Not because of AI
Trials and deployments where AI was measured against something and did not help, or made things worse. They sit here under the same grades as the wins. A registry that only records successes is advertising. A clean null tells you more about what AI does than most press releases.
An AI stethoscope rolled out to 205 NHS GP practices did not significantly increase heart-failure detection at 12 months.
Caveat: The tool worked when used; the deployment didn't get it used — a workflow failure, not a model failure.
Null on intention-to-treat
Displaying AI deterioration-risk scores on a cardiology ward did not improve patient outcomes across 10,422 admissions.
Caveat: Passive display with no mandated response, run during COVID-era staffing; sicker patients were preferentially moved into AI-monitored beds.
No change in primary outcome
A GPT-4o tutor built into a university learning platform was associated with lower grades in a 2,379-student randomized trial.
Caveat: Working paper, not yet peer-reviewed; the estimate is the effect of being offered the tutor, not of using it, at one university.
Lower grades (RCT)
AI-personalized computer-assisted learning did no better than non-personalized software for 8,647 rural Chinese primary students.
Caveat: One AI product, one province; the active-control design is exactly what the Sierra Leone trial lacked.
Null vs active control
A GPT-4o assistant built into the medical record improved Kenyan clinicians' notes and plans but did not reduce treatment failure for 9,691 patients.
Caveat: The event rate was lower than planned, so the trial could miss a modest benefit; one private urban network with a high baseline standard of care.
Null on patient outcome (RCT)
Nothing matches.