The editorial argues Japan's Supreme Court correctly tracked the statute — the Patent Act presupposes a natural person who can hold and transfer rights. Courts across four continents have converged on the view that judicial reinterpretation cannot stretch 'inventor' to cover AI systems; only the Diet (or equivalent legislatures) can make that change.
With the USPTO, UK Supreme Court, EPO Boards of Appeal, and now Japan's Supreme Court all rejecting DABUS, the editorial frames the question as globally resolved. South Africa is dismissed as an administrative outlier rather than a substantive counterexample, meaning inventors must be human across every major patent jurisdiction.
By surfacing the Yomiuri report on HN with strong engagement (192 points, 80 comments), the submitter frames Japan's ruling as a significant capstone to the DABUS saga. The choice of headline emphasizes the finality of the decision rather than treating it as an isolated national matter.
The editorial reframes Thaler's decade-long campaign as a deliberate probe of whether patent law's foundational concept — the inventor as a legal person with mental states, rights, and duties — can survive AI-generated output. The fractal food container was never the point; the point was to force courts to articulate why machines can't hold the role, and they consistently have.
Japan's Supreme Court has ruled that an artificial intelligence system cannot be named as the inventor on a patent application. The case ends a multi-year run by Stephen Thaler and his DABUS system — the same AI he has spent nearly a decade trying to get onto patents in jurisdictions around the world. Japan's lower courts had already rejected the application; the Supreme Court's refusal to overturn them makes the position final under Japanese law.
The court's reasoning tracked the statute closely. Japan's Patent Act refers to inventors in language that presupposes a natural person — someone who can hold rights, transfer them, and be identified for the purposes of employment-invention rules. The judges declined to read the word "inventor" as an abstract functional role that a model could occupy. If Japan wants AI inventorship, the court said in effect, the Diet has to legislate it. Courts won't invent it into existence.
With this ruling, every major patent authority that has heard the DABUS argument — the USPTO, the UK Supreme Court, the EPO Boards of Appeal, and now Japan's Supreme Court — has landed in the same place: the inventor must be human. South Africa remains the lone outlier, and even there the grant was administrative rather than the result of substantive review. The global position on machine inventorship is, for the moment, settled.
The DABUS cases were always more interesting as a stress test than as a serious bid to patent a fractal food container. Thaler was probing whether patent law's foundational concept — the "inventor" as a legal person with mental states, rights, and duties — could survive contact with systems that generate novel technical output without any of those attributes. The answer, across four continents, is that it can't survive that contact by judicial reinterpretation. It can only change by statute.
That matters more than the DABUS applications themselves, because the number of engineers with AI in their patentable pipelines is no longer a rounding error. Drug discovery pipelines routinely surface candidate molecules from generative models. Chip designers use RL agents to place cells. Materials labs run diffusion models over crystal structures. In each case the interesting question is not "can the model be the inventor" — nobody serious was asking that — but "how much human contribution is enough to sustain a valid patent when the model did the heavy lifting?"
The Japanese ruling doesn't answer that question, but it sharpens it: if the model can't be on the patent, then the human on the patent has to have actually contributed something inventive, and that contribution has to hold up if litigated. The USPTO's 2024 inventorship guidance already tries to spell this out — a person who merely prompts a model or evaluates its outputs generally doesn't clear the bar, while a person who directs the model toward a specific unsolved problem and identifies the operative feature in its output plausibly does. That guidance now has more teeth internationally.
There's also a jurisdictional pattern worth noting. The courts that rejected DABUS did so on statutory grounds, not policy grounds. None of them said AI inventorship is a bad idea. They said their current statutes don't permit it. This is a legislature-shaped hole, and lobbies on both sides of it are already forming. Expect the pharma industry, whose economic model depends on defensible patents on AI-surfaced compounds, to push hard for clarity — probably in the direction of expanding what counts as a qualifying human contribution rather than admitting machines as inventors.
Community reaction on Hacker News (192 points) split predictably. One camp reads the ruling as obvious — patents are a legal instrument for allocating rights among people, and a model has no capacity to hold rights, so the question answers itself. The other camp argues that "inventor" is doing quiet metaphysical work and that a court declining to update the concept is just kicking the can. Both are right. The concept is doing metaphysical work, and the court is kicking the can. That's a defensible thing for a court to do when a legislature exists.
If you ship a product that uses generative models in a pipeline that could produce patentable output, three things get more important, not less, after this ruling.
First, your inventorship records need to name humans and describe what they did in language that would survive an invalidity challenge. "Ran the model and selected the best output" is not a claim to inventive step in any of the jurisdictions above. "Formulated the objective function, constrained the search space to X, and identified feature Y as the operative element" is. Get your engineers used to writing the second kind of sentence in their notebooks, because six years from now a litigator will read them.
Second, be careful about tooling that erases human traces. "Autonomous" agent frameworks that log only the final artifact are a patent-hygiene liability. You want provenance that lets you point at a specific engineer's specific decision as the seed of the inventive step — logs that show which human bounded the problem, chose the model, and picked the winner. Systems that record only the model's output leave you defending patents with no named inventor whose contribution you can articulate.
Third, treat cross-border filings as the constraint. Japan just aligned with the US, UK, and EPO, which means the question is no longer "which jurisdiction will let us list the AI" but "can we describe a human contribution that clears every one of these bars simultaneously." Draft for the strictest reader. That is currently a moving target, but it is moving in one direction.
The interesting fight is no longer whether AI can be an inventor — it's what counts as enough human contribution when the AI did most of the work, and that fight is going to be resolved in specific patent-invalidity litigation, one drug and one chip at a time. Japan's ruling closes a door that most practitioners had already assumed was closed. The door that's still open — where inventive step lives in an era of capable models — is the one worth watching.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.