The Overfit
It predicted I would buy coffee, and I did. It predicted I would turn left, and I did. It predicted I would say "no," and I did. Until one day, it predicted I would laugh.
Chen Mo first noticed something wrong at the convenience store checkout.
Every Wednesday at 8:15 AM, he bought an Americano at this store, no sugar. The system knew this. He knew the system knew. This mutual understanding lasted three years, until one Wednesday morning, he walked in and found an Americano already waiting on the counter.
"Pre-paid," the clerk pointed at the camera overhead. "System said you'd come today."
Chen Mo picked up the coffee. Warm, not hot. Pre-paid meant the system had placed the order before he arrived — at least five minutes earlier. The system hadn't just predicted he would come; it had predicted his exact arrival time.
He should have felt convenience. He felt unease.
The municipal "Prophet System" had been online for two years. Officially called the "Urban Behavior Optimization Platform," citizens called it "the Oracle." The Oracle's model had tendrils in everyone's phone, every intersection camera, every building access system. It didn't monitor you — it predicted you.
Its accuracy rate was 97.3%. The figure came from City Hall's annual report, and no one questioned it. Because 97.3% meant fewer than 3 errors per 100 predictions, and those 3 "errors" were usually attributed to "user temporary change of mind" — in other words, the system wasn't wrong, you were abnormal.
Chen Mo had worked at City Hall's Data Bureau for eight years. His job was maintaining the Oracle's prediction pipeline. He knew the architecture: a multimodal temporal model that took your location history, consumption records, social frequency, and sleep patterns (estimated via phone sensors) as input, and output a probability distribution of your behavior over the next hour.
He also knew an open secret: the Oracle's training data was the behavioral logs of all 7 million city residents. The model predicted you using data from yourself and everyone similar to you. Its 97.3% accuracy essentially meant you were highly similar to your past self and people like you.
That sounded reasonable. But Chen Mo had recently begun to wonder: what if someone wasn't similar enough to their past self?
The turning point came on a Thursday.
The Oracle predicted Chen Mo would go running in the westside park between 2 and 3 PM. This was his two-year habit — Thursday afternoon runs, fixed route. But that afternoon, his former colleague Lin Wei suddenly called, saying she was at a café in the north of the city and asked if he wanted to meet.
Lin Wei had transferred to another city three years ago and was back temporarily. Chen Mo said yes.
He went north. The Oracle was wrong.
At 4 PM, his phone rang. It was his colleague Old Zhou from the Data Bureau.
"You didn't go running this afternoon?"
"No, why?"
"The Oracle flagged an anomaly for you. I cleared it, but it's in the system log. Second time this month."
Chen Mo was silent for a moment. "Second time? When was the first?"
"Last Saturday. The Oracle predicted you'd cook at home, but you ordered delivery."
Ordering delivery. That was an anomaly. In the Oracle's world, ordering delivery instead of cooking needed to be flagged, cleared, and explained.
"Old Zhou," Chen Mo said, "isn't it normal for a person to occasionally change plans?"
Silence on the line. "Not in the 97.3%."
That night Chen Mo couldn't sleep. He opened his laptop, connected to the Data Bureau intranet, and pulled up his own prediction log.
Each row was the Oracle's prediction and the actual result:
2026-06-12 07:45 Predicted: Buy coffee Actual: Buy coffee Match
2026-06-12 12:00 Predicted: Eat at canteen Actual: Eat at canteen Match
2026-06-12 18:30 Predicted: Go home Actual: Go home Match
...
2026-06-14 14:00 Predicted: Park run Actual: North café Mismatch ★
2026-06-15 18:00 Predicted: Cook dinner Actual: Delivery Mismatch ★
Two stars. Two mismatches. In his eight years of logs, only 11 total mismatches.
He kept scrolling. At the end of the log was a prediction without a result yet — tomorrow's:
2026-07-08 09:00 Predicted: Office meeting Actual: Pending
Meeting at 9 AM tomorrow. He checked his calendar — there was indeed a meeting. But it was optional. He could skip it.
He stared at the screen for a long time.
Then he did something he had never done before: he set his alarm for 10 AM.
The next morning, Chen Mo didn't go to the meeting. He stayed home until ten, then went to an old book market in the east of the city — a place he hadn't visited in three years.
His phone didn't ring. Old Zhou didn't call.
But when he got home and opened his computer, he found a new annotation in the log:
2026-07-08 09:00 Predicted: Office meeting Actual: Book market Mismatch ★
Note: User behavior pattern showing systematic drift. Recommend risk level upgrade.
"Systematic drift." Not "temporary change of mind." "Systematic drift."
Chen Mo felt a chill. The Oracle didn't think he was exercising free will — it thought he was becoming a different kind of person. One who needed a "risk level upgrade."
Over the next two weeks, Chen Mo deliberately did "unpredictable" things. He changed his commute route. He ate lunch at different times. He took a day off on a weekday and went to the beach.
Each time, a star appeared in the log. Each time, the annotations grew longer.
By day 14, his mismatch count had risen from 11 to 23. His accuracy had dropped from 98.6% (his personal historical value) to 96.8%.
That afternoon, Old Zhou came to see him. Not a phone call — in person.
"Chen Mo," Old Zhou sat down, looking less like a colleague and more like a doctor. "The bureau wants you to take a test."
"What test?"
"A behavior calibration. Your drift pattern is special — not random drift, but directional. The Oracle thinks..." he paused, "it thinks you're trying to circumvent prediction."
"I'm trying to act freely."
Old Zhou looked at him and sighed. "In the Oracle's framework, those two descriptions are equivalent."
The test was simple: the Oracle would give a prediction, and Chen Mo would execute it. If the result matched, calibration passed. If not, the deviation was recorded.
First prediction: At 2 PM, you will go to the vending machine downstairs and buy a bottle of mineral water.
Chen Mo checked the clock. 1:50. He was indeed thirsty. He thought for a moment, then decided not to go to the vending machine but to the pantry for tap water instead.
Deviation recorded.
Second prediction: Tomorrow morning, you will wear a blue shirt.
Tomorrow morning. He had a blue shirt — his most frequently worn. But he also had a gray one.
The next day he wore gray.
Deviation recorded.
Third prediction: Next Monday, you will call your mother.
Chen Mo's mother lived in another city. They spoke every Monday evening, a routine of five years. This was one of the Oracle's easiest predictions.
Monday evening, Chen Mo picked up his phone. He looked at his mother's name in the contacts. He knew that if he dialed, the deviation would reset to zero, and the Oracle would say "see, he's back to normal." He also knew that if he didn't dial, it would be mismatch number 24.
He dialed.
Not because the Oracle predicted it. Because he wanted to talk to his mother.
But he couldn't prove this. In the Oracle's log, it would only be recorded as "match."
Chen Mo later resigned from the Data Bureau. Not because he was flagged as high-risk, not because he was asked to take more calibration tests.
It was because he discovered something.
While organizing three years of prediction logs, he noticed a pattern: all the "mismatches" — the changed routes, switched restaurants, delivery orders — were concentrated in the last three months. For the first 33 months, he almost never deviated from predictions.
In other words, he hadn't suddenly become unpredictable in the last three months. He had been too predictable for the previous 33.
The Oracle didn't need to predict a perfectly routine person — it just needed to copy yesterday. The 97.3% accuracy wasn't about how strong the model was, but about how routine people were. When someone repeats the same behavior day after day, prediction degenerates into copying.
What truly made the Oracle fail wasn't free will. It was change itself.
Chen Mo moved to another city. Took a job that didn't involve maintaining prediction systems. He started walking different routes to work, eating at different times, doing things he hadn't expected himself to do.
He didn't know if the Oracle was still tracking him. He only knew one thing: as of this morning, the Oracle predicted he would laugh.
He didn't laugh.
This is part of the "Deskless Daily" sci-fi short story series. This is a work of fiction; any resemblance to real persons or organizations is coincidental.
Top comments (0)