Your coding assistant wrote half of the latest release, the pitch-deck images came out of a generator, and the product itself has a model at its core. That is an ordinary description of a 2026 startup, and it raises at least four separate intellectual property questions. Some now have settled answers: the UK Supreme Court has ruled twice on AI and patents since December 2023. Others are genuinely open, and in March 2026 the government said it would gather more evidence before changing copyright law. Here is the map.
Protecting what you build: patents and trade secrets
Start with the settled part. An AI system cannot be named as an inventor on a UK patent application. In Thaler v Comptroller-General [2023] UKSC 49 the Supreme Court held that an inventor under the Patents Act 1977 must be a natural person, so applications naming an AI machine called DABUS as inventor were rightly treated as withdrawn. In practice the rule catches almost nobody. Using AI as a tool in research and development does not stop a patent; the human who made the inventive contribution is simply named as inventor in the usual way. Where that line sits, and what records help you hold it, is the subject of our guide on patenting AI-assisted inventions.
The larger question for AI companies is whether the software itself is patentable, since the Patents Act excludes a program for a computer “as such”. This ground moved in February 2026. In Emotional Perception AI Ltd v Comptroller-General [2026] UKSC 3 the Supreme Court held that a trained neural network is a computer program, yet allowed the appeal. The music-recommendation invention in issue should not have been refused under the exclusion, and the application went back to the UK Intellectual Property Office (UKIPO) for the rest of examination. The court also discarded the four-step test examiners had applied for some twenty years, in favour of an approach aligned with European Patent Office practice. Under UKIPO guidance issued on 14 July 2026, an invention that uses technical means clears the first hurdle, and scrutiny falls on whether its features contribute to a technical solution of a technical problem. AI inventions are no longer turned away at the door for being software. The argument has moved to technical substance, which is where careful drafting earns its keep.
Many AI businesses still keep their most valuable assets out of the patent system altogether. A patent application is published around 18 months after filing, so your method becomes public whether or not a patent is ever granted. Model weights, training pipelines and data-curation know-how are often held as trade secrets instead, and those last as long as secrecy lasts. The catch sits in the definition: under the Trade Secrets (Enforcement, etc.) Regulations 2018, information qualifies only if it has been subject to reasonable steps to keep it secret. Loose habits can dissolve the right, which is where the hygiene section below comes in.
Who owns what your AI tools produce
Ownership of AI-generated content is the least settled area of the lot, and the government has said as much. Section 9(3) of the Copyright, Designs and Patents Act 1988 provides that the author of a computer-generated work, one with no human author, is taken to be the person who made the arrangements necessary for its creation. But copyright also requires a work to be original, and the modern test for originality is the author’s own intellectual creation, a test built around human creative choices. The government’s December 2024 consultation acknowledged the contradiction. It is unclear how a wholly machine-made work can be original, and no court has decided the point. The report that followed in March 2026 proposed removing this special protection in due course, absent evidence of its value, while confirming that copyright continues to protect works created with AI assistance. For now s.9(3) remains law, and the uncertainty remains with it.
The practical reading is straightforward. Treat the strength of your copyright as roughly proportional to the human creative input: a designer who selects and reworks AI output into a finished asset is on much firmer ground than a prompt used as-is. Then look at the contracts, because they decide where whatever rights exist end up. The leading AI tools’ terms generally confirm that you own the output so far as the law allows, wording that stops short of promising there is any copyright to own, and at least one major provider states plainly that output may not be unique and that other users may receive similar output. What you get can also differ by subscription tier. Inside the company, the housekeeping is familiar: employment and contractor agreements that deal with IP expressly, and copyright assignments, transfers of ownership, in writing and signed, without which they are simply not effective.
Using AI on other people’s content
If you train or fine-tune models on material you did not create, the headline question is whether AI training on copyright works in the UK is lawful without a licence. No UK court has answered it. The High Court’s November 2025 judgment in Getty Images v Stability AI [2025] EWHC 2863 (Ch) left it open, because Getty abandoned its main training claim mid-trial. There was no evidence the model had been trained in the UK, and UK copyright does not reach acts done abroad. Where a model is trained matters, and that point carried much of the case.
What the court did decide was narrower. Getty’s secondary infringement claim, which turned on whether the model itself was an infringing copy, failed: the model does not store the works it learnt from, although the court accepted that an intangible electronic copy can be an “article” for these purposes. Getty’s wins were limited to trademark findings about watermarks reproduced in outputs of older model versions, findings the judge described as “both historic and extremely limited in scope”. Getty was granted permission in December 2025 to appeal on the secondary infringement point, and no Court of Appeal decision had been handed down as at July 2026.
Nor does any statutory exception cover commercial training. Section 29A of the Copyright, Designs and Patents Act 1988 permits text and data mining, copying a work for computational analysis, only for non-commercial research. The government’s earlier preference for a broader opt-out exception was shelved in the March 2026 report. Businesses that build on third-party material commonly deal with the gap by licensing what they train on, and by keeping records of what went into the model.
Everyday hygiene with AI tools
A lot of AI-related IP damage is self-inflicted. Someone pastes source code or an unfiled invention write-up into a public chatbot to save an afternoon. The National Cyber Security Centre is blunt about what happens next: queries submitted to a public AI tool are visible to the provider, they are stored, and they will almost certainly be used to develop the service at some point. Set that against the trade secret definition above. Information typed into a public tool is very hard to describe as protected by reasonable steps, and once secrecy has gone the right has gone with it.
Many businesses now run a short internal AI-use policy covering which tools are approved and what must never be entered into them. Vendor terms deserve the same attention as the tools. The NCSC’s advice is to read a provider’s terms of use and privacy policy before submitting material, and the points that matter most are what the provider retains and whether your inputs can be used to train its models.
AI logos and your brand
One narrower point. A logo made with an AI image tool can normally still be registered as a trademark, because registration under the Trade Marks Act 1994 turns on whether the sign is distinctive and whether the owner of an earlier mark objects; the UKIPO does not ask who, or what, drew it, and does not check copyright ownership. The weak spot sits behind the registration: if copyright in the artwork is doubtful for the reasons above, stopping copying of the artwork itself and assigning clean rights on investment or exit both become harder. Our guide on trademarking an AI-generated logo works through the fixes.
Where this leaves you
None of the open questions needs to stall a build. The habits that matter are cheap now and expensive to retrofit: dated records of who contributed what whenever AI sits in the development loop; visible human creative input in any asset you may one day need to enforce; vendor terms and contractor agreements read before your confidential material or your brand depends on them; and a working assumption that anything pasted into a public tool is no longer secret. A business that can show those four things has already answered most of what an investor’s due diligence will ask.
If AI runs through your product or your workflows and you are not sure what you own, or what you may have given away, ScalaxIP can help you map it. Get in touch.
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