Anthropic Is Now Fighting Copyright Wars on Three Continents. Canada Just Opened a Fourth Front.

Postmedia et al. v. Anthropic, Ontario Superior Court, Filed 6 October 2026

The facts are becoming familiar, but the pattern is not yet settled law. A coalition of Canada's largest news publishers - Postmedia, Torstar, The Globe and Mail, and The Canadian Press - filed suit against Anthropic and its Canadian subsidiary in the Ontario Superior Court on 6 October 2026. The action, first reported by CARTT.ca and confirmed nationally by The Canadian Press on 19 October, alleges that Anthropic used the publishers' copyrighted journalism, without licence, to develop and continuously improve its Claude large language model.

Strip away the boilerplate and what you have is this: the publishers claim Anthropic deployed automated web crawlers, including one identified in the claim as "ClaudeBot," to scrape their websites, circumvent paywalls, and collect content for commercial training purposes in breach of the sites' contractual terms of service. The Canadian Press alone alleges 4.2 million works were scraped. Postmedia and PNI Maritimes together put their figure at 1.75 million. The publishers are seeking injunctive relief under the Copyright Act to prevent Anthropic from accessing, scraping, or using their current and future works without written consent.

What Was Actually Decided, and What Has Not Been

Nothing has been decided yet. This is a filed claim, not a judgment. The Ontario court has not ruled on jurisdiction, liability, or damages. Anthropic has not responded publicly to the action. What the claim does is crystallise a legal theory that has been circulating in US litigation for three years: that training-data ingestion is ongoing reproduction, not a one-time historical act, and that each new Claude version trained on scraped content restarts the infringement clock.

That theory matters because it defeats one of the AI industry's most reliable early-stage defences: that the training happened years ago, in a different jurisdiction, under a different regulatory environment. If the publishers are right that Anthropic "scrapes content on an ongoing basis because of ongoing user prompts," the infringement is not frozen in 2021 or 2022. It is happening now.

Canadian copyright law also matters here. The Copyright Act (RSC 1985, c. C-42) has no fair use doctrine. It has fair dealing: a narrower, enumerated set of permitted purposes. Research, private study, education, parody, satire, criticism, review, news reporting. Generative AI training does not map cleanly onto any of them. The AI industry cannot simply import its US fair-use playbook across the 49th parallel and expect it to survive a Canadian court.

The Anthropic Problem in Full

This Ontario action is the latest in a documented pattern. In January 2026, music publishers filed a lawsuit seeking damages "in excess of $3 billion" over what they characterised as "flagrant piracy" of more than 20,000 works, a figure they said they discovered through the discovery process in the earlier Bartz v. Anthropic class action. In September 2025, Anthropic agreed to pay authors $1.5 billion to settle Bartz, in what would rank as the largest copyright resolution in US history, pending final court approval. A separate music publishers' action, naming both Anthropic CEO Dario Amodei and co-founder Benjamin Mann as defendants, is actively proceeding.

The Broadcast Dialogue report on the Ontario filing made the point that distinguishes Anthropic from almost every competitor: "in contrast to nearly every major artificial intelligence company, including OpenAI, Anthropic has yet to execute any form of public content licensing agreement with major publishers." That is not a minor commercial oversight. It is a strategic posture, and increasingly an expensive one.

For comparison: News Corp's deal with OpenAI is reported at up to $250 million over five years. Its deal with Meta, signed in March 2026, is reported at up to $50 million per year for three years. Brazil's Grupo Folha converted its 2025 copyright suit into a licensing agreement in May 2026. The licensing market is functioning. Anthropic is simply not participating in it.

The UK and EU Context: Why Jurisdiction Still Determines Everything

The Ontario filing arrives in a legal environment where training-data liability remains unresolved in every common-law jurisdiction. In the UK, Getty Images v. Stability AI produced a November 2025 High Court judgment dismissing Getty's secondary copyright claim on the basis that AI model weights imported into the UK were not "infringing copies" within the meaning of the Copyright Designs and Patents Act 1988. Getty has been granted permission to appeal, and the Court of Appeal's review of how "infringing copy" applies to an AI model weight will be one of the most important copyright hearings of 2027.

The core question, whether training an AI model on copyright-protected works in the UK constitutes infringement under the CDPA, was never answered. Getty dropped its primary claim after being unable to establish that Stability AI's training took place on UK soil. That geographical escape hatch is precisely what the Canadian publishers appear to be targeting: Anthropic and its Canadian subsidiary are squarely within Ontario's jurisdiction.

The EU has moved further. The EU AI Act's training-data disclosure obligations entered full enforcement in August 2026. The DSM Directive's text-and-data mining framework, with its Article 4(3) opt-out mechanism, has no direct UK equivalent. British developers selling into the EU market may face obligations that domestic statute does not impose. That asymmetry shows how far the statute is from being drafted for this technology.

The Trademark Signal That Sits Underneath All of This

There is a quieter data point worth logging. AIBD's analysis of UK Intellectual Property Office trademark data shows only 3,094 Class 42 filings in 2026 Q4, a drop of 62.2% against the prior period. Class 42, which covers scientific and technological services and, since 1 January 2026, AI-as-a-Service under the WIPO Nice Classification 13th Edition, is the class in which every AI company filing brand protection for its software services should be active. A collapse of that magnitude in a quarter where AI litigation is expanding globally suggests either the litigation chill is real and companies are deferring brand investment, or the market has overshot and a correction was due. Neither reading is reassuring.

NCL 13-2026 explicitly added "artificial intelligence as a service (AIaaS)" to Class 42 for the first time, alongside AI consulting services and AI research and development. Any AI business that filed its Class 42 marks under legacy specifications before 1 January 2026 should be checking coverage against the new terminology. An overly narrow specification does not automatically invalidate a trademark, but it weakens the mark in non-use cancellation proceedings: precisely the kind of proceedings that well-resourced defendants initiate to neutralise a claimant's IP portfolio while litigation is pending.

The Practical Impact: Who Changes Behaviour, and How

For AI developers, the Ontario filing reinforces what the Bartz settlement should already have established: the training-data problem is not going away through attrition. Each new jurisdiction that proceeds to trial without a licensing framework in place adds both litigation risk and transaction costs. The publishers bringing this Ontario action also have an active suit against OpenAI from 2024, which has not settled. They are not reluctant litigants.

For publishers, the strategic split is now documented and stable. Some sign: News Corp, the Financial Times, the Associated Press. Some sue: the New York Times, the Danish DPCMO, the Canadian coalition. Some do both: the Times is a litigant against OpenAI and a licensee of Amazon. The split is not random. It maps onto each publisher's assessment of whether licensing revenue at current prices compensates for the traffic destruction that AI-answer products are already causing.

For counsel advising AI companies, the structural lesson from the Ontario filing is precise: terms of service violations as the predicate for bypassing technological protections are the cleanest possible theory for a plaintiff. No fair use. No fair dealing. Just breach of contract plus copyright infringement, both arising from the same act of automated scraping.

Next Steps

Anthropic must file a response in the Ontario Superior Court. Expect a jurisdictional challenge as an opening move. If jurisdiction is confirmed, the discovery phase will be the critical battleground: the publishers will seek internal documentation of ClaudeBot's scraping parameters and the training datasets used for each Claude version. That is the same discovery that unlocked the $3 billion music publishers' claim in the US, where a shadow library was found inside Anthropic's training corpus that the company had not initially disclosed.

Watch for a jurisdiction ruling in Q1 2027. If the court confirms it will hear the case, a licensing negotiation becomes Anthropic's lowest-cost exit. At current market rates, that conversation starts at eight figures annually.