2 %.
Training data.
No paper from 2025 and no paper's main text is in the detector's training data. The abstracts of 4 papers from 2022 are, as human text, at most 3 % of each paper. Without those papers the 2022 result is the same, with none of 102 flagged.
Sources.
Pangram’s figures are from the NeurIPS post AI-generated papers in the NeurIPS 2026 position paper track of 2 June 2026. NeurIPS and Pangram Labs, Inc. have not reviewed this comparison.
More results.
See how ParaTrace compares with Pangram on 21 public benchmark results, and our benchmark report.