Pangram

AI-text detectors have a reputation to live down: GPTZero once flagged the US Constitution as machine-written, and OpenAI quietly retired its own detector as unreliable. Pangram is the entrant trying to escape that history — its degree-of-editing method was accepted as a paper at ICLR 2026, and independent testing at the University of Chicago's Booth School put its accuracy at or above 99.8% across most models. Then Pangram turned the same classifier on the conference that accepted it.

Running it over roughly 70,000 ICLR 2026 peer reviews, Pangram estimates about one in five were fully AI-generated, and more than half showed some AI involvement. The reviewers judging AI papers, in other words, were outsourcing the judgment to AI — and those machine-written reviews handed out higher scores than the human ones did. It is the research pipeline auditing itself and finding its own referees asleep.

The catch is what happens to a single person caught in the net. On the podcast, host Nathan Labenz ran 400 essays through Pangram; it flagged four as entirely AI, and one of those was a genuine human essay with fifty-plus minutes of documented editing — scored 0% human. A detector can be the best on the market, trustworthy as an aggregate signal, and still mislabel a real student or job applicant. That gap between a population statistic and an individual verdict is the whole problem: it is why a flag is evidence, not proof, and why the pressure to treat it as proof — in classrooms, hiring screens, and journals wired into these tools — is the risk worth watching.