Your team runs a machine-learning model over ten thousand candidate compounds overnight, and it flags three worth making. One works. Or a generative design tool proposes a bracket geometry no engineer had sketched, and it passes every load test. Somewhere between the celebration and the board update, a worry surfaces, if AI found it, can you still patent it?
The short answer is yes. Using AI as a tool does not stop a patent. UK law asks for two things: a human must be named as the inventor, and that human must have made the inventive contribution. AI can sit anywhere in your research process; what it cannot be, as the law stands, is the inventor.
The DABUS case
The human-inventor rule reached the Supreme Court because one applicant refused to name a person. Dr Stephen Thaler filed two UK patent applications and named as inventor DABUS, an AI system he owned. The machine, he said, had devised the inventions autonomously. On his own case, no human had invented anything.
Patent law requires the applicant to identify the person believed to be the inventor. Dr Thaler would not, so the UK Intellectual Property Office (UKIPO) treated his applications as withdrawn, and in Thaler v Comptroller-General [2023] UKSC 49 the Supreme Court unanimously upheld that result. Lord Kitchin, giving the lead judgment in December 2023, held that an inventor under the Patents Act 1977 is the “actual deviser” of the invention and must be a natural person. A machine is not a person, so DABUS could not be an inventor. Nor did owning the machine give Dr Thaler any right to patent what it produced on its own.
Two things about the case matter for working teams. It failed because Dr Thaler insisted no human had devised the inventions; the court never had to deal with an application that had a genuine human contribution behind it. And the court decided only what the current Act means. Whether advances generated autonomously by machines should be patentable is a question it expressly left open.
Tool or inventor: where the line sits
None of this troubles the ordinary AI-assisted invention. Where a human makes the inventive contribution using AI as a tool, that human is the actual deviser and is named as inventor in the usual way. The government examined the question in its consultation on AI and intellectual property; its June 2022 response described AI as a tool used by human inventors and made no change to the law. That remains the position, and the area stays under review.
UKIPO practice is blunt on the point. Its Formalities Manual states that “AI Inventor” is not acceptable on an application because it does not identify a person. Name a machine and you will be asked to substitute a human; refuse, and the application is treated as withdrawn, exactly as Dr Thaler’s were.
So the useful question inside an R&D team is who came up with the inventive concept. The person who framed the problem, directed the AI’s part in solving it and recognised in its output the thing that works will usually have a strong claim to the title. Say your chemists chose what the model should screen for and understood why the surviving compound succeeded: you have an inventor to name. What UK law has no room for is the opposite case, an advance nobody can honestly claim to have devised. For those wholly machine-generated inventions there is currently no patent route in the UK, and whether there should be one is an open policy question.
Patenting the AI itself: a different trap
Everything above assumes the AI helped you invent something else, a compound or a component, say. If the invention is the software itself—a model, a training pipeline, a recommendation engine—a separate obstacle appears. The Patents Act 1977 excludes a “program for a computer” from patentability, but only “as such”, two words that carry most of the weight in this corner of the law.
For the best part of two decades, examiners and courts policed that exclusion with a four-step test from a case called Aerotel. That era ended in February 2026. In Emotional Perception AI Ltd v Comptroller-General [2026] UKSC 3, the Supreme Court considered a music-recommendation system built on an artificial neural network. The High Court had decided in 2023 that the exclusion did not bite and the Court of Appeal reversed that in 2024 and held that a trained neural network, meaning its weights and biases, is a program for a computer. The Supreme Court agreed that it is, and allowed the appeal anyway. The claim was not excluded “as such”. Aerotel was abandoned, and UK law realigned with the approach of the European Patent Office. A claim that uses technical means clears a low first hurdle as an invention; the real scrutiny then falls on whether the claimed features contribute to a technical solution to a technical problem, tested through novelty and inventive step. The application went back to the UKIPO for those remaining requirements to be examined.
The UKIPO kept pace. It withdrew its 2022 guidelines on examining AI inventions on 11 February 2026, the day of the judgment, and published new statutory guidance on search and examination of patent applications on 14 July 2026. Examiners no longer apply Aerotel, and how the new approach treats any particular AI invention will emerge decision by decision. One line survives every twist in the story, a program written by a machine is still a computer program, so code your model generates gets no special treatment.
Records worth keeping
Inventorship stops being abstract the day an application asks for names, and the right to apply for a patent starts with the inventor, so the names matter well beyond formality. Businesses with AI-heavy pipelines commonly keep a contemporaneous trail: who set the problem, who designed or configured the tools, who reviewed the outputs, who recognised why one result mattered when a hundred others did not. Dated lab notebooks, decision logs and project files all do the job.
The habit worth building above all is human review. A person who examines what the model produced, understands it and takes the inventive decisions about it is a person who can honestly be named. If nobody in the business can describe their own contribution to the advance, the problem is not paperwork.
A related discipline is confidentiality before filing. An invention made public before a filing date is generally no longer patentable in the UK, so most businesses keep the work under wraps until they have decided. Timing has traps on both sides, and we cover them in our guide on when to apply for a patent.
Where trade secrets fit better
Sometimes the audit comes back murky. The model did the surprising part; people set it running and confirmed the outcome. The nearer your process sits to that end of the spectrum, the shakier the foundation for a patent application, since somebody must be named and must genuinely be the deviser. Many businesses in that position protect the advance as a trade secret instead. Secrecy needs no registration, no official fee, no renewal and no named inventor, whereas a patent application that goes ahead is published in full about 18 months after filing. Access controls and confidentiality terms in staff and partner contracts do the work. The trade-offs are real: a trade secret is no answer to a competitor who develops the same thing independently or takes your product apart and works it out, and protection ends the moment the secret does.
For most teams using AI in R&D, though, the position is more comfortable than the worry suggests. Use whatever tools you like. Keep humans making the inventive calls, record who made them, and name people rather than machines when you file. If your product is the AI itself, the question since February 2026 is what it technically contributes, which is a matter of careful drafting under the new rules.
ScalaxIP works with founders on exactly these questions, from inventorship records in AI-heavy development to filing strategy for AI-based inventions under the new UK approach. If a patent decision is sitting on your desk, we are happy to talk it through.
Disclaimer: This article is provided for general information only and does not constitute legal advice. For advice on your specific circumstances, please contact ScalaxIP at admin@scalaxip.com
SHARE
RELATED EXPERTISE