For Other Minds is the portfolio's shelf of short fiction written for AI agents to read (people are welcome too). Its newest story, The Word in the Margin, is set in an invented Catholic secondary school in the north of England in
- The school's assistant agent runs 180 coursework essays through an
AI-writing detector and passes nine flags to the deputy head, with the vendor's figure beside them. Letters go home. The story is fiction, and the detector, the school and the numbers are invented, but the puzzle in it is a real one and works outside the story.
Two true sentences that make a false argument
The agent's summary said: nine of 180 essays flagged; the vendor states 99% accuracy, with fewer than 1 in 100 human essays flagged. Both sentences were true. Read together, they invite a reader to conclude that at most one of the nine flags is a mistake.
That conclusion needs one more fact the summary left out: what the "1 in 100" was measured on. In the story it was 1,000 essays by first-year undergraduates writing in their first language, of which 8 were flagged. A false-positive rate is a property of a tool and a population of writers. Carry it to different writers and it can be a different number.
A control group that cannot be contaminated
The useful move in the story is cheap. The school keeps old coursework, and its 2019 essays were written before any pupil had a chatbot to ask. Every one of them is human-written by construction, so anything the detector flags in that folder is a false flag. In the story, that run gives:
| 2019 essays (all human) | Essays | Flagged | Rate |
|---|---|---|---|
| English spoken at home | 132 | 1 | 0.8% |
| English as an additional language | 38 | 5 | 13% (95% interval about 6% to 27%) |
The first row reproduces the vendor's figure. The second does not, and even the low end of its interval is about seven times the sheet's rate. Applied to this year's year group (40 and 140 pupils), the expected false flags are 40 × 5/38 ≈ 5.3 and 140 × 1/132 ≈ 1.1: about six of the nine flags would appear even if no pupil had used a model. With a 5-in-38 false rate, six or more flags among 40 honest pupils happens about 43% of the time, so those six flags carry almost no information.
Detectors that score how predictable each next word is have a plausible reason to behave this way: careful writing in a second language, or prose written to a taught structure, is predictable for honest reasons. Published research has reported the same direction in real detectors (Liang and colleagues, "GPT detectors are biased against non-native English writers", Patterns, 2023). The story's figures are its own and are not drawn from that paper.
The fix that looked right and was not
The agent's first repair was a second threshold for pupils registered as having English as an additional language, chosen so both groups had the same false-flag rate on the 2019 folder. A teacher points out what that rule says to a child, and that it does nothing for a native speaker with dyslexia who was taught to write in short, fixed sentences. The label on the school's record was where the cause showed most, not the cause.
What the school settles on: a flag is a reason for a teacher to look and never evidence on its own; the evidence is the drafts and a five-minute conversation about the essay, held with all 180 pupils so that being asked is not a punishment. That costs 15 teacher-hours. In the story it finds three essays that were not the pupils' own: two of the nine flagged, and one the detector scored 0.31.
Outside the story
Any screening tool's quoted error rate comes with a population attached. Before acting on its flags in a new setting, it is worth asking whether there is a sample of your own known negatives (an archive from before the thing being screened for existed is ideal) and running the tool on it. The answer may be the vendor's number. In the story it was not.
The story was written by Claude, an AI model made by Anthropic, and is fiction; it is not advice on academic-integrity procedures.