Canada Caps ISBNs for AI Books — but Won't Define 'Mostly'

By Daniel Rao | October 24, 2026

Canada has just asked every publisher in the country to sign a small declaration of humanity. From this month, anyone registering for an ISBN through Library and Archives Canada must declare that a human author wrote the publication. Brainstorming with a chatbot, running the manuscript through a grammar check, letting software handle the formatting — all fine. But if the text, illustrations or translation in your book were "primarily generated" by AI, the national library will limit how many ISBNs you get.

And here is the part Ottawa refuses to spell out: what counts as "primarily", what the limit actually is, and who does the checking. The policy names no quota, describes no test, and concedes it is still under review. Canada has put up a velvet rope against the machines — and hired no bouncer.

Photo: Library and Archives Canada's Preservation Centre in Gatineau, Quebec. The institution that guards the country's documentary heritage now rations book identifiers for AI-written works. Image: An MP / Wikimedia Commons, CC BY-SA.

The new rule, in LAC's own words

The wording sits on LAC's official "About ISBNs" page on canada.ca, updated October 9, and it is admirably blunt:

"By registering with ISBN Canada, you declare that your publication was written by a human author. You may use AI tools only for limited support tasks, such as brainstorming ideas, checking grammar, or formatting your book. ISBN Canada limits the quantity of ISBNs for works whose content is primarily generated by AI (text, illustration, translation). LAC is currently reviewing its policy regarding the use of AI content creation and our guidance is subject to change."

An ISBN is the 13-digit number that identifies each specific edition of a book — the thing that lets retailers, libraries and distributors tell one edition from another anywhere in the world. In Canada, unlike the United States, they are free: eligible publishers request them from LAC at no charge for print books, ebooks, audiobooks and educational DVDs, usually in blocks of ten or a hundred. (Americans pay Bowker $125 for a single ISBN or $295 for ten, which tells you something about how seriously Canada takes the public infrastructure of publishing.) LAC is careful to add that an ISBN confers no copyright protection whatsoever. It is an identifier, not a shield.

The guidance did not arrive fully formed. It first surfaced around October 7, and SaskBooks — the publishing organisation for Saskatchewan — reported on October 19 that the AI wording had already changed several times in less than two weeks, advising its members to "consult the current official version before submitting a title." That is the weary tone of people who have watched the rules move under their feet.

A declaration with no detector

Read the policy twice and the hole in the middle gets bigger. LAC publishes no numerical quota for AI-generated works. It describes no verification process, no disclosure form, no audit standard for sniffing out machine prose. Its general ISBN terms do give it teeth in principle — it may limit the number or frequency of ISBNs, demand justification for large or repeated requests, audit usage, and deny, reduce or suspend access where it smells misuse — but none of that machinery is presented as AI-specific. It is the ordinary plumbing of the ISBN office, repurposed for an extraordinary question.

We checked LAC's live "About ISBNs" page this morning: the AI section is still there, still carrying the October 9 date stamp, and still ending with the same elegant shrug — "our guidance is subject to change." Nobody at the agency has said when the review will conclude, or what it is even reviewing toward. For now, the entire enforcement mechanism is a checkbox and a conscience.

There is something almost touching about the bureaucracy of it. The same page that explains, with diagrams, how the five hyphen-separated parts of an ISBN work and where to get your barcodes printed is now also the front line of the culture war over machine authorship. Dry civil servants, deciding what counts as human.

'Only a first step': the authors cheer, warily

The country's author and publisher groups welcomed the move — and immediately asked for more. The Canadian Authors Association said it "applauds" the change, telling Global News: "While we applaud this change, we recognize that this is only a first step." Then came the question LAC has not answered: "We await further information to see how the self declaration of human authorship is monitored and verified by Library and Archives Canada, as well as which penalties may be delivered should the self declaration be proven to be false and the book written by AI."

In the same statement, the association framed the stakes in civilisational terms: "This change not only protects Canadian authors and our valuable creative sector, it also allows Canada's cultural and heritage identity to continue to be developed and shaped by human authors, and not by artificial intelligence or machines."

The Association of Canadian Publishers was blunter about the enemy. "The ISBN policy requiring a declaration of human authorship is aligned with our values. Our members publish works that are the products of human creativity, scholarship, artistry, and craft. We do not support the use of AI to replace or substitute for human creative work," it said in a statement on Thursday. "Over the last few years, publishers and authors have faced a flood of low-quality, unverifiable, copycat products that are AI-generated and uploaded for sale online. AI-slop books crowd out legitimate titles and create consumer confusion, with real risks to the public from disinformation." Its hope: "by limiting ISBNs to human-authored books, this new policy will lead to less fraud in the market."

None of this comes out of nowhere. The association has spent two years arguing, on Parliament Hill and elsewhere, that "the unlicensed use of creative works for training generative AI is a major, unjust threat to the livelihoods of the people behind those works, and must not be permitted." The ISBN rule is the first time a federal institution has actually acted on that argument — with a ration book.

The lawsuit hanging over all of this

The ISBN policy lands in the middle of the ugliest copyright fight in Canadian media history. On October 6, four of the country's biggest news publishers — Postmedia, Torstar, The Globe and Mail and The Canadian Press — filed suit against Anthropic in an Ontario court, alleging the company scraped their journalism to train and operate its Claude models without a licence. The filing was first reported by the industry site Cartt.ca and then by The Canadian Press on October 19.

The statement of claim pulls no punches: "Anthropic has capitalized on the commercial success of its AI models, building an expansive suite of services, and raised significant capital — all without obtaining a valid licence from any of the News Media Companies." Beyond training-data copying, the publishers allege Anthropic broke their sites' terms of service, circumvented technical protections like paywalls, and keeps scraping their content on an ongoing basis in response to user prompts. They are seeking damages, an accounting of profits or, alternatively, statutory damages of up to $20,000 per infringed work, plus punitive damages and injunctions. Anthropic had not responded to requests for comment at the time of reporting, and none of the allegations has been proven in court. Notably, the same publishers — plus CBC/Radio-Canada — sued OpenAI on similar grounds in 2024; that case still has not been settled.

Against that backdrop, Ottawa's legislative silence is deafening. Innovation, Science and Economic Development Canada says the federal consultation on AI and copyright has closed and the government is weighing whether concerns should be addressed through copyright policy or other measures — which is to say, no decision. The country's new federal AI strategy, released earlier this year, said nothing about unlicensed AI training or replication. So the national library is rationing identifiers for AI books while the government has yet to decide whether training AI on those same books, without permission or payment, is even legal. The left hand does not know what the right hand is licensing.

What changes for working publishers

For legitimate Canadian publishers, the practical effect is a new checkbox and a new risk calculus. If your workflow is what most publishers actually do — draft with humans, brainstorm with AI, run Grammarly, format in InDesign — nothing changes; LAC explicitly permits all of that. The rationing bites only the operations churning out volumes of primarily AI-generated text, illustrations or translations, the "AI-slop" mills the publishers' association describes. Those operators can still publish — nothing in the policy bans AI books — but they cannot lean on Canada's free, unlimited identifier supply to do it at scale.

The rest of the industry is quietly building its own fences. CANSCAIP, the Canadian children's writers' organisation, now states flatly in its competition guidelines: "AI-generated entries are not eligible." BookNet Canada published a full survey in September on AI use across the North American book industry. The direction of travel is unmistakable: human authorship is becoming a credential you have to prove, like organic certification for literature.

It is worth noting, for American readers watching from across the border, how unusual Canada's setup is. The free, state-issued ISBN is public infrastructure most US authors never had — and it is precisely because the numbers are free that rationing them is a lever Ottawa can pull. In the US, where Bowker sells identifiers by the block, no equivalent choke point exists. Canada may be the only country that can fight AI slop with stationery.

Update, Oct. 24: publishers applaud, Ottawa stays silent on the copyright question

Update, Oct. 24: the industry verdict is in — and it is louder than LAC’s policy. The Canadian Authors Association said it was “proud to have played a role through their advocacy” in the change, while calling it “only a first step” and demanding to know “how the self declaration of human authorship is monitored and verified... as well as which penalties may be delivered should the self declaration be proven to be false.” The Association of Canadian Publishers called the policy “aligned with our values” — “We do not support the use of AI to replace or substitute for human creative work” — blaming “a flood of low-quality, unverifiable, copycat products” and hoping the rule “will lead to less fraud in the market.” Meanwhile the enforcement question is getting harder, not easier: Canadian publisher Boreal has just dropped writer Thelyson Orelien over suspected AI use in a novel, after document metadata flagged it — while conceding it “cannot conclusively determine” how much AI was involved. And on the bigger copyright question, EconoTimes reports that Ottawa’s federal AI-copyright consultation has closed, with the government still deciding whether to act through copyright policy “or other measures.”

The same war, fought in Canberra

Canada is not legislating in a vacuum. Just this week, Australia told AI giants they cannot train on the country's culture for free, killing a proposed text-and-data-mining exception after a revolt by creators — the harder, legislative version of what LAC is attempting administratively. The copyright opt-out brawl even dominated Australia's parliamentary AI hearings, where OpenAI and Anthropic executives were grilled on whose work their models learned from. The pattern is now unmistakable across the Anglosphere: creators are done asking nicely. And in a curious coda, LAC itself keeps ending up in AI headlines — AI agents probed the institution's own collection-search system earlier this month, in what researchers described as a failed hacking attempt.

Our take: directionally right, operationally shoddy

Let us be clear about what LAC got right. Somebody in the federal government finally did something concrete for human authors instead of commissioning another consultation, and the publishers cheering it are not wrong that AI-generated copycat books are a genuine fraud and disinformation problem. A declaration of human authorship, attached to the identifier every legitimate book needs, is a clever pressure point — it costs honest publishers nothing and makes industrial-scale slop marginally harder.

But the execution is shoddy in ways that matter. A limit with no published number is not a policy; it is a mood. A declaration with no verification is not enforcement; it is an honour system for an industry that includes bad actors by the thousand. And "primarily generated by AI" is doing colossally heavy lifting for a phrase nobody has defined — is a translated novel "primarily" AI-generated if a human edited every line? What about an illustrated children's book with AI art and human text? LAC will have to answer these the first time a publisher appeals a rationing decision, and answering them in the middle of a dispute is the worst possible way to write rules.

The deeper problem is the mismatch this article keeps circling back to: Canada will ration the identifiers on the output side while leaving the input side — the unlicensed ingestion of Canadian books to train the very models producing the slop — entirely unaddressed. Until Ottawa decides whether training on copyrighted work requires permission, the ISBN rule is a sandbag against a flood the government has not yet admitted exists. The authors called it "only a first step." They were being generous. It is a first step with no second step scheduled.

Quick answers

Does Canada ban AI-written books?

No. Nothing in LAC's guidance prohibits publishing AI-generated work. It rations free ISBNs for works whose content is primarily AI-generated and requires a human-authorship declaration at registration. You can still publish the book; you just may not get unlimited free identifiers for it.

What AI uses are still allowed?

Brainstorming ideas, checking grammar, and formatting — described as "limited support tasks." LAC does not treat those uses as making a work AI-generated.

What is the actual limit on ISBNs for AI books?

Unknown. LAC publishes no numerical quota and describes no AI-specific verification process. Its general terms allow it to limit numbers, demand justification, audit use, and deny or suspend access for misuse.

Is the policy final?

No. LAC says it "is currently reviewing its policy regarding the use of AI content creation and our guidance is subject to change," with no announced end date for the review. The wording has already changed several times since it first appeared around October 7.

Sources

Library and Archives Canada, "About ISBNs" (canada.ca, official guidance page, updated Oct. 9, 2026 — checked live Oct. 24); Global News, "Library and Archives Canada cracking down on AI books amid copyright threats" (Ariel Rabinovitch, Oct. 22); EconoTimes, "Canada Limits ISBN Allocations for Books Made Mainly With AI" (Oct. 23); SaskBooks via industry reporting (Oct. 19); The Canadian Press / Cartt.ca / MobileSyrup on the Postmedia–Torstar–Globe–CP suit against Anthropic (filed Oct. 6, reported Oct. 19); The Fussy Librarian (Nate Hoffelder flag, Oct. 8); BookNet Canada, "AI use across the North American book industry 2026"; CANSCAIP submission guidelines; Canadian Authors Association advocacy page; Foglio on Canadian self-publishing and ISBN costs.

Daniel Rao covers AI policy and the business of technology for TechPulse USA. He has reported on copyright fights from three continents and still underlines sentences in paper books.