Staff were already using AI. Most of the sites were not vetted.
Between April and August 2025, about 12,000 Ontario Public Service (OPS) staff accessed roughly 400 AI-related websites.
The Auditor General scored those sites with Microsoft Defender, on a scale of zero to ten where a low score means higher risk. Sites scoring five or lower were treated as unsafe or unsecured. That was 244 of them, about 60%. Of the sites scoring five or lower, 15% were also not work-related and carried inappropriate content. (p. 12)
The audit's first recommendation asks the OPS to review and block access to unsafe and unsecured AI websites.
AI websites accessed by OPS staff, by Defender safety score
The report gives “roughly 400” and “about 60%”; the bar is drawn at 244 on a 0–400 scale, so the unfilled remainder is approximate.
The tool the province chose drew about 6% of use.
From April to August 2025, other popular generative AI websites made up 94% of OPS staff's usage. Microsoft Copilot Chat, the tool the OPS provides, made up about 6%.
The audit follows this with two recommendations: set and report adoption targets for Copilot Chat (Recommendation 2, p. 15), and block Copilot Chat on browsers other than Microsoft's while using training to explain the risk (Recommendation 3, p. 16).
OPS staff usage: Microsoft Copilot Chat versus other popular GenAI websites
The AI-risk course exists. Almost nobody had taken it.
By August 2025, 1,800 of 55,000 OPS staff, about 3%, had completed the responsible-AI training. The course was not mandatory. (p. 13)
Recommendation 1 asks the OPS to put measures in place so that all staff take the training, with refreshers as its AI use grows.
OPS staff who had completed AI-risk training
Twenty approved vendors, two test recordings, and a health-care record at stake.
AI scribes listen to a clinician's conversation with a patient and draft the note. The Auditor General tested the 20 vendors on the approved list using two simulated audio recordings. (p. 22)
Nine of the 20 produced notes with fabricated information, including suggested treatment steps such as a therapy referral or blood tests that were never mentioned. Twelve captured a different drug from the one the doctor prescribed. Seventeen missed key details about the patient's mental health. (p. 23)
On paperwork, 11 vendors did not submit SOC reports, HITRUST certification or ISO 27001 certification, and five did not submit threat risk assessments and privacy impact assessments. (p. 25)
AI scribe vendors flagged, by check
How to read this. The report is anonymised and gives aggregate counts only; it has no vendor-by-vendor table. Marks fill from the left in each row, so position says nothing about which vendor, and the report does not say whether the same vendors appear in more than one row. The document rows are grouped as reported: audit and certification reports (11), and threat risk and privacy impact assessments (five). The testing used two simulated recordings, so these are findings from a small test, not a measure of error rates in clinical use.
Ten recommendations, twenty actions.
The report closes each finding with a numbered recommendation. By our count they carry 20 separate action items.
The summaries below are Northern Signal paraphrases with report page numbers; read the report for the exact wording. This build does not include the government's responses or later progress.
The Auditor General's ten recommendations
| # | What it asks (paraphrased) | Actions | Page | Theme |
|---|
The statute lets Ontario regulate public-sector AI. The regulations are the missing piece.
The Enhancing Digital Security and Trust Act, 2024 (from Bill 194) enables regulation of AI use in the Ontario public sector. (pp. 5–6)
Under it, the Minister can set technical standards and the Lieutenant Governor in Council can set requirements. According to the report, no AI regulations had been finalized when the audit was written. The Assembly's bill page describes the bill's rules as depending on regulations yet to be made, and lists topics they may cover: public disclosure of AI use, accountability and risk management, human oversight, prohibited uses and technical standards.
In a nutshell..
How many AI websites did Ontario public servants visit, and how many did the Auditor General judge unsafe?
About 12,000 OPS staff visited roughly 400 AI-related websites between April and August 2025. Scored with Microsoft Defender, 244 — about 60% — came in at five or lower, meaning unsafe or unsecured; 15% of those also carried inappropriate, non-work content.
How much of OPS staff's AI use went through Microsoft Copilot Chat, the province's approved tool?
About 6%. The other 94% of usage went to other popular generative AI websites the OPS hasn't vetted or approved.
What share of OPS staff had completed AI-risk training?
About 3% — 1,800 of 55,000 staff — by August 2025. The course wasn't mandatory.
How did the approved AI medical-scribe vendors perform in testing?
The Auditor General tested all 20 approved vendors against two simulated recordings. Nine produced notes with fabricated information, 12 recorded a different drug than the one prescribed, and 17 missed key mental-health details. Eleven vendors hadn't submitted required audit or certification reports, and five hadn't submitted threat-risk or privacy-impact assessments.
How many recommendations did the report make, and what do they cover?
Ten recommendations totalling 20 action items, covering staff web access and training, Copilot Chat adoption targets, procurement validation and bid-scoring criteria, and AI-scribe vendor oversight.
Does Ontario currently regulate AI use in the public sector?
The Enhancing Digital Security and Trust Act, 2024 gives the province the power to set AI rules, but as of the audit no regulations had been finalized — the Act's requirements depend on regulations that haven't been written yet.
How to use this page
This is a reference build of a report published on 12 May 2026, not a news update. Everything shown is an audit finding by the Auditor General, from fieldwork carried out between January and November 2025, and should not be read as a description of Ontario's systems today.
Sources
- Office of the Auditor General of Ontario, Use of Artificial Intelligence in the Ontario Government, special report, 12 May 2026, 48 pp. PDF · HTML edition. © King's Printer for Ontario. Figures and page references cited with attribution; the document is not reproduced.
- Office of the Auditor General of Ontario, news release, 12 May 2026.
- Legislative Assembly of Ontario, Bill 194, 43rd Parliament, 1st Session. Sources accessed 30 September 2026.