A thread currently circulating on Lobsters AI focuses on Pangram — a small company (around 20 employees, ~$4 million in seed funding) that claims to have solved one of the hardest problems in modern AI use: actually knowing whether something was written by a human or a machine.

It isn't necessarily the product itself that's drawing attention. It's the claims behind it.

Pangram says its Pangram 3.0 model achieves a false positive rate of 1 in 10,000 — meaning only one in ten thousand texts written by a human is incorrectly flagged as AI. If accurate, that figure is genuinely remarkable. Chicago Booth researchers Brian Jabarian and Alex Emi reportedly concluded in 2025 that Pangram is "the only detector" that meets rigorous low-error-rate requirements without sacrificing detection capability.

If Pangram's numbers hold up, this is one of the most underrated AI tools out there right now.

But — and this is where the community discussion gets interesting — even an error rate of 0.01% starts to look serious at scale. Consider: universities applying this tool to hundreds of thousands of submissions. An expert cited in connection with Pangram makes exactly this point: the tool is solid at the aggregate level, but in individual cases — where the stakes are high — even "near zero" errors is too many.

This is the core tension the underground is latching onto. AI detection is a field where mistakes aren't just statistics — they can mean a student fails unfairly, or a journalist loses their credibility.

AI Detection Climbs: Pangram Claims 1 in 10,000 Error Rate - Bilde 1

Source assessment: This is an early signal based on community discussions and the company's own published figures. The numbers have not been independently verified by 24AI, and Pangram's own studies should be read with that in mind. The Chicago Booth research is real, but as far as we know it has not been peer-reviewed in the traditional sense.

Why is this worth following? Because the race between AI generation and AI detection is one of the most underrated technological contests right now. Pangram is not alone in the market, but it is one of the few players publishing concrete error rates and inviting scrutiny. If its methods withstand that scrutiny, this could become a standard tool in media and education far sooner than anyone expects.

Worth keeping an eye on.