The hammer rang once. A spark jumped off the anvil and did not go out. It waited. They followed.

It was a Tuesday in late October, and the Whetstone was nearly empty. Tara had the back table, the one by the radiator, with her chemistry notes spread out and a cup of tea she had stopped drinking an hour ago. She liked the back table because from there she could hear the whole café without anybody expecting her to join in. She was good at that. Her mother called it listening. Her brother called it lurking.

Leyla came in with her coat still buttoned and sat down opposite her without saying hello.

Leyla had moved here two years ago, from a city Tara had only seen in photographs. Her English was better than most people's in their year, but it was a careful English, the English of someone who had learned it properly: full sentences, no slang, every comma where the rules said it should go. Teachers loved her essays. Tara had read some of them. They were plain and exact and a little bit stiff, like a pressed shirt.

Leyla put a sheet of paper on the table and turned it round so Tara could read it.

It was a printout from the school's new system. At the top was the title of Leyla's history essay, the one about the salt marches. Underneath, in a red box, it said: Likely AI-generated (97%). And in the corner was a little logo with a sticker beside it, the kind the company put on everything: 99% accurate.

"Mr. Okafor wants to see me on Thursday," said Leyla. "With my mother."

Tara read the page twice. Then she looked up.

"Did you?" she said, very quietly, and immediately wished she hadn't.

Leyla looked at her for a long moment. "No," she said. "I wrote it at my desk. Three evenings. I have the notes. I have the books from the library with the pages folded down. My mother watched me do it." She pressed her lips together. "And now you are asking me too."

"I'm sorry," said Tara. "I didn't mean it like that."

"Everybody means it like that," said Leyla. "That is the problem. It says ninety-nine percent. Who argues with ninety-nine percent?"

By Wednesday lunchtime, everybody in their history class knew.

It was not that anybody was cruel about it. That was almost worse. People were kind and sorry and certain. Josh, who sat at the back and had once been caught copying a lab report word for word, said that the detectors were "basically perfect now." Hana, who was class rep and liked rules because rules were fair, said that the school had to use something, because people really were handing in essays they had not written, and that wasn't fair on the people who did the work. That was true. Tara knew two people it was true of.

"It's not personal," said Hana. "It's just a number. Ninety-nine percent accurate means it's right ninety-nine times out of a hundred. So if it says Leyla's essay is AI, it's ninety-nine percent sure. Those are the odds. I feel awful for her, but that's maths."

Several people nodded. It sounded like maths.

Tara opened her mouth. Then she closed it again. She had a feeling that something was wrong with what Hana had said, but she could not have said what. It was like a word on the tip of her tongue. So she did what she always did. She listened.

Malik, who was eating an apple on the window sill, did not nod.

Malik ran the statistics club, which had four members and met in the maths room on Fridays. He was tall and slow-moving and spoke as if every sentence cost him money.

"That's not maths," said Malik. "That's a shortcut."

Hana turned round. "It's literally what the number means."

"It's literally what the number says," said Malik. "What it means is a different question." He took another bite of his apple. "There are two ways to be wrong here. Has anybody said which one they're willing to risk?"

Nobody had.

"Well," said Hana, after a pause, "the school is risking letting cheaters get away with it if they don't use the detector."

"That's one," said Malik. "What's the other?"

The bell went before anyone answered. But Tara noticed that Hana did not repeat the thing about the odds on the way out. She was frowning, the way people frown when a door they thought was shut turns out to be ajar.

That evening there were four of them at the back table of the Whetstone: Tara, Leyla, Malik and Danya.

Danya was three years younger and was not supposed to be there, but he had followed Tara from the Green, and once Danya had decided to be somewhere it was easier to let him stay. Danya's whole personality was the question "how do we know?" He asked it about history, about adverts, about the weather forecast, about whether the café's cakes were really homemade. Most people found it exhausting. Tara found it restful, because Danya asked out loud the questions Tara only asked in her head.

Malik drew a line down the middle of a napkin.

"Every test like this can be wrong in two directions," he said. "On this side: the essay was written by a machine, and the detector says it was written by a person. It misses. On that side: the essay was written by a person, and the detector says it was a machine. It accuses."

"Which is worse?" said Danya.

"That's the question," said Malik. "That's a choice, not a fact. A school that's terrified of cheating sets the dial so it misses almost nothing, and then it accuses more people. A school that's terrified of accusing someone innocent sets it the other way, and some cheating gets through. You can't make both go to zero. You can only choose which mistake you'd rather make."

"And our school chose?" said Tara.

Malik shrugged. "Our school read the sticker. The sticker doesn't say which mistake it makes. It just says ninety-nine."

Leyla was holding her tea with both hands. "So it could be wrong about me," she said. "That is what you're saying."

"It could be wrong about anyone," said Malik. "The question is how often. And how often depends on something nobody in our class has thought about."

"What?" said Danya at once.

But Malik only shook his head. "I'm better with dice than with essays," he said. "Ask Prei. Prei builds spam filters for fun. This is a spam filter with a uniform on."

Prei turned up the next afternoon with a laptop and a packet of crisps and an air of someone who had been looking forward to this.

"Okay," said Prei, sitting down. "Two rules. One: a flag is not a fact. Two: we use a big round number, because percentages lie to your brain and counts don't. Let's say ten thousand essays. That's a lot of schools for a year."

Prei drew a box on a fresh napkin and wrote 10,000 in it.

"Now," said Prei. "How many of those are actually written by a machine? Nobody knows exactly. Mr. Okafor thinks it's going up. Let's be generous to the detector and say one in a hundred. One percent."

"That seems low," said Danya.

"Maybe," said Prei. "We'll change it in a minute and see what happens. For now, one in a hundred. So out of ten thousand essays, a hundred are machine-written." Two boxes now: 100 on one side, 9,900 on the other.

"The detector is ninety-nine percent accurate. Let's be generous again and say it's ninety-nine percent accurate both ways: it catches ninety-nine out of every hundred machine essays, and it clears ninety-nine out of every hundred human ones." Prei wrote under the 100: caught: 99. Then under the 9,900: wrongly flagged: 1 in 100 of these.

"What's one percent of nine thousand nine hundred?" said Prei.

Tara did it in her head. "Ninety-nine."

"Ninety-nine," said Prei, and wrote it down. "So. How many red boxes does the detector print, in total?"

"Ninety-nine and ninety-nine," said Danya slowly. "A hundred and ninety-eight."

"And of those hundred and ninety-eight red boxes," said Prei, "how many are real?"

Nobody said anything for a moment. Leyla had gone very still.

"Ninety-nine," said Tara.

"Ninety-nine out of a hundred and ninety-eight," said Prei. "Of everything it flagged, how many were real? Half. Half of the red boxes are people like Leyla." Prei tapped the napkin. "Ninety-nine percent accurate. And if it flags you, it's a coin toss."

Danya stared at the napkin as though it had done a magic trick. "That can't be right."

"Count it again," said Prei, and pushed the pen across.

Danya counted it again. It was right.

"Now change the one in a hundred," said Prei. "Say cheating gets more common. Say it's one in ten. Do it."

Danya did it, muttering. A thousand machine essays; nine hundred and ninety caught. Nine thousand human essays; ninety wrongly flagged. "Then most of the flags are real," he said. "Nine hundred and ninety against ninety."

"Same detector," said Prei. "Same sticker. Totally different meaning for the person holding the red box. The ninety-nine percent didn't change. What changed?"

Tara knew the answer. She could feel the shape of it, the way you feel a step in the dark. But she did not have a word for it yet, and she did not want to say something half-made. So she said nothing, and wrote the two napkins' numbers carefully into the back of her chemistry notebook.

On Thursday morning Corah found Tara at her locker.

Corah wrote for the school paper and was the kind of person who read the terms and conditions before clicking accept. She had heard about Leyla from Malik, who had heard about it from Prei, which was how news moved in their year.

"I read the fine print," said Corah. "On the company's website. Read it with me."

She held up her phone. Under the big 99% accurate there was a small grey link, and under the link there was a page, and near the bottom of the page there was a sentence: Accuracy measured on a benchmark set of essays.

"Whose essays?" said Corah. "That's what I wanted to know. Who did they test it on? So I went looking." She scrolled. "There's a study. Real scientists, published in a real journal in 2023. They took detectors like this and gave them essays by native English speakers and essays by people writing in a second language. The detectors did fine on the first lot. On the second lot, they flagged a huge share of perfectly human essays as machine-written."

Tara read it. Then she read it again. "Why?"

"Because the detectors look for writing that's very plain and very predictable," said Corah. "Simple words, tidy sentences, nothing surprising. Which is how machines write. It's also how a lot of careful people write when they learned the language from a textbook and are trying very hard not to make mistakes."

Tara thought of Leyla's essays. Plain and exact and a little bit stiff, like a pressed shirt.

"So the one in a hundred," said Tara slowly, "isn't one in a hundred for everyone."

"It's not spread evenly," said Corah. "For some people it's a lot worse than one in a hundred. Who got left out of the testing changes the answer." Her voice was steady, but Tara could hear that she was angry, the quiet kind of angry that has decided to do something about it. "And the sticker doesn't mention that either."

The meeting was at four o'clock, in Mr. Okafor's room, and somehow it was not only Leyla and her mother.

It was not supposed to be a crowd. But Leyla had asked Tara to come, and Tara had asked whether Malik and Prei and Corah could wait outside, and Mr. Okafor, who was not a villain, had looked at the four of them in the corridor and sighed and said, "Fine. Come in. All of you. Sit at the back and don't interrupt." Danya had somehow come too, and nobody had the energy to send him away.

Mr. Okafor was tired. He had been teaching for twenty years. This term, he told them, he had received essays that were so obviously written by a machine that they had described events that never happened, quoted books that did not exist, and in one case thanked him at the end for "this helpful prompt." He had sat at his kitchen table at midnight feeling like a fool. And the students who had worked hard, really hard, deserved better than to be marked on the same scale as someone who had typed a question and pressed a button.

"I don't want to accuse anyone," he said. "I want to be fair to the ones who did the work. That includes you, Leyla. That's why I'm asking."

It was the strongest version of his side, and Tara found she believed every word of it.

Leyla's mother spoke next, in a careful English that was very like her daughter's. She described three evenings at the kitchen table, the library books, the folded pages, the drafts in pencil. Leyla put her notes on the desk. They were good notes.

Mr. Okafor looked at them for a long time.

Then he looked at the printout, with its red box and its little sticker.

"The trouble," he said, mostly to himself, "is that the system says ninety-nine percent."

Malik, at the back, did not say anything. He had said his part on the window sill. Prei did not say anything either, though the napkin was in Prei's pocket. Corah had her phone face down on her knee, the study open on the screen.

And then Mr. Okafor did something nobody expected. He turned and looked at the back row, and his eyes went past Malik and Prei and Corah and stopped at Tara, who had not said one word all week.

"Tara," he said. "You've been very quiet. You usually are. But you're Leyla's friend, and you've clearly been thinking about this. What do you think?"

The room went still. The radiator ticked. Somewhere down the corridor, a door closed.

Tara had listened to all of them. Hana, who wanted it to be fair. Malik, with his two kinds of mistake. Prei, with the napkin and the coin toss. Corah, with the fine print and the people left out. Mr. Okafor, at his kitchen table at midnight. Leyla, holding her tea with both hands.

Everyone in the room was partly right.

Tara opened her mouth.

 

 

(Write the sentence Tara says.)

 

The spark dimmed to a warm dot and settled back on the anvil. The Green was just the Green again.

Back on the Green

Two weeks later, on the Green, Danya's little cousin came running up with a leaflet. This test spots the rare disease in 99 out of 100 people who have it! it said, over a picture of a smiling doctor.

Danya took the leaflet and read it.

"How rare is the disease?" he asked.

Tara, who was sitting on the anvil steps with her chemistry, smiled without looking up. She knew exactly what Danya was going to do next. He was going to get a napkin and write a big round number in a box.

Afterword. The idea in this case has a name: base rates. How common or rare a thing is to begin with changes what a warning about it means. When the thing a test looks for is rare, even a very accurate test will raise many false alarms, because there are so many more chances to be wrong about the ordinary cases than right about the rare ones. Doctors, judges, spam filters and schools all face it. You can play with it in PriorForge and OddsForge, and with the two kinds of mistake in FilterForge.