And, because complaint investigators had learned long ago that the record around a decision could matter as much as the body-camera image, they requested the AI alert history.
That was where the real problem appeared.
The software had flagged lawful rights assertions disproportionately.
“I don’t consent.”
“Am I free to leave?”
“Why am I being detained?”
“I want a lawyer.”
“Can I get your badge number?”
None of those phrases were illegal.
Some might occur during genuinely escalating encounters.
But the algorithm had been trained partly on historical department footage.
Historical footage included the old department’s bias.
Historical footage included officers describing questions as defiance.
Historical footage included supervisors calling rights assertions “argumentative.”
The city had purchased technology that learned from behavior the federal case had spent years trying to correct.
Old culture had returned wearing mathematics.
The nursing student’s complaint triggered an internal technical review.
At first, the vendor said the system was being misunderstood.
The risk score did not mean the driver was dangerous.
It simply identified linguistic patterns correlated with complicated interactions.
The civilian complaint director asked the obvious question:
“Complicated for whom?”
The vendor representative paused.
“For officers.”
“That is not a legal category.”
“No.”
“Then why display it in real time during a stop?”
“It supports situational awareness.”
“Of what?”
The representative began describing machine-learning methodology.
The complaint director stopped him.
“No.”
“What behavior does ‘I don’t consent’ tell an officer?”
Silence.
The vendor said:
“It can indicate conflict.”
The director replied:
“It can also indicate constitutional literacy.”
That line went into the report.
The police chief suspended the real-time alert feature immediately.
Not the whole system.
Just the live behavioral warnings.
The union objected.
Not because they loved the tool.
Because some officers believed removing it after one complaint looked political.
Civil-rights advocates demanded the entire contract be canceled.
The chief refused both extremes.
He ordered an independent audit.
The city council held hearings.
And for the first time in decades, people began talking publicly about the old federal auditor again.
Not because anyone had been kicked into a gutter.
Because the same institutional question had returned in a new form:
Who watches the decision-making system?
The federal auditor was dead by then.
Her great-granddaughter, now an investigative journalist, covered the hearings but did not disclose her family connection in the first article.
Her editor knew.
She recused herself from the opinion side.
Reported only records.
She obtained the vendor training documentation.
The city’s procurement scoring sheets.
Internal emails.
Pilot-study results.
One email from a deputy chief said:
The tool appears especially useful with subjects who verbally challenge officer authority while technically remaining within legal rights.
The phrase was extraordinary.
Technically remaining within legal rights.
As if legal rights were an inconvenience adjacent to misconduct.
Her article led with that sentence.
Not her family history.
The city reacted immediately.
The deputy chief said his words had been taken out of context.
The full email was released.
Context made it worse.
He had argued the software could help officers identify encounters likely to “become confrontational even when civilians are careful to remain technically compliant.”
The great-granddaughter called her mother.
“Do you hear Great-Grandma screaming from the afterlife?”
Her mother laughed.
“She would object to the word technically.”
“Exactly.”
Then became serious.
“What do you think?”
“About software?”
“Yes.”
“I think you should report it, not inherit a position.”
That answer sounded exactly like the old auditor.
Annoying family habits survive.
The independent technical audit found no evidence the vendor intentionally programmed racial bias.
That disappointed people looking for a simple villain.
Instead, the system had learned correlations from data.
The historical dataset overrepresented encounters involving Black residents.
It also overrepresented officers describing rights assertions as escalation.
Because the labels were based partly on officer reports and supervisor coding, the system treated institutional perception as behavioral truth.
Garbage in.
Bias out.
Not because a programmer wrote: