“Anything containing coercion terms routes human.”
“We need technical change.”
“Then call Ethan.”
“Already awake?”
“He will be.”
Ethan answered on second call.
He sounded exhausted before anyone explained.
Meaning his team had already found it.
“Model overweights dominant-language detection.”
Ava said:
“English.”
“Spanish in this case.”
“You know what I mean.”
“Yes.”
“Why?”
“We trained for conversational continuity.”
“If user switches languages, it avoids interrupting.”
“Who decided continuity mattered more than uncertainty?”
Ethan went quiet.
Ava could hear typing.
Finally:
“We did.”
“Why?”
“User abandonment rates.”
There it was.
Metric.
Not malice.
People hung up when systems repeatedly asked which language.
Product team optimized.
That optimization made system less likely to admit uncertainty.
Ava closed her eyes.
Her mother would have understood immediately.
Not because AI.
Because it was the same old temptation in modern clothes.
Do not interrupt.
Do not inconvenience.
Do not make customer repeat.
Keep flow smooth.
Smoothness had hidden risk.
Ethan said:
“We are disabling legal intake globally.”
Ava opened eyes.
“You said counties—”
“I know.”
“We’ll pay for human overflow.”
“How?”
“Emergency contractor network.”
“That will cost.”
“Yes.”
“Your investors going to like?”
“No.”
“Good.”
He laughed tiredly.
Ava almost smiled.
Mateo was eventually located.
Safe.
He agreed to speak with human interpreter.
The case expanded.
His employer faced civil penalties and criminal investigation for coercion.
Language Access House apologized directly.
Not a press release first.
To him.
Mateo said:
“I told machine.”
Meera answered:
“Yes.”
“It did not hear me correctly.”
“No.”
“Then why did it say it understood?”
No one in the room had an answer they liked.
That question became the center of redesign.
Not accuracy percentage.
Not translation quality.
When should a system admit it does not know?
BridgeSpeak changed interface.
When language confidence fell below threshold in any high-risk context, software stopped pretending continuity.
It said:
I may not be understanding you accurately. A human interpreter is joining.
Simple.
Annoying.
Slower.
Better.
Six months later, Ethan stood before Language Access House board.
No slides about ninety-eight-point-seven percent.
Instead first screen read: