AI Meeting Notes in Canada: Consent, PIPEDA and Accuracy

One-party consent under the Criminal Code, what PIPEDA and the OPC's generative AI principles add, and a bilingual test showing how French clauses break transcripts.

By · · 11 min read

Illustration for AI Meeting Notes in Canada: Consent, PIPEDA and Accuracy

Canada edition. Written for Canadian readers: Criminal Code section 184 instead of US state wiretap laws, PIPEDA consent and the OPC's generative AI principles, and an accuracy test with a French clause in an English meeting. US edition

Can you let an AI note taker record and transcribe a meeting in Canada? Under the Criminal Code, a participant in a conversation can consent to it being recorded, so Canada is often described as a one-party consent country. That answers the criminal question only. If your organization is recording clients, customers or (in federally regulated workplaces) employees, privacy law asks a different question: did people know what was being collected, why, and who would see it? And whatever the law says, a transcript of a meeting where people switch between English and French can be badly wrong in exactly the sentences that matter.

This edition covers section 184 of the Criminal Code, what PIPEDA and the Office of the Privacy Commissioner of Canada (OPC) expect when an AI tool is involved, and a word error rate test with a bilingual sentence using the Speech to Text and Note Taker tools. It explains the rules in general terms; for a recording you may rely on in a dispute, an investigation or a regulated field, or for a Quebec or other provincial question, get legal advice.

The criminal law: one participant's consent

Section 184(1) of the Criminal Code makes it an offence to knowingly intercept a private communication by means of an electro-magnetic, acoustic, mechanical or other device; on indictment, the maximum is five years. Section 184(2)(a) says the offence does not apply to a person who has the consent, express or implied, of the person who originated the communication or of the person it was intended for.

In practice that means a participant in a call can record it, or consent to a note taker recording it, without everyone else agreeing. It does not make a recording by an outsider lawful: a bot that joins a meeting nobody on it invited, or a tool that records a call its user is not part of, is outside that exception. Because the Criminal Code is federal, the consent rule is the same in every province, so a call between Toronto and Vancouver does not raise the state-by-state question US calls can. Provincial privacy laws are a separate matter.

Privacy law: notice, purpose and consent

An AI note taker collects personal information: voices, names, opinions and anything said about other people. For private-sector organizations doing commercial activity, the Personal Information Protection and Electronic Documents Act (PIPEDA) applies except where a province has a substantially similar law: Alberta, British Columbia and Quebec have their own private-sector privacy laws, though PIPEDA still applies to personal information that crosses provincial or national borders, as the OPC explains. For employee information, the OPC explains that PIPEDA covers only federal works, undertakings and businesses (such as banking, aviation, telecommunications and interprovincial transport), and that several provinces have their own laws for employee information.

The OPC's Guidelines for obtaining meaningful consent set out what people must be able to see quickly before they consent: what personal information is collected, which parties it is shared with, for what purposes, and the risks of harm. They say consent should generally be express when the information is sensitive, when the use is outside a person's reasonable expectations, or when it creates a meaningful residual risk of significant harm. Health and financial information are generally sensitive.

For meetings, that has practical consequences:

  • A client call about money or health is likely to involve sensitive information, so a clear, spoken request for consent before recording is the safer choice than relying on an on-screen icon.
  • Where the audio goes is one of the key elements. If your note taker is a third-party service that stores audio and transcripts on its own servers, or uses them to improve its models, that is sharing with another party for a purpose people may not expect. Say so.
  • A meeting that is not about the recorded person's business (a job interview, a performance discussion, a support call) still collects personal information, and the purpose for recording it has to be one a reasonable person would consider appropriate.

What the OPC's generative AI principles add

On December 7, 2023, Canada's federal, provincial and territorial privacy commissioners published Principles for responsible, trustworthy and privacy-protective generative AI technologies. Several apply directly to organizations using an AI note taker:

PrincipleWhat it asks of an organization using the toolIn a meeting
Legal authority and consentWhere consent is the authority, make it valid and meaningful; avoid deceptive designAsk at the start, and switch the bot off if someone declines
Necessity and proportionalityEvaluate the validity and reliability of the tool for the intended purpose; tools should be accurate throughout their intended lifecycleTest it on your own meetings, including bilingual ones, before relying on it
OpennessTell affected people whether generative AI is used and howSay the summary is AI-assisted when you send it
Limiting collection, use and disclosureKeep only what the purpose needs; set retention schedules for prompts and outputsDecide how long recordings and transcripts are kept, and delete the rest
AccuracyPersonal information should be as accurate as the purpose requiresCheck names, figures and decisions before the notes go out

The accuracy and reliability rows are where a Canadian team's test needs to be different.

Worked example: a bilingual sentence breaks the transcript

Many Canadian meetings switch languages mid-sentence. An English-only recognizer will try to turn French into English words, and the result is far worse than a few misheard names. Write a reference passage the way your meetings actually sound:

Mathieu will send the revised forecast by Thursday. We are not renewing the Laval lease. Le lancement est reporté au 13 avril and the rebate stays at thirteen percent.

  1. Paste it into the Note Taker. The word counter shows 29: that is N.
  2. In Speech to Text, pick a language and read it aloud at meeting pace. The picker offers English (US) and French, but no English (Canada) or Canadian French, so test both choices on your own speakers.
  3. Compare the transcript with the reference word by word, or use the Online Diff Tool with one word per line and Ignore case ticked.

Suppose the English recognizer returns this (an illustrative result, not a measurement of any product):

Matthew will send the revised forecast by Thursday. We are renewing the Laval lease. Love on some on a report to 13 April and the rebate stays at thirty percent.

Part of the passageReference wordsErrors (substitutions / deletions / insertions)Word error rate
English sentences only222 / 1 / 03 ÷ 22 = 13.6%
The French clause76 / 0 / 28 ÷ 7 = 114.3%
Whole passage298 / 1 / 211 ÷ 29 = 37.9%

Word error rate is substitutions, deletions and insertions divided by the reference words, as NIST's scoring tool counts it, so it can go above 100% when a recognizer adds words, as it did here. Seven French words produced eight errors and lost the launch date's meaning entirely ("reporté", postponed, became "report to"). The English errors are fewer but worse: the dropped "not" reverses a decision about the lease, and "thirteen" became "thirty".

A summary written from this transcript would confidently say the Laval lease is renewed and the rebate is 30%. That is the accuracy problem the OPC principles point at: a tool that is reliable in one language can be unreliable in the meetings you actually hold.

A five-minute check before the notes go out

  1. Keep your own list during the meeting of decisions, figures, dates and owners. The Note Taker saves as you type and nothing it holds is recorded audio.
  2. Check each decision in the summary against your list, and anything else against the transcript or recording.
  3. Search for "not", "unless", "only if", "ne", "pas" and "sauf" near each decision, and for every number.
  4. Check names against the invitation, especially French names an English model anglicizes ("Mathieu" to "Matthew").
  5. Remove what should not travel, including side remarks and anything said after someone asked to go off the record.
  6. Send it as a draft: "AI-assisted notes; reply with corrections by Friday."

The Speech to Text tool uses your browser's speech service; in Chrome and Edge the audio goes to the browser maker's servers, so the same notice and consent questions apply if you use it in a meeting with others, and it is not the place for confidential material.

Tools for this job

  • Speech to Text: read a test passage with your team's names, numbers and French phrases to see how a recognizer handles them.
  • Note Taker: word counts for the reference passage, and your own record of decisions during the meeting.
  • Online Diff Tool: compare the reference and the transcript one word per line.

The Web Speech API guide explains where browser speech recognition sends audio, and the text diff guide explains how the diff lines up two texts.

Frequently asked questions

Is Canada a one-party consent country for recording?

For the Criminal Code, yes: section 184(2)(a) exempts a person who has the express or implied consent of the person who made the communication or the person it was meant for, so a participant can consent. Privacy laws such as PIPEDA can still require organizations to tell people and, for sensitive information, to get their express consent.

Do I need consent to use an AI note taker with clients?

If your organization is subject to PIPEDA, the OPC's consent guidelines expect people to know what is collected, who it is shared with, why, and the risks, and to give express consent for sensitive information such as financial or health details. Asking at the start of the call, and turning the bot off if someone declines, is a practical way to address that expectation; it does not by itself make a recording compliant.

Can my employer record meetings with an AI tool?

It depends on who your employer is. PIPEDA covers employee information only in federally regulated businesses such as banking and aviation; elsewhere, provincial law and any collective agreement apply. The OPC's workplace guidance says employers should limit collection to what is necessary for identified purposes.

Why are bilingual meetings transcribed so badly?

A recognizer set to one language maps the other language's sounds onto words it knows. In the worked example, a seven-word French clause produced eight errors. Test with the language mix your meetings really have, and check every decision by hand.

Sources and further reading

Browse all tools