Almost everyone who asks this question asks about the wrong statute. And the people who have actually extracted money from AI developers have not won anything resembling a claim over identity. They won, or settled, on ordinary copyright in ordinary files. Anthropic agreed to pay US$1.5 billion — roughly US$3,000 a book — and the finding that produced that number had nothing to do with what its model could imitate. It had to do with where the books came from.
What copyright actually covers
Under the Copyright Act, copyright subsists in original literary, dramatic, musical and artistic works that have been fixed in some material form. Three words in that sentence do all the work. Original — it must be yours and not copied. Fixed — it must exist as a file, a page, a recording. Work — a defined thing with edges.
A persona has none of those edges. Your voice is not a work; a recording of your voice is. Your visual identity is not a work; the photograph is. Your way of explaining things is not a work; the fourteen essays where you explained it are. Copyright has never protected style, method or identity, and the arrival of a machine that can absorb all three has not changed the statute. In March 2026 the U.S. Supreme Court declined to hear Stephen Thaler's appeal against the Copyright Office, leaving standing the rule that a work needs a human author — a reminder that copyright is a system about authorship, not about resemblance.
What you already own, without filing anything
Copyright arises the moment the work is fixed. No registration, no notice, no fee. Registration with the Canadian Intellectual Property Office is cheap and is worth doing for the pieces you would actually sue over, because the certificate raises a presumption of ownership that saves you an evidentiary fight later. But your back catalogue is already protected.
The problem is usually ownership, not subsistence. Three traps catch people who thought they owned their own material:
- Work created by an employee in the course of employment belongs to the employer by default. If a company you once worked for paid for the writing, the company owns it.
- Work created by a contractor — your videographer, your ghostwriter, your podcast editor — belongs to the contractor unless your agreement says otherwise. Since 2012 this includes commissioned photographs. The photographer owns the shoot.
- Platform terms routinely take a broad licence, and increasingly an express training licence, over everything you post. You keep the copyright and give away the exclusivity that made it worth having.
Fix these before you need them, not after. An assignment signed today costs a paragraph; an assignment you are trying to obtain from a former contractor after you have found your work in a model output costs whatever they decide to charge.
How an infringement claim is actually built
Every serious claim against an AI developer has been assembled from the same five parts. If you want the option, build them now.
- An inventory. Titles, first-publication dates, URLs, file hashes. Canadian statutory damages are elected per work, so the schedule of works is the claim. The Canadian news publishers suing OpenAI pleaded up to $20,000 for each.
- A clean ownership chain. Assignments from every contractor, confirmation for anything created during employment, and CIPO registration for the flagship pieces.
- Evidence of copying. This is where most claims die. Getty had 12 million images in the training set and still lost in London, because it could not show reproduction in the UK. What survives are outputs: verbatim regurgitation, a reproduced watermark, a passage that could only have come from your file.
- Technological protection measures. Circumventing a paywall, a login or a rights-management measure is its own cause of action under the Act, independent of whether the copying was fair.
- Terms of use. A scraper that accepted your terms is in contract, and contract does not care about fair dealing. This is the quietest and most underrated of the five.
Which right reaches which problem
| If the problem is… | The right that reaches it | What it will not do |
|---|---|---|
| A model was trained on your articles, photos or videos | Copyright in those works; TPM circumvention; breach of your terms of use | Nothing stops the model from having already read you |
| Your face or voice is used to sell something | Appropriation of personality; BC Privacy Act s. 3; Quebec Civil Code art. 36 | Requires commercial use and that you are identifiable |
| Your name or catchphrase identifies your business | Registered or common-law trade-mark; passing off | Only reaches confusing commercial use |
| Your work is distorted or falsely associated with a product | Moral rights — integrity and association — which cannot be assigned, only waived | Gone if you signed a standard waiver |
| You want to stop training on your material going forward | Contract: licence terms, platform terms, engagement terms | Binds only the party that agreed |
What the recent cases decided
Judge Bibas got there first. In February 2025 he held that Ross Intelligence's use of Westlaw headnotes to train a legal research tool was not fair use, largely because Ross was building a competing product and because there is a market for training data. The Third Circuit heard the appeal in June 2026 and has not ruled.
Four months later Judge Alsup split the question in two. Training Claude on books, he held, was transformative and fair. Downloading and keeping a pirate library of those same books was not. The second half of that sentence is what produced the US$1.5 billion settlement. Two days later Judge Chhabria reached a fair-use result for Meta on the record in front of him, and took a different view of the downloading. Both judges were careful to say they were deciding the cases before them.
In London, Mrs Justice Joanna Smith delivered 759 paragraphs and gave Getty almost nothing. Getty abandoned its training and output claims mid-trial, because it could not establish that the copying happened in the United Kingdom, and its remaining secondary infringement claim failed on a finding that will echo: the model weights contain no copies of the works. What Getty won was a narrow, historic trade-mark point about watermarks appearing in generated images — which is to say it won on branding, not on copyright.
The Canadian case has not reached any of this. On 7 November 2025 the Ontario Superior Court dismissed OpenAI's jurisdictional challenge and held that Toronto Star Newspapers Ltd. v. OpenAI Inc., brought by CBC/Radio-Canada, Torstar, Postmedia, The Globe and Mail and The Canadian Press, can proceed in Ontario. The merits are untouched. But the venue matters more than it sounds, because Canadian law is not American law on the one question that decides these cases.
Why Canada is a different problem
The United States has fair use: an open-ended, four-factor balancing test that a court can stretch to cover a use nobody contemplated in 1976. Canada has fair dealing, which is a closed list. The dealing must first fall within an enumerated purpose — research, private study, education, parody, satire, criticism, review or news reporting — and only then is it tested for fairness. Commercial model training is not obviously any of those, and Parliament has not added a text-and-data-mining exception. The government consulted on exactly that question in 2023 and has not legislated.
So Alsup's reasoning does not travel. A Canadian court asked whether training on scraped Canadian journalism is fair dealing has to start by finding a listed purpose, and "building a commercial product" is not on the list. That is a meaningfully better starting position for a Canadian rights holder than American headlines suggest.
And the persona itself
Outside copyright, the tort of appropriation of personality has been available in Canada for fifty years. Bob Krouse, a Tiger-Cats defensive back, lost in 1973 because the photograph on Chrysler's football spotter showed him from behind in a crowd; the Court of Appeal recognised the tort and then held he was not the subject of it. George Athans, a water-skier, won four years later because a line drawing of him in his signature pose was identifiable and sat in a summer camp's brochure. In 1986 an amateur bodybuilder photographed from neck to waist lost in British Columbia for the same reason Krouse did.
Two elements, and both are doing real work against synthetic media. The use must be commercial, and you must be clearly and primarily captured. A convincing deepfake of you endorsing a product satisfies both far more easily than Chrysler's spotter ever did. Glenn Gould's estate marks the limit: in 1998 the Ontario Court of Appeal let a book about Gould stand, because biography and commentary about a person of public interest are not appropriation. British Columbia and Quebec add statutory versions of the same right, and the United States is close to a federal one — the NO FAKES Act cleared the Senate Judiciary Committee unanimously in June 2026 and is still a bill.
What to do this quarter
Inventory the works you own and record their publication dates. Register the handful you would actually sue over. Get assignments from every photographer, editor and contractor who has ever touched your material, and check what you signed about moral rights. Put a training restriction in your own licensing and terms of use and your client engagement terms, so that anyone who deals with you is in contract rather than merely in the vicinity. Preserve outputs when you find them — screenshots, prompts, dates — because the evidentiary gap is what sank Getty. If you have any of this in place and want to know which pieces are load-bearing, that is a short conversation — and if the dispute is already live, it may not need a courtroom.
None of it will stop a model from having read you. That question was decided before anyone thought to ask it.
Common questions
Can I copyright my face, voice or personality?
No. Copyright protects original works fixed in a material form — a photograph, a recording, an article. Identity is not a work. The rights that reach identity are appropriation of personality, the BC Privacy Act, Quebec's Civil Code, and trade-mark law.
Is there any way to stop an AI model from training on my content?
Not by registration. You can restrict it by contract — your own terms of use, licence terms and engagement terms — and you can assert copyright against a developer that copied your works to get there. Neither prevents a model that has already been trained from having absorbed your material.
Do I have to register copyright in Canada to sue?
No. Copyright arises automatically on fixation. Registration with CIPO is optional, inexpensive, and worth doing for the works you would actually litigate, because the certificate creates a presumption of ownership and removes an argument the defendant would otherwise get to make.
How much are statutory damages for copyright infringement in Canada?
For commercial infringement, section 38.1 of the Copyright Act allows $500 to $20,000 per work, elected instead of proving actual damages and profits. For non-commercial infringement the range is $100 to $5,000 for all works together. Courts can go below the minimum where the total would be grossly disproportionate.
Does Canada allow AI training as fair dealing?
Unsettled, and structurally harder than in the United States. Fair dealing requires the use to fall within an enumerated purpose — research, private study, education, parody, satire, criticism, review or news reporting — before fairness is even assessed. Canada has no text-and-data-mining exception. The Ontario case against OpenAI will be the first real test.
Who owns the photographs and videos on my website?
Usually the photographer or videographer, unless your contract says otherwise. Commissioning and paying for a shoot does not transfer copyright. Employees are the exception — work created in the course of employment belongs to the employer by default.
What did the Getty v. Stability AI decision actually decide?
Less than the headlines suggested. Getty abandoned its training and output copyright claims during trial over where the copying took place, and the remaining secondary infringement claim failed because the court found the model weights contain no copies of the works. Getty won only a narrow, historic trade-mark finding about watermarks in generated images.
Can my estate stop AI from using my likeness after I die?
Possibly. Canadian courts have suggested personality rights are descendible property, though the point is not settled. In the United States the NO FAKES Act would expressly pass the right to heirs for up to seventy years; it advanced out of Senate Judiciary in June 2026 but is not law.