60% of AI note-taking tools record errors: Ontario audit report warns automation practitioners

A new report from the Ontario Audit Office shows that 60 per cent of AI medical note-taking tools have serious factual errors, including mixing up medications and fabricating patient descriptions. This is a warning to all practitioners who use AI for automation: establishing a content verification system is more important than ever.

Core conclusion

A new report from Ontario's Audit Office finds that an AI note-taking system used by hospitals and clinics frequently makes serious factual errors. Audit inspections show that 60% of AI medical record tools make errors when recording prescription drug information, and some systems even go so far as to write patient A’s medication records in patient B’s medical record. This news is an important warning sign for all teams using AI to automate content production and information aggregation.

Key Points

Source of the incident: The Auditor General of Ontario issued a special audit report in May 2026
Error Rate: 60% of evaluated AI note-taking systems had prescription drug confusion errors
Scope of Impact: Multiple medical institutions in Ontario, Canada, directly affecting patient medication safety
Deep Problem: The hallucination problem is common in AI summary/note-taking applications, and the consequences are particularly serious in high-risk scenarios such as medical care.

Background: Why are AI medical note-taking tools so popular?

In the past two years, AI-driven medical note-taking tools (AI Scribe) have rapidly become popular in the North American medical industry. Doctors use these tools to automatically record consultations, generate medical record summaries, and extract medication information—theoretically reducing administrative burdens for doctors and allowing them to focus more on patients.

However, in the actual implementation process, AI has repeatedly exposed its fatal weakness in "seemingly simple" summary tasks such as meeting records and medical notes: fabricating facts. An investigation by Ontario's Audit Office revealed a harsh reality - these systems not only obfuscate drug names but also add descriptions of symptoms that patients never mentioned.

One HN user commented: "The AI recording tool used by our company will write things that the CIO has not said as 'commitments', causing tension in supplier relationships. Without recording playback to check, you won't find out at all."

Key Impact

Dimensions	Change	What it means to us	Recommended actions
AI credibility	Medical-grade AI note-taking tools have a deviation rate of up to 60%	Do not blindly trust any AI content production/summarization tool	Add human review nodes to the content pipeline
Industrialization risks	AI errors will amplify exponentially after scale	The more automated assembly lines there are, the faster errors accumulate	Introducing regular verification + cross-validation mechanism
Regulatory expectations	Countries may follow up to review the reliability of AI in production environments	Compliance costs will rise, and pure AI content may face more stringent scrutiny	Establish a content traceability and review system in advance
User trust	The problem of AI illusion is no longer a technical problem, but a crisis of trust	Transparent statements (such as ContentDisclaimer) will become standard	Clearly mark the degree of AI participation and provide verification paths

Adaptation suggestions: How to avoid errors in your AI workflow

1. Add a verification layer to the content

Don’t treat the AI’s output as the final result. Refer to the logic of Ontario’s audit:

Key data fields (product name, price, date) are constrained by structured schema instead of allowing AI to generate them freely
Perform regular verification on sensitive fields (such as amount format, URL structure)
Establish a "human-machine collaboration" process: AI generates first draft → rule engine verification → manual spot check

2. Introduce cross-validation

Use two different models to summarize the same content and compare the differences
Attach original links or screenshots to statistics and citation sources required
Establish a "fact review" database: every time AI outputs a specific number or name, it automatically records the source

3. Draw boundaries for AI tools

The error in medical notes stems from letting AI "play freely" instead of filling in formatted templates
Constrain AI output with templates + fixed fields, don’t let it generate important data from scratch
Explicitly specify "Do not add information not mentioned in the original text" in the AI Agent prompt

Code Example: Content Validation Framework

python


# AI
import re

def validate_ai_content(content):
    warnings = []
 #
    numbers = re.findall(r'\d{8,}', content)
    if numbers:
 warnings.append("，")
    return warnings

AI content validation workflow

AI

** AI **。 Claude Code 、 n8n 、 AI ——AI 。

，

**/， AI **
**， AI **
** AI ——，**

OpenAI、ChatGPT、Claude、DeepSeek、n8n、Claude Code、Gemini、Hermes Agent

Related extended information

Internal link guidance

Want to build a reliable AI automated workflow? Watch: AI Agent-Driven Content Automation: n8n MCP Building Guide from Scratch
Real case: He used an AI automation system to achieve a monthly income of $12,000, but only if he knew the limitations of AI: He Built an AI Automation Stack with Claude + n8n — $4K to $12K/mo in 6 Months