Most AI note-takers approved for medical workers by the Ontario government had errors in their testing, the province’s auditor general found in a report released Tuesday.

Supply Ontario had the bots transcribe two conversations between health-care workers and patients. Most of the vendors that were approved had inaccuracies in their results, including “incorrect information, AI hallucinations and incomplete information,” Auditor General Shelley Spence’s report notes.

Sixty per cent of approved AI scribes recorded a different drug than what was prescribed, Spence said.

Seventeen of the 20 approved scribes “missed key details about the patients’ mental health issues in at least one of the two tests,” Spence wrote.

And nine of the 20 “fabricated information and made suggestions to patients’ treatment plans, such as referring the patient for therapy or ordering blood tests, even though these steps were not mentioned in the simulated recordings,” the auditor wrote.

Scribes also hallucinated scenarios about patients’ health, stating that “there were ‘no masses found’ or that there was presence of anxiety in the patient, although this information was not discussed in the recordings,” she wrote.

The province did not put much weight on accuracy in its testing. “Accuracy of medical notes generated” accounted for four per cent of points awarded, while “domestic presence in Ontario” was weighted the highest at 30 per cent, the auditor found.

“Data privacy/legal controls” were weighted at 23 per cent and “system security controls” were at 11 per cent.

Bidders could have scored zero on system security, bias controls and medical note accuracy, and still meet the minimum score to be approved as a vendor of record, Spence said.

The tests also did not have to be done live, in front of the evaluators. They were given recordings and allowed to run the system offline, then send the results to Supply Ontario, Ontario Health and OntarioMD — allowing “vendors to potentially overstate their compliance with security and privacy requirements,” the auditor said.

“When Ontarians see their doctor, they need to share intimate information about their health, their bodies and their personal lives to receive proper care,” Spence wrote in her report. “Ontarians expect this extremely personal information to be kept private and confidential. Using AI to assist in providing health care must not come at the cost of compromising privacy.”

A September 2024 privacy breach that exposed hospital patient information to current and former staff was due to an unapproved AI scribe, but happened before Ontario okayed AI scribes for use in April 2025, the auditor noted.

Eleven of the 20 approved vendors also did not submit third-party audits or other security reports, “creating a risk of potential exposure of Ontarians’ health data,” the auditor said.

Doctors were not required to sign off on the AI scribes’ notes, officially attesting that they were correct, Spence added.

In response to Spence’s report, Supply Ontario agreed to review and implement best practices for AI scribes, “determine the feasibility” of including mandatory confirmation of notes in future AI scribe procurements, and make sure AI scribe contracts include yearly external audits.

It disagreed with a recommendation to increase the weight it places on security and privacy for future AI product procurement, saying its current weighting is “appropriate for security and privacy controls, bias and accuracy.”