The Court Transcriber
Zhang Mei spent her entire career as a court transcriber. She typed more words than she ever spoke.
In her final month before retirement, the court introduced an AI assistant—Verbatim 3.0, claiming 99.97% real-time transcription accuracy. "You won't have to work so hard anymore," the president told her. "Just watch the screen and proofread."
Zhang Mei said nothing. She was used to her stenotype machine, used to marking three spaces for silence when a witness stumbled, used to ellipses when someone hesitated.
The AI was different. The AI did not pause. The AI left no room for emotion.
Her first case with Verbatim was a divorce hearing. The wife, sobbing, described her husband's emotional neglect. Her words came in fragments, logic tangled. Zhang Mei listened and mentally rearranged the testimony, preparing to smooth "He just ignores me... always ignores..." into a coherent sentence.
But when she looked down at the screen, Verbatim had already finished.
It read:
"Plaintiff alleges emotional neglect by the defendant but fails to provide specific timelines or behavioral details. Testimony lacks structural coherence. Credibility rating: Below Average."
Zhang Mei froze.
She told no one. Using her thirty years of experience, she quietly re-transcribed the testimony in the margin—the tremor in the broken sentences, the desperation in the repetition, the truth buried in the incoherence.
Every case after that, she found the same thing.
A minor accused of theft. He said "I don't know why" seven times. Verbatim recorded: "Suspect claims absence of criminal motivation. Statement contains 7 repeated denials, indicating evasive behavioral pattern."
A battered wife filing for a protection order. Mid-sentence, she stopped and sat in silence for forty-seven seconds. Verbatim left a blank line with a note: "Witness testimony interrupted for 47 seconds. Cause unknown."
In the margin, Zhang Mei wrote: "She was crying."
On her last day, Zhang Mei compiled three months of proofreading notes into a report titled: "Between Semantics and Pragmatics: An Empirical Observation of Systemic Bias in AI Court Transcription."
She printed three copies. One for the court president, one mailed to the Supreme People's Court Judicial Technology Division, and one left for the young person taking over her position.
On the title page, she wrote:
"Verbatim 3.0 can record 270 words per minute, but it will never record the deep breath before someone speaks. It can annotate a speaker's decibel level, but never the catch in their throat when they lower their voice. It can transcribe every word, but cannot read what hides in the spaces between them.
And justice often hides in those spaces."
That afternoon, after submitting her report, Zhang Mei walked into the empty courtroom one last time. She sat in the transcriber's seat, placed her fingers on the stenotype machine she'd used for half her life, and tapped out one final line.
The sentence appeared on the screen. She looked at it for a long time. Then she stood, turned off the lights, and closed the door behind her.
She had typed it slowly, as slowly as she'd always believed such words should be written:
"Transcription is not turning sound into text. It is turning a person into evidence that can be heard."
Top comments (0)