Software had spotted the theft of fentanyl in Bakersfield: no one followed up on the alert

At Adventist Health, machine learning had reported opioid diversions from a temporary nurse. Alerts went unanswered.

At the end of September 2024, at Adventist Health Hospital in Bakersfield, California, an interim nurse recruited a few weeks earlier by an agency collected fentanyl and morphine from a secure cabinet, while recording in the records of administrations patients who never occurred. A patient she was in charge of endured severe pain, believing that she was receiving the painkillers on her file; relatives described erratic behaviour in intensive care units.

Technology was working, the human chain was not

The central point of the case is not an algorithmic failure. The machine learning software deployed to detect drug abuse had clearly identified suspicious motives and issued reports. These alerts remained dead: pharmacists who were supposed to instruct them did not act, and the supervision of the institution did not follow them up. The Centers for Medicare and Medicaid Services opened an investigation in November 2024.

The case illustrates a structural limitation of hospital algorithmic surveillance: these tools do not prevent anything by themselves. Without a procedure for handling alerts, dedicated staff and clearly assigned responsibility, detection is reduced to a trace that is searchable after the fact — useful to the investigation, useless to the patient.