The AI proctoring system was called MingCha. It ran from ceiling cameras, watching all 180 students simultaneously.
Lin Shuhao sat at seat 37, unaware of the machine's existence. He was taking his final exam — data structures and algorithms. He had crammed the night before and his mind was blank.
At minute 45, he reached question 8. A graph traversal problem. Nothing came to him.
He instinctively looked up.
On the notebook of the student ahead, he could faintly make out the letters "BFS."
Lin Shuhao hesitated for three seconds.
Then he looked down and wrote "BFS."
On the ceiling, MingCha logged the event: head raised 17 degrees, lasting 2.1 seconds, gaze direction toward the seat ahead. Behavior matched threshold for "reference copying" alert, confidence 0.78. Flagged as "pending human review."
A week after exams, Lin Shuhao received a notice from the academic office: materials referred to the academic integrity committee. Results within 15 business days.
He found his advisor and appealed.
The advisor pulled up the monitoring footage and showed him the 2.1-second screenshot.
Lin Shuhao was quiet for a long time. "I did look up. But I didn't copy anything."
"MingCha only flagged the behavior. Humans made the final call. We compared your answer with the student ahead — both wrote BFS, but your derivation is completely different."
"So the result is?"
"No academic misconduct. But the behavior stays on record. If triggered again within three years, it becomes automatic determination."
Lin Shuhao stood in the hallway after leaving the office.
He thought about high school exams. The proctor was a woman near retirement who always dozed off by the midpoint. Cheating wasn't caught then — not because no one cheated, but because no one was watching.
Now someone was watching. Not a person. A machine that never dozed off.
He wondered: was this more fair, or less fair?
He had looked up. He hadn't copied anything — just borrowed a direction, letting his own brain remember the answer.
Did that count as cheating?
Four years later, Lin Shuhao joined a company building AI systems. His first project: upgrading exam proctoring for a provincial education bureau.
He added one parameter to the code: confidence_threshold=0.85.
Raising the trigger threshold from 0.78 to 0.85.
The test engineer asked why he changed it.
Lin Shuhao said: "Reduce false positives."
More sci-fi stories: wdsega.github.io
Top comments (0)