The case in three lines. A school's AI-writing detector, "99% accurate", flags Leyla's essay. Everyone treats that as 99% proof, until Malik points out there are two ways to be wrong, Prei counts what the flags mean in a big round number, and Corah finds who the detector was tested on. At the meeting, the teacher asks Tara, who has said nothing all week, what she thinks, and the case stops there.

One question to ask. "What did Tara say?" Then listen. There is no right sentence. A good one names something true about at least two of the people in the room.

One thing to do at the table. Take a napkin. Write 10,000 in a box. Pick how rare the thing is (one in a hundred? one in ten?) and how often the test is right, and count the flags, real and false. Change only one number at a time and watch what happens to the share of false flags.

The grown-up layer. The case is about base rates: a test's accuracy alone does not tell you what a positive result means; you also need to know how common the thing is in the people being tested. The same arithmetic sits under medical screening, airport security and spam filters. The second idea is that the errors are not spread evenly: a detector trained and tested on one kind of writing can be wrong far more often for another, and a published 2023 study found exactly that for non-native English writers. Both ideas are worth talking about with a teenager who will spend years being sorted by systems like this.

The case leaves Tara's sentence blank on purpose. Ask for it before you offer yours.