For three years British courts have sidestepped the question of whether the company that builds an AI model answers for what its users produce with it. A written defence filed with a British court has now put that question on the record.
SpaceX Super Intelligence, the company formerly known as xAI and now part of Elon Musk’s SpaceX, submitted the document on Oct. 29, and Reuters and other outlets reported its contents on Nov. 19. The claim was brought by Jess Asato, a Labour member of parliament.
The defence rests on two arguments. The company says the images at issue are recognisable at a glance as fabricated and therefore cause no actionable invasion of privacy, and that it did not determine the content of the prompts users submitted. It also concedes something: the filing acknowledges that material breaching its terms has been removed, including a video that depicted the lawmaker sedated and being prepared for a sexual assault.
She is asking the court to order measures that would be effective and permanent, rather than the takedown-after-complaint that has become standard practice. That remedy, more than damages, is what the case is about, and it would bind the company’s product design rather than its moderation queue.
The mechanism matters here. Image models generate from a text prompt, and the developer ships filters that estimate whether a request or an output crosses a line. Those filters are probabilistic, they are tuned for volume rather than for any one person, and they are routinely updated after a failure becomes public.
The defence’s first limb treats obviousness as a shield: if no reasonable viewer believes the image, the harm is thin. Its second limb treats the prompt as the cause: the developer supplied a general tool, and a user chose the target.
Neither limb is comfortable for a court. Obviousness speaks to belief, and the circulation of fabricated sexual imagery does its damage among people who already suspect the subject, whether or not they credit the file. Causation speaks to control, and the argument that a developer cannot foresee a prompt sits awkwardly beside a product that generates likenesses on request.
British statute offers little guidance. The Online Safety Act governs user-to-user services and the content they host, not the developers upstream. Data protection law applies where personal data is processed, and the case will turn on whether generating a likeness counts as processing and whether the developer acts as a controller of it.
Parliament did move in 2025. The Data (Use and Access) Act added offences for creating sexually explicit deepfakes and placed a duty on internet services to take reasonable steps to prevent their creation, enforced by the communications regulator. The duties are written for services, and they do not name the companies that build the generators.
That gap is partly deliberate. Ministers have spent two years courting AI investment and have repeatedly delayed promised legislation on model safety, while the European Union’s AI Act already imposes transparency and documentation duties on general purpose systems. A British ruling would land in the space the statute left open.
The remedy has a price. Hash matching, prompt blocking and output classifiers cost money to run at the scale of a consumer product, and each produces false positives that users notice. A court order to implement them permanently would set a floor for the industry, and a floor is cheaper to meet before a judgment than after one.
Timing favours the defendant in one respect. British civil claims of this kind typically take two years to reach trial, and the technology under examination will have been rebuilt at least twice by then. It favours the claimant in another, because each rebuild creates a new record of what the company knew.
Lawyers following the case expect it to be cited well beyond Britain. American litigation runs into a liability shield written for platforms that host other people’s content, which maps poorly onto a company that makes the generator. Defamation and privacy claims in several jurisdictions are testing the same ground in parallel.
The commercial consequences would be concrete: insurance cover for model providers, warranty language in enterprise contracts, and the diligence questions that buyers of AI services now ask. None of that depends on the lawmaker winning outright.
What the filing has already changed is the record. A company that once described its systems as neutral instruments has told a court, on paper, where it believes responsibility stops, and it has put the removal of the material in the same document.
The judgment will decide whether that boundary holds. If it does, the effective remedy for anyone targeted by a generated image remains a request to a company that has already said it decides such cases itself.


