IT EN
Start a trial

Operational notes Observatory

$500,000 for an image no one ever shared: xAI takes Minnesota to court

7 min read

A camera lens diaphragm closed in a spiral around a central hole, shot in macro, in black and white
The blades close around a single hole: they decide how much light passes, not whether the scene exists. A nudification law tries to do the same to an image generator.

On 27 July 2026 xAI filed suit in the United States District Court for the District of Minnesota against Keith Ellison, the state’s Attorney General, to block a law it considers unconstitutional. The law, House File 1606, creates a new section, 325E.91, in the Minnesota Statutes, and bans anyone who owns or controls a website, app or service from letting users produce “nudified” images — altered to show an identifiable person’s intimate parts without consent, realistically enough to pass as genuine. It takes effect on 1 August 2026: five days after the suit was filed.

The target of the suit is the company’s exposure, not the principle behind the law. In the complaint’s first paragraph, xAI states that it already bans its users from generating nude or sexualised images without consent, and that it has itself sued users who evaded its technical blockers to produce them anyway. The company’s problem is how the law is written: strict liability, with no requirement to prove intent, which applies — the complaint states — “even if the depicted persons consented — or created the image themselves — and even if the image is never shared”. There is no exemption for a company that has already deployed state-of-the-art technical controls.

The facts, in order

  • Since 2024: xAI has been developing Grok Imagine, the image- and video-generation and editing feature built into Grok. On 5 August 2025, Musk said on X that the product was generating more than 20 million images a day.
  • 3 June 2026: Reuters reports that British Labour MP Jess Asato has sued xAI in a UK court, alleging that Grok generated fake sexualised images of her; Prime Minister Keir Starmer publicly backs her case.
  • 28 July 2026: Reuters reports that Asato is seeking a court order to stop Grok from generating that kind of image; the BBC reports that court filings discussed at the hearing describe it as a “gross violation” of her dignity.
  • 2026: Minnesota enacts Chapter 72 of its session laws (House File 1606), creating Minn. Stat. §325E.91.
  • 27 July 2026: xAI files X.AI LLC v. Ellison, case no. 0:26-cv-03425, asking the court to declare the law unconstitutional and to bar its enforcement against the company.
  • Same filing: absent an injunction, xAI will restrict Grok Imagine’s image-editing feature for Minnesota users “in various ways” before 1 August — not the whole product, just the editing function.

What the law bans, and why xAI is challenging it

Minn. Stat. §325E.91 lets the Attorney General seek up to $500,000 per violation (subd. 5(a)), alongside a private right of action for anyone depicted: treble damages, punitive damages, injunctive relief and attorneys’ fees (subd. 4). It requires no proof that the image was distributed, and no proof of intent on the part of whoever runs the tool: it is enough that the image was produced. The only exemption covers tools that require “the technical skill of a user” to make the alteration — manual photo-editing is out, an automated generator is in.

This is a deliberate design choice, not an oversight. Minnesota already has two sexual-deepfake statutes — 617.262 and 604.32 — that require proof of distribution, lack of consent and intent, plus a federal analogue, the TAKE IT DOWN Act (47 U.S.C. §223h), with exceptions for content of public concern or for medical or educational purposes. xAI’s complaint cites precisely these two regimes to argue that “less restrictive alternatives” already exist: if they were enough, the company argues, Minnesota did not need a broader tool. This is an argument built on American constitutional law — the First Amendment — that an Italian or European court would not apply. But the underlying question — how broad a ban needs to be to reach people who never intended any harm — is the same one facing any lawmaker, European ones included, when drafting an obligation for providers of generative tools.

As of this article, Ellison has not made any formal response public: we flag that as a missing data point, not as an absence of position. For now, the state’s position is the one written into the statute itself: a strict, intent-free prohibition, adopted — judging by the comparison xAI itself draws in its complaint — after the earlier laws proved hard to enforce precisely because of the burden of proving intent.

The lesson for buyers and integrators of image generators

Three points, independent of how the case is decided.

First: a local law can force a global product to change within days, not quarters. Five days passed between the filing of the suit and the law’s entry into force. Anyone who has built a workflow — marketing, prototyping, client content — around a consumer provider’s image-generation feature needs to plan for that feature narrowing overnight, by decision of a single state, with no contractual notice.

Second: a provider’s stated moderation policies are technical measures, not a legal shield for whoever uses them downstream. xAI states that it already bans, by contract and by technical blocker, exactly what Minnesota’s law bans — and is still being sued over the case where a user evades those blockers. Anyone reselling or embedding an image-generation model inherits the same problem: a provider “banning” a use in its own terms does not mean whoever distributes it downstream is covered under the law that applies in their own jurisdiction.

Third, the European angle: there is, for now, no nudification law here with comparable strict-liability thresholds, but the AI Act still imposes a transparency duty on AI-generated or altered content — Article 50, in application from 2 August — and the split between whoever builds the system and whoever puts it in front of their own users carries its own consequences for who answers for what. It is not the same lawsuit, but it is the same question: who answers for a generated image, the model’s builder or whoever makes it available? This is not the first time a public buyer has had to weigh up an xAI product under active litigation: it happened a year ago too, when Grok’s antisemitic outputs preceded a $200 million Pentagon contract by six days.

What to do before signing

  1. Ask the provider for the list of jurisdictions where a feature could be restricted or suspended, and with what contractual notice — not the moderation policy, the service-modification clause.
  2. Check who bears liability if you resell or embed the tool in a product for your own clients: the provider’s policy does not automatically cover you under the law that applies to you.
  3. Distinguish stated controls from verifiable ones: blockers, filters and lawsuits against users who evade them are evidence of effort, not guarantees of outcome.
  4. For content aimed at the public, keep a log of who requested what, and when: that is the proof you will need, whichever jurisdiction ends up challenging the use.

The technical point behind the case is simpler than the constitutional dispute: an automated tool lowers the skill needed to produce an alteration, and that is exactly what Minnesota’s law chooses to target — not manual photo-editing, automation. Anyone building an image-generation capability into a business process or a public service should ask the same question as the legislator: who really decides what the tool produces, with what verifiable log. That is why we deliver generative AI systems in only two modes: on-premises, in the client’s own environment, with the moderation policy and usage log under their direct control; or a dedicated cloud — an environment reserved for a single client, with a dedicated VPN, a data centre in Italy, and premises directly staffed by us. In both, who is entitled to generate what is written into the contract, not a policy that changes with the user’s jurisdiction: the same question addressed in hardware and open-weight models.

Integrating or reselling an image- or video-generation tool, and you haven’t yet mapped which jurisdictions could restrict it? Talk to us for thirty minutes.

Sources