The Inadvertent Infringer: UK Firms Face Growing IP Liability as AI Tools Outpace Legal Literacy

The Problem Nobody Wants to Own

On 1 October 2026, IP partner Oliver Fairhurst of Lewis Silkin published a detailed breakdown of the legal risks accumulating around generative AI use in UK businesses. The piece drew on the firm's AI Agenda 2026 conference, held in London on 16 September. It is not a dramatic filing or a court ruling. But it is, in its quiet way, one of the most practically important documents this desk has seen in recent months.

The core argument is this: AI tools are now capable of generating infringing content - copied images, near-identical text, reproduced brand assets - at a cost of less than a cup of coffee per instance. The legal exposure that follows does not care whether the operator knew what was happening. Intention is not an element of copyright infringement in UK law. It is not an element of trade mark infringement either, under most of the heads of claim available to a brand owner. Ignorance, in short, is not a defence. It is merely an aggravating factor in the damages conversation.

Fairhurst's session addressed three distinct areas of IP law: copyright, trade marks, and patents. Each is in a different state of flux. Understanding the differences matters.

Copyright: The Appeal That Rewrites the Map

The copyright position in the UK turns, for the moment, on Getty Images v. Stability AI [2025] EWHC 2863 (Ch), the first, and still the only, copyright judgment on generative AI in this jurisdiction. Mrs Justice Joanna Smith's November 2025 ruling found that Stability AI was not liable for secondary copyright infringement arising from the download of copies of Stable Diffusion into the UK from overseas servers. The reasoning was narrow: the conduct did not satisfy the statutory definition of 'infringing copy' for the purposes of importation.

Getty has been granted permission to appeal. The Court of Appeal will hear arguments, likely in late 2026 or early 2027, on precisely that statutory question: what does 'infringing copy' mean when the 'copy' is an AI model whose weights may or may not encode protected works.

The stakes are not abstract. If the Court of Appeal reverses, then any foreign-trained AI model downloaded and deployed in the UK could constitute an act of secondary copyright infringement under the Copyright, Designs and Patents Act 1988, even if the model contains no extractable copies of protected works in a traditional sense. Every UK business using a US-developed image model would need to reassess its exposure overnight.

The Lewis Silkin analysis also raises the question of AI model weights as potential copies of training data. This is the same theory being tested in Andersen v. Stability AI Ltd., No. 3:23-cv-00201 (N.D. Cal.), which went to jury trial before Judge William Orrick on 8 September 2026, the first US case to put the generative AI copyright question directly to a lay fact-finder. The central theory, stripped of its technical scaffolding, is that the model's weights constitute unauthorised copies of the training images. If a jury accepts that, distributing model weights is distributing infringing copies. Every downstream user who runs the model is potentially making one.

No verdict has been publicly confirmed at time of publication. The docket last recorded activity on 28 September 2026. Watch it.

Meanwhile, the UK government's copyright consultation response, promised for spring 2026, has still not landed in final form. The Secretaries of State for Science and Culture signalled a reset in January 2026, walking back any apparent preference for an extended text-and-data mining exception. The government now says it is gathering further evidence and monitoring international developments. Translation: Whitehall is watching the courts, in both jurisdictions, before committing.

One specific statutory provision is already on notice. The protection for computer-generated works under section 9(3) of the Copyright, Designs and Patents Act 1988, the provision that currently gives UK-generated AI outputs a sui generis copyright protection not available in most other major jurisdictions, may be removed. The government has said it will continue monitoring, but the direction of travel is clear. Businesses building products around AI-generated content should not be designing compliance frameworks around a provision that may not exist in 18 months.

Trade Marks: Where the Risk Is Immediate

The trade mark picture is, if anything, more immediately dangerous for UK businesses than the copyright one. The Getty v. Stability AI judgment did find limited trade mark infringement, concerning watermarked synthetic images, even while dismissing the copyright claims. The court emphasised that liability in the generative AI context depends heavily on specific circumstances: the user prompts, the model design, and the extent to which the output is plausibly associated with the claimant's registered marks.

That is a conditional liability framework. Its practical consequence is that a business deploying an AI image tool that produces outputs bearing a competitor's watermark, logo, or brand identity has generated an infringing article, regardless of whether the user intended to do so. The Advertising Standards Authority has separately taken the position that deepfake advertising content implicating real brands is actionable under its codes. The legal and regulatory risk is stacking.

Against this backdrop, the decline in UK Class 35 trademark filings is striking. AIBD's analysis of IPO data recorded just 323 filings in Q4 2026, a fall of 97.2% against the prior period. Class 35 covers advertising, business management, and office services, precisely the service categories most exposed to AI-generated output liability. Whether that collapse reflects genuine market retraction, a filing-date anomaly in early Q4, or businesses pausing to reclassify under the 13th Edition of the Nice Classification (in force from 1 January 2026, which introduced 'artificial intelligence as a service' expressly into Class 42) is not yet clear. But the direction is notable. Brand owners are not rushing to lock in protection in the service classes most relevant to AI-adjacent businesses. That is a risk management failure.

The UKIPO's planned abolition of series trade mark applications, under which applicants could file up to six similar marks in a single application, will add cost and friction to an already stretched filing environment once the new digital service launches. Each variant will require a separate application. The timing could not be less convenient.

Patents: One Question Settled, One Wide Open

On patents, the position is the clearest it has been in years, at least on one question. UK law now confirms that an AI tool cannot be named as an inventor. Recent decisions have reinforced that only natural persons can hold inventorship. The US Patent and Trademark Office reached the same conclusion in November 2025 guidance, and the US Supreme Court declined certiorari in Thaler v. Perlmutter on 2 March 2026, leaving the Federal Circuit's ruling intact. The question is closed, for now, in both jurisdictions.

The Lewis Silkin analysis flags a second, less settled question: whether artificial neural networks are themselves patentable subject matter under UK law. Recent decisions suggest they may be, as a form of technical process, provided the application clears the computer program exclusion under section 1(2) of the Patents Act 1977. This is not academic. AI model architectures are being filed for patent protection. The validity of those filings, and the scope of the monopolies they would confer, has not been tested at the appellate level in the UK.

The Practical Reckoning

Think of it this way: the UK IP framework for AI is like a Victorian water main running beneath a modern city. The pipes mostly hold, but they were not designed for this pressure, and nobody is quite sure where all the joints are.

For UK businesses using generative AI tools, the Lewis Silkin analysis identifies the core risk clearly: both deliberate and inadvertent infringement are live. The practical guidance, review data provenance, check whether your AI supplier's training practices are contractually warranted, do not assume fair dealing exceptions cover your use case, is sound but insufficient if counsel are not integrated into procurement decisions from the outset.

For brand owners, the immediate priority is monitoring AI output in the channels where their marks are active. The ASA will not wait for the Getty appeal to conclude before taking action on deepfake advertising.

The next procedural milestone is the Getty v. Stability AI Court of Appeal hearing, expected no earlier than late 2026. The Andersen jury verdict, whenever it arrives, will set the American frame. The UK government's final copyright consultation response will follow on its own schedule, which is to say, no one knows when.

The statute wasn't drafted for this technology. And it shows.