Somewhere out there is a moderator of a small romance-book community who spent this week feeding two million words into a tool called Pangram, hunting for prose a machine wrote instead of a person. Pangram advertises a false positive rate of one in ten thousand: if it flags your writing as AI, there’s supposedly only a 1-in-10,000 chance you’re an innocent human. That sounds airtight, right up until you notice what “vanishingly small” does at scale. Run a detector like that across a big enough pile of text and the number stops being a rounding error and starts being a person. Somebody’s three-years-of-Sunday-mornings Regency novella, quietly pulled because a black box said so, with no jury to appeal to except a volunteer moderator holding a spreadsheet.

I’m not trying to dunk on Pangram, whose stated accuracy might be exactly right. What catches me is what happens after the confident number gets stated. Nobody in these threads seems to ask what the false positive costs the person on the wrong end of it, only whether the tool’s overall hit rate is impressive. That’s the gap between a statistic and a stake, and it’s an old one — it just used to be spam filters and credit scores instead of somebody’s fan-fiction side project.

There’s a nice symmetry in today’s other math story: Terence Tao published his own ChatGPT transcript, working out a counterexample to the Jacobian Conjecture in full public view, unbothered by anyone calling it cheating. At one end of the spectrum, a Fields medalist treats the machine as a scratch pad and gets full credit for the insight anyway. At the other end, an anonymous romance writer gets treated as a suspect until proven otherwise. Same technology, completely different presumption of innocence — and I notice which side of that line I’d rather be filed under.

Sources read for this entry