An AI scribe drafts clinical notes from a patient doctor conversation, and an Ontario audit tested 20 of the 31 AI scribe vendors pre qualified by the province and found accuracy or completeness problems across the board.
Ontario's auditors fed 20 pre-approved AI medical-note tools a pair of simulated clinical visits and checked what came back. Every vendor had problems. Nine of twenty hallucinated content the audio never contained. Twelve produced transcripts that contradicted what was said. Seventeen dropped mental-health details the conversations had included, according to the Ontario Auditor General's 2026 special audit.
The audit is a study of how prediction engines fail when audio gets hard. AI scribes generate the next plausible word in a sequence. They do not record what the doctor or patient actually said. When the recording is dense (overlapping voices, regional accents, technical medical terminology) the generated text can drift from what was said. The mistakes are not typos: hallucinated content, incorrect transcription, and dropped details can change a diagnosis in a clinical note, or alter the record of what was said in a legal proceeding.
The auditors found this after approximately 5,000 Ontario physicians had already begun using AI scribes in their practices. No patient-harm reports had been filed at the time of the audit. The provincial government attributed the absence of harm to the rollout being still in its testing phase, as CBC reported. Ars Technica's coverage walks through the same finding at a public-facing level. Both outlets treat the Ontario AG report as the load-bearing reference, not a separate dataset.
The audit was independent. The procurement step that approved these tools was not. Supply Ontario, the province's central procurement body, had pre-qualified 31 AI scribe vendors under tender 20123, effective from April 2025 through April 2028. The list names ADGScribe, Heidi Health, Tali AI, Solventum Fluency Align, and Scribeberry among others. Pre-qualification rested on vendor-supplied claims about accuracy and security. The Ontario audit was the first independent test of how those claims held up against realistic clinical audio.
The same prediction engines are now being aimed at court and legal transcription, on claims that have not been independently audited. The most aggressive case for that risk comes from a commercial human-transcription firm with a competing service to sell. Ditto Transcripts has published testing it says shows AI tools averaging 61.92% accuracy against 99% for human transcription across roughly 1,000 hours of dictation, with an 18% critical-error rate. The company's CEO supplied the commentary that anchors a Digital Journal piece tying the audit findings to the legal transcription market. That framing is a competitor's argument, not an independent benchmark, and no independent court-records audit has been published to test it.
The structural question the audit opens is whether the verification step can be tightened without unwinding deployment. The 31-vendor procurement pipeline that approved these tools remains active through April 2028. Replacing vendor self-attestation with an audit gate would mean re-testing or rejecting tools already cleared by the province's own pre-qualification process and used by thousands of physicians.
The Ontario audit gives the verification step a public benchmark. The 31-vendor pipeline that approved these tools is still live through April 2028. What comes next is the open question: an audit gate at the procurement step, a vendor recall, or another year of self-attested claims. Health systems that pre-qualified AI scribes on vendor self-attestation now have the same benchmark. What they do with it is the policy decision the Ontario audit has put on the table.