On April 26, 2026, the World Intellectual Property Organization (WIPO) marks World Intellectual Property Day, and the annual campaign is arriving in a different register. Patent offices, copyright registries, commercial courts and trade negotiators are being pressed to answer a question that used to be confined to policy seminars: when an algorithm produces a text, an image, a candidate compound or a piece of audio, who owns the right — and who gets paid.
World IP Day was established in 2000 and falls on the date the WIPO Convention entered into force in 1970. WIPO has 193 member states, and each April 26 governments, law firms and creator groups run events to explain how patents, copyrights, trademarks and designs operate. Past campaigns have focused on women inventors, small businesses, youth and the Sustainable Development Goals. This year the debate is not waiting for a unifying theme. It is already inside courtrooms, licensing negotiations, model terms of service and trade ministry briefings.
The question underneath every AI and IP dispute
Patent and copyright law presume a person at the centre. An inventor must be named, a creator must make choices, an owner must hold a transferable right, and a public record must explain the boundary. Generative AI systems — large language models and image or audio generators — interrupt that chain. They produce outputs from statistical patterns learned over enormous datasets, often with no human directing the exact expression. The legal machinery of originality and inventive step was built for minds, not models.
That is why the current fights are not only about licensing fees, although those are substantial. The deeper issue is whether the core concepts of intellectual property can handle a tool that generates candidate inventions and expressive works at speed and at scale, with no admitted subjective intent. WIPO counts 193 member states. Thaler's DABUS applications reached more than a dozen patent offices, and the base rate among major jurisdictions has been consistent: inventors must be human.
DABUS and machine inventorship
Stephen Thaler listed an AI system called DABUS as the inventor on patent applications in multiple countries, forcing the issue past theory. The United States Patent and Trademark Office refused, and the US Court of Appeals for the Federal Circuit ruled in Thaler v. Vidal in 2022 that the Patent Act requires inventors to be natural persons. The United Kingdom's Supreme Court reached the same conclusion in December 2023, holding squarely that an inventor must be a person and that DABUS itself cannot be one. The UK Supreme Court judgment in Thaler remains one of the clearest judicial statements for both documentarians and practitioners.
South Africa granted a patent listing DABUS as inventor in 2021, a result that drew wide coverage but did not reset global practice. Germany, Australia, the European Patent Office and the United States ultimately held to the human-inventor rule after early administrative signals suggested some uncertainty. WIPO's own process has been mapping the fault lines since 2019 through its multi-session Conversation on Intellectual Property and Artificial Intelligence; the WIPO AI and IP portal collects those discussions. No major jurisdiction has yet named an AI system as a legal inventor.
What remains open is the harder question Thaler's filings exposed. If a human makes a material contribution to an AI-assisted invention, patent protection may still be available. Patent attorneys now spend a large share of drafting time documenting exactly what the human changed, selected, verified or rejected. The line is not between AI and no AI; it is between material human contribution and mere operation.
Copyright and the shrinking circle of human authorship
Copyright has drawn an even brighter line. The US Copyright Office issued guidance on March 16, 2023, stating that it will register AI-generated material only where a human author has exercised sufficient creative control. A federal district court later affirmed the Office's refusal to register Thaler's AI-generated image A Recent Entrance to Paradise, calling human authorship a bedrock requirement. The US Copyright Office's AI guidance page now tracks the evolving policy.
The distinction sounds clean but collapses quickly in practice. A prompt alone may not create authorship; a photographer who uses an editing tool to remove a power line may still own the result. The Office has said protectable authorship depends on the degree of human selection, arrangement, modification and curation, not on whether AI was involved. For employers and rights holders that rely on creative assets, that means every AI-assisted work needs a short record of what the human changed and decided. Without that record, the copyright may be weaker than the team assumes.
Courts in China have split on this question. A 2023 Beijing decision recognised copyright in an AI-generated image because the user's detailed prompts showed sufficient original choices, while other Chinese courts have held that pure machine output has no author. These splits matter to exporters, publishers, collecting societies and platforms that distribute content across borders.
Training data: the fight that could settle more than ownership
Behind authorship sits the largest commercial question: whether copyrighted works used to train AI models require licences at all. Getty Images sued Stability AI in London in 2023 over millions of images it says were copied without permission; preliminary rulings allowed parts of the claim to proceed. The New York Times sued OpenAI and Microsoft in December 2023 over news content used for training and output generation, one of several publisher actions alongside authors and the Authors Guild. In music, Universal Music Group, Sony Music, Warner Records and other rightsholders sued AI music firms Suno and Udio in June 2024 over alleged mass ingestion of sound recordings. Reuters has documented the New York Times complaint and why it matters for the wider industry.
Not all plaintiffs want the same remedy. Some seek licensing fees; others want outputs restricted or training datasets deleted. The legal questions turn on whether training is fair use, fair dealing, a statutory exception or a separately licensed act. In the United States the analysis leans on fair use; in the United Kingdom and European Union the exceptions are narrower and collective licensing is more established. That is a key reason cases are clustered in London and New York rather than only in Silicon Valley.
Regulators are moving before courts finish
The European Union's AI Act, in force since August 1, 2024, imposes transparency duties on providers of general-purpose AI models. Providers must publish a sufficiently detailed summary of the content used to train the model, while protecting trade secrets. The AI Act tracker has become a standard reference for legal teams because the enforcement provisions phase in through 2027.
Japan runs a different policy, with a broad text and data mining exception that has been read as friendlier to model training. The United Kingdom has consulted on similar commercial text and data mining rules, but proposals have repeatedly stalled after creator groups objected. The United States has no federal AI-specific training-data law, leaving the issue to fair use litigation and sectoral state rules. In practice, a dataset that is defensible in Tokyo may create substantial liability in London, and a model released in Paris may have disclosure duties its New York competitor avoids.
These regulatory splits sit next to other AI policy fights. Export controls on advanced AI chips, for instance, shape which organisations can train the most capable models at all. For a wider account of that front, see this analysis of the US-China AI chip restrictions.
What this means for your lab, studio or legal team
The base rate for practitioners is straightforward: assume an AI system cannot be an inventor or sole author in the major markets that matter, but expect the line to move when a human makes meaningful creative or technical choices. Document those choices. A lab notebook entry that records the model proposed compound X, and notes which team member selected and validated it because of Y, is far more useful than a vague AI-assisted discovery notation. The same discipline applies to design studios, newsrooms, film teams and music labels that increasingly use generative tools.
For anyone publishing or selling AI output, the practical risk is not that a machine will claim ownership — the decisions so far block that. The risk is that ownership falls to no one, or to a service provider whose terms you accepted without reading. Review the terms of generative AI platforms before using them for commercial work. Some providers assign output to the user; others reserve rights, run content audits, cap commercial use, or promise only what local law permits, which may be very little.
- Keep a short record of human decisions in any AI-assisted patent filing or creative work.
- Check that training data licences cover the specific use you are making; a research licence rarely permits commercial training.
- Track the EU AI Act's disclosure obligations if your product is marketed in the EU, even when the model was trained elsewhere.
- For litigation or licensing, identify the dataset's origin, the model developer, the deployer and the end user; each jurisdiction attaches its own rules.
The next concrete step
No single court ruling or WIPO session will settle these debates in 2026. The useful next action for a creator, researcher, newsroom editor or legal team is to stop waiting for a final rule. Pick one AI-assisted project your organisation is currently shipping and run it under the strictest standard among the markets you serve: human-inventor or human-author documentation, training-data provenance, user-facing disclosure and commercial licence review. Running one project under that strict standard costs little now and produces the evidence you will need when the next ruling lands.
World Intellectual Property Day has always been about explaining why rights exist. In 2026 the harder job is explaining who, or what, actually holds them.
Photo by Google DeepMind on Unsplash