When the detector says you didn't write it — lessons from the NeurIPS 2026 desk rejections
July 5th, 2026, by Guy Harris
On June 2, 2026, the organisers of NeurIPS, the machine learning field's flagship conference, published a blog post announcing that 178 submissions to its Position Paper Track (18.4% of the total) had been desk-rejected on the basis of scores from an AI-detection tool called Pangram. There was no appeal.
A further 123 authors received a different notice: document your "pre-AI and post-AI writing checkpoints" by June 15, or face rejection. In other words, authors were asked, after the fact, to produce evidence that a human wrote their paper.
Detectors get it wrong
The backlash centred on false positives. Several authors reported that manuscripts they had written almost entirely themselves were flagged as AI-generated. The reported triggers were surprisingly superficial: heavy use of em-dashes and certain punctuation patterns pushed scores upward.
More troubling still, the results swung with the detector's settings. Depending on the detection window applied, the flag rate moved between 42.7% and 12.7%. Authors could not see these settings. Which side of the line a paper landed on was decided by parameters invisible to the person who wrote it.
For researchers writing in English as a second language, this is not an abstract risk. A 2023 study in Patterns showed that GPT detectors disproportionately misclassify text by non-native English speakers as AI-generated. The more carefully correct the English, the more machine-like it can appear to a classifier.
The age of "prove it"
The episode points to two new realities.
First, conferences and journals have begun using detector scores as a gate on submissions, while the detectors themselves remain imperfect. The verdicts are running ahead of the accuracy.
Second, the burden of proving human authorship is shifting onto authors. NeurIPS asked 123 authors for records of their writing process. How many researchers could produce such records on request, today?
What authors can do now
Nothing dramatic is required. Ordinary habits are the defence.
- Keep a record of your writing process. Dated draft versions, file history, correspondence with co-authors. A visible writing history is the strongest evidence there is.
- Read the venue's AI policy before you submit. Does it require disclosure, or prohibit use, and of what exactly? Policies differ sharply between venues, as we have discussed here before.
- Reread your final draft for machine habits. Whatever tools you used along the way, the submitted text is yours. Uniform rhythm, stock phrases, mechanical punctuation: human writing is less tidy than that.
The detector era has begun, ready or not. The best position an author can be in is to be able to tell the story of how the paper was written.
