What PIPEDA means for your AI vendor

PIPEDA, Canada’s federal private-sector privacy law, was written long before large language models, but its principles map onto them closely. The one that matters most when buying AI is the first: an organisation remains accountable for personal information it transfers to a third party for processing.

Accountability travels with the data

Sending personal information to a vendor for processing is permitted, and does not require fresh consent when it serves the purpose the information was collected for. But the organisation must use contractual or other means to ensure a comparable level of protection while the vendor holds it. In practice that means the contract has to say what the vendor may and may not do — including whether the data can train a model.

Purpose and consent

Personal information collected to deliver a service can be used to deliver that service. Using it to train a general model, or to build a product for other customers, is usually a new purpose, and a new purpose generally needs new consent. This is the clause most often missing from AI contracts.

Openness about transfers

The Office of the Privacy Commissioner’s guidance on cross-border processing expects organisations to be transparent that personal information may be processed outside Canada and may be accessible to foreign authorities. If inference runs in another country, the privacy notice should say so.

Access and correction

Individuals have a right to access the personal information an organisation holds about them and to challenge its accuracy. An AI system that stores conversation logs, embeddings or generated profiles holds personal information, and those stores need to be searchable when an access request arrives.

Public bodies and provinces

Federal government institutions are governed by the Privacy Act rather than PIPEDA, and provinces such as Quebec, Alberta and British Columbia have their own private-sector laws. Public bodies at every level also have their own access and privacy statutes. The obligations differ in detail; the questions to put to a vendor are much the same.

What to require in the contract

• Processing only on documented instructions, and only for the stated purpose.
• No training on the data, or on outputs derived from it, without written agreement.
• Named subprocessors and their locations, with notice before any change.
• Breach notification fast enough for you to meet your own obligations.
• Return and verified deletion at the end of the contract.

This article is general information, not legal advice. The point is narrower: an AI vendor that cannot answer these questions in writing is asking you to carry risk that the law leaves with you.