Today I spoke at Osgoode Professional Development’s 4th Annual Forum on AI and the Law. My topic was the legal aspects of intellectual property for AI-generated content and for training AI models. This post summarizes the main themes and the cases that are shaping the law including cases from Canada, the U.S., the U.K., Germany, China and India. An copy of my presentation can be accessed below.
| Key takeaways |
|
|
|
|
|
|
|
Does copyright subsist in AI-generated content?
In Thaler v Perlmutter (D.C. Cir. 2025), the court confirmed that “the Copyright Act requires all work to be authored in the first instance by a human being.” The court did not decide how much human input is enough when a work is created with AI assistance.
European courts are starting to answer that question. In February 2026, the Munich District Court held that AI outputs can be protected where creative prompting determines the output enough that it reflects the prompter’s personality. In April 2026, the Düsseldorf Higher Regional Court reached a similar conclusion. Both courts held that simply choosing one image from several AI suggestions is not enough.
China is taking a similar approach to prompts. In a case of first impression (Shanghai, November 2025), the court held that six sets of Midjourney prompts were unprotectable ideas and “stock” elements without originality. It also expressed a policy concern about the “excessive privatization of linguistic resources.” In Feng v Zhangjiagang (2025), prompts for AI-generated chair designs were found not to be copyrightable. In Canada, the Federal Court challenge in CIPPIC v Sahni to a registration naming an AI as co-author is the case to watch.
Patents follow the same principle. In Thaler (Re), 2025 CACP 8, the Canadian Patent Appeal Board and Commissioner confirmed that an “inventor” under the Patent Act must be a natural person, so DABUS cannot be named.
When do AI outputs infringe?
Output-based claims are getting past motions to dismiss:
- Disney v MiniMax (C.D. Cal. 2026): prompting with character names allegedly produced close reproductions of characters such as R2-D2 and C-3PO. Both the training and the output infringement theories were found plausible.
- Alcon v Tesla (C.D. Cal. 2026): a literal-copying claim about an AI image resembling a Blade Runner 2049 still was allowed to proceed.
- Advance Local Media v Cohere (S.D.N.Y. 2025): the publishers’ “substitutive summaries” theory survived, supported by 75 examples of alleged verbatim copying and close paraphrasing.
- In re OpenAI (S.D.N.Y. 2025): ChatGPT summaries of books were found to be plausible “abridgments” of protected setting, plot and characters.
- China: defendants were criminally convicted for using a “picture-to-picture” AI tool to make only minor changes to others’ artwork for commercial jigsaw puzzles.
Training AI models: fair use, fair dealing
This is where courts diverge the most.
- United States: In Bartz v Anthropic (2025), training was held “quintessentially transformative” fair use, but the use of pirated library copies was not. A class settlement followed. Kadrey v Meta (2025) also found fair use, but suggested a “market dilution” theory of harm. In September 2026, the U.S. government filed a Statement of Interest in In re OpenAI calling LLM training “extraordinarily transformative” and calling Kadrey’s dilution theory “deeply flawed.” In contrast, the Third Circuit in Thomson Reuters v ROSS (September 29, 2026) held that copying Westlaw headnotes to train a competing non-generative legal research tool was not fair use. The court expressly distinguished generative AI.
- United Kingdom: In Getty Images v Stability AI [2025] EWHC 2863 (Ch), the court held that a trained diffusion model is not an “infringing copy,” because its weights do not store the training images.
- India: In ANI v OpenAI, the Delhi High Court took the prima facie view that training falls within the “private or personal use, including research” fair dealing exception and the training was a fair use.
- Germany: In GEMA v OpenAI (Munich, November 2025), memorized song lyrics in a model were held to be reproductions. In GEMA v Suno (Munich, July 2026), the court held that memorization during training falls outside the text and data mining exception. It also concluded the use would not be fair use under U.S. law.
- Canada: In Toronto Star v OpenAI, 2025 ONSC 6217, the court took jurisdiction over the publishers’ claims but decided none of the substantive issues. OpenAI is appealing.
AI and trademark infringement
In Getty v Stability AI the court held that synthetic Getty watermarks in outputs from certain model versions infringed under sections 10(1) and 10(2)(b) of the U.K. Trade Marks Act, and treated those outputs as Stability’s own commercial communications. In Cohere, Lanham Act claims over hallucinated articles bearing publishers’ marks also survived a motion to dismiss.
How using generative AI can cost you your rights
- Privilege and confidentiality: In United States v Heppner (S.D.N.Y. 2026), a defendant’s exchanges with a public AI platform were held not to be privileged or work product. Morgan v V2X (D. Colo. 2026) and Assini v Hayward (N.Y. 2026) went the other way on work product for litigants’ own AI use. In Munir (UKUT), the U.K. tribunal warned that putting client documents into non-enterprise versions of AI tools breaches confidentiality and waives privilege. In Fry, the court refused to fully seal documents that had been uploaded to a consumer AI tool.
- Trade secrets: In Trinidad v OpenAI (N.D. Cal. 2026), the court held that “protocols and frameworks” developed in ChatGPT were not trade secrets, because the plaintiff had disclosed them under the platform’s terms.
- Governance and contracts: In Fortis Advisors v Krafton (Del. Ch. 2026), a CEO who followed ChatGPT’s strategy over his legal advisors’ advice was found in breach of contract, and he admitted deleting the chat logs. ASIC v Bekier (FCA 2026) commented on company directors’ use of AI to review board materials.
Who is responsible when AI agents access websites?
In Amazon v Perplexity (9th Cir. 2026), the court held that it is the user, not the AI agent’s provider, who “accesses” a website under the Computer Fraud and Abuse Act (CFAA). This is an early and important ruling on responsibility for agentic AI.
Practical diligence on AI vendor terms
Reviewing the terms of popular generative AI platforms shows consistent patterns:
- Some vendors use customer inputs and outputs for training.
- Vendors give no warranties on outputs.
- Most require the customer to indemnify them.
- Vendors rarely disclose their training data, even though infringement risk tracks the training data.
Even where a vendor assigns rights in outputs, purely AI-generated content may carry no enforceable rights. To build protectable assets, add human-authored content, modifications, or creative selection and arrangement. Platform terms also change regularly.
The bottom line
Courts around the world are moving quickly but not in lockstep. Organizations should:
- treat purely AI-generated content as unprotected
- assess output infringement risk
- keep confidential and privileged material out of public AI tools
- read vendor terms closely before relying on generative AI.
Frequently asked questions
Can AI-generated content be protected by copyright?
Not if it is generated entirely by AI. Courts in the U.S. (Thaler v Perlmutter), Germany and China require human creative contribution. Works created with AI may be protected where there is sufficient human authorship.
Are AI prompts protected by copyright?
Rarely, but it depends. In 2025, a Shanghai court held that short, keyword-style Midjourney prompts were unprotectable ideas. German and Chinese courts accept that detailed creative prompting can make an output protectable.
Is training AI on copyrighted works fair use?
It is unsettled. Bartz and Kadrey found fair use for generative AI training, and the U.S. government supports that view. An court in India found fair use for generative AI training. ROSS rejected fair use for a non-generative tool, and a German court rejected a fair use defense where outputs substantially copied musical works training tata.
Can using ChatGPT or Claude waive privilege?
It can. Heppner held that a defendant’s use of a public AI platform was not privileged. Other U.S. courts have protected litigants’ AI-assisted work product. Use enterprise tools for confidential material.
Who is liable when an AI agent accesses a website?
Under Amazon v Perplexity (9th Cir. 2026), the user, not the AI provider, accesses the site for CFAA purposes.
A copy of my paper can be accessed below.
