Copyright does not simply attach to anything new and useful — it attaches to a work with an author. A business relying on AI-generated material needs to know where that requirement leaves it, and the honest answer is: it depends on how the material was made.
Key takeaways
Section 5 of the Copyright Act sets the condition for protection to exist at all: copyright subsists in an original work if, among other conditions, the author was a citizen, subject or resident of a treaty country at the time the work was made. Notice what the test assumes — that there is an author, and that the author is a person with a nationality or residence. A work with no human author at any point does not fit neatly inside a test built entirely around one.
Some jurisdictions have answered this more directly. The United Kingdom’s copyright statute contains a specific provision that names the person who made the arrangements necessary for a computer-generated work to be created as its author. Canada has no equivalent provision anywhere in the Copyright Act — a full-text search of the statute confirms it. Importing the UK approach by analogy is a common instinct and an incorrect one: Parliament has not enacted anything like it here, and describing Canadian law as though it had is simply wrong.
The gap above is at its widest for material that comes out of a single prompt with no further human shaping. It narrows considerably where a person is doing real creative work around the AI tool: selecting and rejecting outputs, editing and arranging them, combining AI-generated elements with their own original contributions, and making the string of judgment calls that separates a curated final piece from a raw model output. treadstonelaw.ca’s explainer on copyright basics covers the general shape of this question — what copyright protects and how it arises — and the more of that human judgment is present, the more the finished piece looks like a traditional case of human authorship with a new tool involved, rather than a novel case with no clear answer.
Section 34.1 of the Act provides that, in civil proceedings where a defendant puts copyright or ownership in issue, the work is presumed to be protected and the named author is presumed to be the owner, unless the contrary is proved. That presumption is a litigation rule, not a general licence to assume protection exists. It helps a claimant who is already in court and has named themselves as author — it does not resolve, in advance, whether a wholly AI-generated work had an author capable of holding copyright in the first place.
The federal government has already tested this exact question against the public. Summarizing what it heard in a 2024 consultation on copyright and generative AI, ISED reported that “existing copyright jurisprudence in Canada suggests that authorship must be attributed to a human who exercises skill and judgment to create a work”, and that participants generally supported keeping human authorship central to protection (Consultation on Copyright in the Age of Generative Artificial Intelligence: What We Heard). No legislative change followed, which is itself informative: the skill-and-judgment standard this article describes is not a gap waiting to be filled, it is the standard Canada has confirmed it is keeping.
Treat material with no meaningful human authorship as unprotected until shown otherwise, and treat AI-assisted work with real human curation and editing as being on the same footing as any other original work. The practical consequence runs in both directions: a business cannot necessarily stop a competitor from copying a purely AI-generated asset it published, and a business also should not assume every AI-touched piece in its library is automatically defensible IP just because a person was involved somewhere in producing it. Where a specific asset genuinely matters — a core piece of branding, a flagship report — the safer path is to ensure a person is doing visible, documented creative work on top of whatever the AI tool produced first.
If the question of whether a specific asset is protected ever actually gets tested — a competitor copies it, or a buyer questions it during diligence — the evidence that matters is contemporaneous: drafts showing what was selected, revised or rejected; version history showing edits made after the AI tool’s first output; and a record of who made those editorial decisions and when. None of this is a formal filing requirement. It is simply the kind of evidence that turns the claim that a person did real creative work here from an assertion made after the fact into something that can actually be shown.
A marketing team generates a first-draft product photo with an AI image tool, then a designer spends several hours adjusting composition, correcting the product’s actual features, adding the brand’s specific colour treatment, and combining it with separately shot detail photography into a finished layout. That finished layout looks very different from a case where the same first draft was published unedited under a marketing deadline. The first is a strong candidate for ordinary copyright protection as a piece the designer authored, using AI as one tool among several; the second is much closer to the unresolved, unprotected end of the spectrum this article describes. Neither outcome depends on which tool the designer used — it depends on what the designer actually did with it, and how well that work is documented afterward.
It depends on how much human authorship went into the finished piece. Content that is essentially raw model output, with no meaningful human selection or editing, sits in the least protected position; content a person substantially shaped, edited and arranged is on much stronger ground.
No single amount of editing is a guaranteed threshold under Canadian law, since the Act does not set one for this scenario specifically. What is true generally is that the more real creative judgment a person exercises — not just light formatting — the closer the result sits to a traditional authored work.
Not automatically. Other rights can still apply — trademark, confidentiality, contractual restrictions in the tool's own terms of service — even where copyright itself may not attach, so “not copyrighted” is not the same as “free of every legal constraint,” and a competitor is not automatically free of every other obligation just because copyright does not reach the specific asset.
A vendor’s terms of service can grant the user whatever rights the vendor actually holds, and Section 13(4) of the Copyright Act still requires that grant to be in writing and signed by the owner — a clickwrap term can do that, but it cannot manufacture an author or a copyright interest that the underlying statute does not recognize as existing in the first place.
Related: what the Copyright Act says about authorship and who owns AI-generated work in Canada.
Whether a deliverable is protectable, and who holds what rights in it, is worth settling before a custom build starts.
