Before asking what Canadian copyright law says about AI output, ask what the tool’s own terms of service already say. Most mainstream AI vendors already answer the ownership question contractually — the harder part is knowing exactly what that answer covers, and what it cannot.
Key takeaways
A clickwrap agreement is still a contract, and a contract is exactly the mechanism Section 13(4) of the Copyright Act requires for moving a copyright interest between parties: it must be in writing and signed by the owner, and accepting a vendor’s terms by clicking “I agree” satisfies that formality for whatever interest the vendor is actually granting. The question worth reading closely is what interest that actually is, because vendors phrase this differently and the differences matter.
Terms that say the vendor “assigns” all rights in output to the user are making a different promise than terms that say the user is granted a licence to use the output as they see fit. The practical difference shows up later: an assignment is meant to leave the vendor with nothing, while a broad licence can leave the vendor holding rights it never gave up, even if the user experiences no day-to-day difference using the tool. treadstonelaw.ca’s comparison of terms of service, privacy policies and EULAs covers this exact distinction for software agreements generally — ownership, exclusivity and scope have to be spelled out, not assumed from how generous the language sounds.
This is the limit that matters most and gets missed most often. If a generated image or block of text incorporates copyrighted material the vendor itself had no right to use — the training-input question covered on a companion piece on this hub — then the vendor’s promise to assign the user “all rights” in the output cannot manufacture a clean title that does not exist. A contractual grant transfers what the grantor has; it does not launder a defect in how the underlying material was produced. This is exactly why some vendor terms include an indemnity for third-party IP claims arising from output — and why the presence or absence of that indemnity is one of the more consequential things a business skips reading.
Where a business commissions a custom AI build rather than using an off-the-shelf product, the ownership question stops being about reading someone else’s standard terms and becomes something the business can actually negotiate. treadstonelaw.ca’s guide to custom software development agreements sets out exactly what should be pinned down in that kind of agreement: who owns the underlying model versus the fine-tuning data versus the output it produces once deployed, what happens to those rights if the relationship ends, and whether the vendor can reuse anything learned from the engagement on a different client’s project.
A single AI vendor agreement often answers two separate ownership questions in two separate places: what happens to the data a business feeds into the tool, and what happens to what the tool produces in return. treadstonelaw.ca’s explainer on data-ownership clauses in SaaS agreements covers the first question in detail — the clause that determines whether a vendor can use a customer’s inputs to improve its model for other customers is a data-ownership and data-use question, not an output-ownership one, and reading only the output clause can miss it entirely.
A business uses a mainstream generative tool under terms that assign it all rights in output, then later discovers the same terms let the vendor use the business’s prompts and outputs to further train the underlying model unless the business opts out in account settings. Both clauses are real and both were accepted by clicking the same box — the output-ownership assignment did not disappear, but it now sits beside a separate data-use permission the business may not have intended to give. Reading the agreement for both questions, not just the one about who owns the output, is what catches this before it matters.
Businesses reading AI vendor terms tend to go straight to the sentence about who owns the output and stop there. Three other clauses usually carry more practical weight: an indemnity clause, covering who bears a third-party IP claim arising from generated output; a confidentiality clause, covering whether anything typed into the tool can be used to train future versions or shown to other customers; and a termination clause, covering whether previously granted output rights survive if the vendor relationship ends or the vendor itself is acquired. An ownership clause that looks generous is worth less if any of these three quietly takes it back.
A design agency generates marketing concepts for a client using a vendor whose terms assign it full output rights, then wants to hand those concepts to the client as fully owned deliverables. Before promising that, the agency needs to check three things against the vendor’s terms: whether the assignment runs to the agency’s account holder and can be passed along to a third party like the client at all; whether the vendor’s confidentiality terms permit sharing the prompts and iterations that produced the final concept; and whether any indemnity for third-party claims travels with the deliverable once it changes hands. An ownership clause that reads cleanly for the agency’s own internal use does not automatically read the same way once the output is being resold or handed off to someone else.
Generally yes, as a contract between the user and the vendor — but it only transfers what the vendor actually held to begin with, and it does not resolve, on its own, whether the output qualifies for copyright protection at all.
Terms of service can generally be updated prospectively, which is why the effective date and any grandfathering language in a vendor's terms matter — a business relying heavily on a specific ownership grant should track whether that grant survives a terms update.
Not necessarily. Free and paid tiers commonly carry different data-use and output-ownership terms from the same vendor, so the specific tier's terms — not the vendor's general reputation — are what actually govern a given use.
Related: who owns AI-generated work in Canada and licensing your own content to an AI firm.
A custom engagement is the point where ownership of the model, the data and the output can actually be pinned down in writing.
