PIPEDA, Canada's federal private-sector privacy law, already applies to AI systems that collect, use, or disclose personal information in the course of commercial activity, even though it contains no AI-specific provision. Privacy regulators across Canada said so in December 2023, and joint investigations of facial recognition, ChatGPT, and the Grok image tool show how they apply consent, appropriate purposes, and accountability to training data and chatbots.
Key takeaways
- The ten fair information principles in PIPEDA apply to AI systems; the Office of the Privacy Commissioner treats generative AI as fully covered.
- In December 2023, federal, provincial, and territorial privacy authorities issued nine principles for generative AI, and many of the points are legal requirements, not suggestions.
- Findings against Clearview AI (2021), OpenAI (May 6, 2026), and X Corp. and xAI (June 11, 2026) rejected the idea that publicly accessible data can be used for AI without consent.
- The Privacy Commissioner can investigate and make recommendations but cannot currently issue compliance orders.
The ten principles applied to AI
Schedule 1 of PIPEDA sets out ten principles. Two statutory rules sit on top of them: subsection 5(3) permits collection, use, or disclosure only for purposes a reasonable person would consider appropriate in the circumstances, and section 6.1 makes consent valid only if the person could reasonably be expected to understand the nature, purpose, and consequences of what they agree to. The table maps each principle to the AI practice that the regulators' December 2023 document and later findings tie to it.
| PIPEDA principle | What it means for an AI system |
|---|---|
| 4.1 Accountability | Governance, impact assessments, traceability, and a named human accountable for decisions the system supports |
| 4.2 Identifying Purposes | A documented, appropriate purpose for training and for use; testing for misuse before release |
| 4.3 Consent | Valid consent or another documented legal authority; inferences about a person count as personal information |
| 4.4 Limiting Collection | Collect only what the purpose needs; filter personal information from training data where possible |
| 4.5 Limiting Use, Disclosure, and Retention | Retention schedules for training sets, prompts, and outputs |
| 4.6 Accuracy | Update models when training data proves wrong; disclose known accuracy limits; verify outputs used in decisions |
| 4.7 Safeguards | Protection against prompt injection, model inversion, and jailbreaking, scaled to sensitivity |
| 4.8 Openness | Tell people what is collected and why; identify AI outputs with significant impact |
| 4.9 Individual Access | Access and correction rights that reach prompts, outputs, and information held in the model |
| 4.10 Challenging Compliance | A working complaint channel for AI-related concerns |
The December 2023 generative AI principles
On December 7, 2023, the Principles for responsible, trustworthy and privacy-protective generative AI technologies were issued jointly by the Office of the Privacy Commissioner of Canada and its provincial and territorial counterparts, including the offices in Ontario, Quebec, British Columbia, and Alberta. The nine principles are legal authority and consent, appropriate purposes, necessity and proportionality, openness, accountability, individual access, limiting collection, use and disclosure, accuracy, and safeguards.
The document states that generative AI is not exempt from Canadian privacy legislation and that scraping publicly accessible personal information remains subject to privacy law. It uses the word "should", but notes that many of its considerations are legally required depending on the organization and its activities. It also flags children as a group facing heightened risk and calls for extra safeguards such as privacy impact assessments.
Findings involving AI that can be confirmed
Four published findings involve AI or AI-adjacent technology. Each is a regulator's conclusion about specific facts, and each company disagreed with at least part of the result.
- Cadillac Fairview, PIPEDA Findings #2020-004, October 28, 2020: the Office of the Privacy Commissioner, with the Alberta and British Columbia commissioners, found that facial-analysis cameras in mall directories collected about 5,061,324 numerical representations of faces without valid consent. The consent and retention issues were well-founded and resolved, because the company had disabled the technology and deleted the data.
- Clearview AI, PIPEDA Findings #2021-001, February 2, 2021: a joint finding with the Quebec, British Columbia, and Alberta authorities that scraping images from public websites to build a facial recognition database required consent and served a purpose a reasonable person would not find appropriate. The matter was well-founded; the offices recommended that Clearview stop offering the service in Canada and delete Canadian images.
- OpenAI, PIPEDA Findings #2026-002, May 6, 2026: see the next section.
- X Corp. and xAI, PIPEDA Findings #2026-004, June 11, 2026: the Commissioner found that the Grok image tool launched without adequate safeguards and without valid consent for sexualized deepfakes of identifiable people. The matter was well-founded and not resolved. The companies declined the recommendation to suspend Grok Imagine but committed to other measures and to quarterly reports.
The Commissioner opened a separate investigation into X Corp. on February 27, 2025 about training AI models on personal information. A published outcome for that strand was not found as of October 31, 2026.
Training data: consent, purpose, and accountability
The joint investigation of OpenAI's ChatGPT is the clearest statement on training data. The offices accepted that developing ChatGPT was a legitimate purpose, but found that the scale and nature of personal information gathered from public websites and licensed datasets to train GPT-3.5 and GPT-4 was overbroad. They rejected implied consent, because the data included sensitive information, people could not reasonably expect the use, and no meaningful choice was offered. The publicly available information exception did not cover indiscriminate scraping of sites such as social media.
The Office of the Privacy Commissioner found the matter well-founded and conditionally resolved, relying on measures OpenAI introduced, including filtering personal information out of training data, adopting retention and deletion policies, and retiring earlier models. British Columbia and Alberta found the consent issue well-founded and unresolved, and Quebec's commission found consent unresolved while conditionally resolving other points.
Chatbots in practice
For conversational tools, the findings translate into concrete duties. The user-interaction side of the ChatGPT complaint was not well-founded, since using people's chats was found necessary and proportionate to the service. The offices still pressed for notice: OpenAI committed to tell signed-out web users, within three months, that chats may be reviewed and used for training, and to warn them against sharing sensitive information.
Accuracy and access apply to outputs. The offices found that people lacked an easy way to access, correct, or delete their information, and that insufficient notice was given about fabricated answers. On accountability, the Grok findings held the companies responsible for content generated on their platforms even though users wrote the prompts, and faulted privacy impact assessments that did not cover the new tool. Organizations deploying a third-party chatbot remain accountable for personal information under their control under Principle 4.1.
What PIPEDA does not provide
PIPEDA has no provision on automated decisions, no duty to explain an algorithmic outcome, and no express privacy impact assessment requirement for private organizations. In June 2026 the Commissioner noted that his office cannot currently issue orders to compel compliance and called for order-making powers and administrative monetary penalties. The offences in section 28 are limited to listed knowing contraventions and obstruction, with fines up to CAD 10,000 on summary conviction or CAD 100,000 on indictment. Bill C-36 would add several of the missing pieces, but it has not passed. Provincial private-sector laws in Quebec, British Columbia, and Alberta apply alongside PIPEDA, which is why those offices joined the investigations.
This page is general information, not legal advice. Findings describe specific facts, and organizations with a particular situation should consult a qualified lawyer or the original reports.
FAQ
Does PIPEDA apply to AI tools?
Yes, when the tool collects, uses, or discloses personal information in the course of commercial activity. PIPEDA is technology-neutral, and the Office of the Privacy Commissioner has applied it to facial recognition, a large language model, and an image generator.
Can personal data be scraped from public websites to train AI?
Not freely. In the ChatGPT findings, the offices held that the publicly available information exception did not cover indiscriminate scraping, and in the Clearview findings, the offices rejected the same argument. Consent or another valid legal basis was still needed, and the purpose had to be appropriate.
Are the December 2023 principles binding?
The document itself does not create new law. It explains how existing privacy law applies, and it states that many of its considerations are legally required depending on the organization and its activities.
Can the Privacy Commissioner fine a company for an AI privacy breach?
Not currently. The Commissioner can investigate, publish findings, and make recommendations, and he has said his office cannot issue compliance orders. Bill C-36 would create a regulator with order-making and penalty powers.
Related reading
- Bill C-36 and automated decision systems
- Quebec's automated decision rules
- Canada's AI regulation overview
- An AI privacy checklist for small businesses
More in AI Law in Canada.
Last verified: October 2026
Sources
- Personal Information Protection and Electronic Documents Act, Justice Laws Website, Government of Canada
- Principles for responsible, trustworthy and privacy-protective generative AI technologies, Office of the Privacy Commissioner of Canada and provincial and territorial privacy authorities, December 7, 2023
- PIPEDA Findings #2026-002, Joint investigation of OpenAI OpCo, LLC, Office of the Privacy Commissioner of Canada with the Quebec, British Columbia, and Alberta authorities, May 6, 2026
- Backgrounder: Summary of joint investigation into OpenAI's ChatGPT, Office of the Privacy Commissioner of Canada, May 6, 2026
- PIPEDA Findings #2026-004, X Corp. and xAI, Office of the Privacy Commissioner of Canada, June 11, 2026
- News release on the Grok chatbot and sexualized deepfakes findings, Office of the Privacy Commissioner of Canada, June 11, 2026
- PIPEDA Findings #2021-001, Joint investigation of Clearview AI, Inc., Office of the Privacy Commissioner of Canada with the Quebec, British Columbia, and Alberta authorities, February 2, 2021
- PIPEDA Findings #2020-004, Joint investigation of Cadillac Fairview Corporation Limited, Office of the Privacy Commissioner of Canada with the Alberta and British Columbia commissioners, October 28, 2020
- News release on expanded investigation into X Corp. and xAI, Office of the Privacy Commissioner of Canada, January 15, 2026