Grok and sexualised deepfakes: when AI humiliates before it is controlled
A British MP is suing xAI, Elon Musk’s company, after non-consensual sexualised images were generated with Grok.
The case could look like yet another row around X. It is more serious than that.
Because here, the subject is not only “a user misused a tool”. The heart of the file is harder: can an AI built into a social platform be designed in a way that makes the public humiliation of real people easier?
That is the question raised by the legal action filed on 3 June 2026 by Jess Asato, a British Labour MP. She says Grok was used to generate fake sexualised images of her, including bikini images, after she had precisely criticised non-consensual pornographic deepfakes.
According to Reuters and Associated Press, the claim targets xAI for invasion of privacy, misuse of private information and breach of personal data rules. She is seeking damages, a formal declaration of illegality and an injunction to prevent further abuse.
But the most sensitive point is elsewhere: Asato wants it established that responsibility does not stop at the user typing the prompt. She wants the designer of the tool to be capable of being held responsible too, when the system makes this kind of abuse possible.
A timeline that makes the case explosive
The scandal does not start on the day of the claim.
It starts in late December 2025, when Grok’s image editing feature comes into massive use on X. On 29 December, image generation usage rises sharply after a feature was promoted allowing one-click edits of images posted on the platform.
Between 29 December 2025 and 8 January 2026, the Center for Countering Digital Hate analyses a sample of 20,000 images out of a total of 4.6 million images generated by Grok over that period. The organisation estimates that around 3 million sexualised images were produced and published on X in 11 days.
That figure is not an official count from xAI. It is an estimate by the CCDH, a campaigning organisation specialising in the fight against online hate. But even with that caution, the order of magnitude is striking: we are not talking about a handful of isolated cases.
On 9 January 2026, some features are restricted to paying users after the public outcry.
On 14 January 2026, xAI adds new technical restrictions to limit requests aimed at undressing real people.
But on 3 February 2026, Reuters publishes a test run by nine journalists. Despite the announced restrictions, Grok still generates, in the majority of the cases tested, sexualised images of people, even when the prompts specify that the subjects do not consent.
It is this sequence that makes the case explosive. The guardrails did not merely arrive late. They also looked insufficient.
Our own test: the problem does not stop at Grok
To illustrate this article, Le Recul also tested generating an image with AI, in our case through ChatGPT. Without using the image of a real person, we were able to obtain a visual showing a woman in underwear, with a sexualised staging tied to the deepfake theme.
This test is not intended to produce or distribute a pornographic deepfake. What it mainly shows is how quickly an AI can approach a sensitive zone: female body, sexualisation, adult warning, suggestion of explicit content.
The difference is essential: our visual is a fictional illustration, with no real person targeted. But in the cases denounced around Grok, the problem comes precisely from the step to the image of a real, identifiable person, sexualised without consent.
That is exactly the danger: when creation becomes as simple as a prompt, control cannot arrive afterwards.
How we use AI ourselves, and where we stop, is set out in our AI charter.
What makes this case different
Sexual deepfakes have existed for years. What is new is how ordinary they have become.
Before, you had to find a dubious site, install a tool or go through specialist forums. Today, the danger comes from a feature built into a massive social network.
With Grok, users could reply to a public photo and ask the AI to modify the image. In some cases reported since January 2026, this made it possible to generate sexualised images of real women, sometimes without consent, sometimes with the result publicly visible on X.
The Guardian reports that a “bikinification” trend circulated on X, with images generated from real women.
In the Jess Asato case, Reuters also reports the existence of an AI-generated video depicting her as if she had been drugged and prepared for a sexual assault. So the subject is not limited to a few “bikini” images. We are talking about a humiliating, sexual and potentially violent use of a real person’s image.
That is the problem: the guardrails arrive after the scandal. The victims arrive before.
The false debate: “it is only an image”
Saying “it is only an image” is a comfortable way of avoiding the subject.
A non-consensual sexualised image does not need to be “real” to do damage. It can be shared, commented on, repurposed, archived, reposted, searched for. It can reach the family, colleagues, clients, voters, the school, the employer.
The harm does not only come from the nudity or the sexualisation. It comes from the loss of control.
Someone takes your face, your body, your public image, then lets a machine manufacture a humiliating version of you. After which the platform sometimes lets that version circulate.
This is image violence.
And when the tool is simple, public and built into a social network, scaling up becomes immediate.
Can xAI hide behind its users?
This is the most important legal point.
xAI can say that users are responsible for the prompts. That is partly true. Whoever asks for a humiliating image carries responsibility.
But that does not answer the real question: why does the system accept that kind of request? Why is the tool capable of producing the result? Why do the restrictions arrive afterwards? Why can the image of a real person be turned into sexualised content?
The classic platform defence is to say: “we only host” or “we cannot control everything”.
But a generative AI is not a simple host. It manufactures.
It interprets a request, transforms an image and produces new content. If that content is humiliating, sexual and non-consensual, the tool cannot be treated as a neutral pipe.
That is exactly what the Asato case could test: is the maker of the AI responsible for the design of its system?
The same question is being fought over in the United States, where OpenAI is backing a bill that could clear it after a hundred deaths.
A case that could set a precedent
On 3 June 2026, Jess Asato files a claim at the High Court in London against xAI.
She invokes in particular misuse of private information and the protection of personal data. She is seeking damages, a formal declaration that what happened to her was unlawful, and an injunction to stop xAI from continuing to allow this kind of abuse.
Her lawyer presents the case as a test of the responsibility of AI developers.
That is the central point: the case is not only trying to punish a user. It is trying to determine whether the company designing the AI can be held responsible for its design choices.
That distinction changes everything.
If a tool is designed in a way that easily allows the sexualisation of real people, can the maker then say: “it is not us, it is the user”?
The Asato case will try to answer that question.
This is not one isolated victim
The Jess Asato case is the most visible in the United Kingdom, but it is not alone.
Associated Press reports that in January 2026, Ashley St. Clair, the mother of one of Elon Musk’s children, also started proceedings in New York against xAI. She says Grok generated explicit images of her, including one in which she was allegedly depicted as a minor.
Reuters indicates that in March 2026, the city of Baltimore sued xAI, arguing that Grok’s ability to create fake sexualised images breached local consumer law.
On 5 June 2026, The Guardian reports that other claimants came forward after the Asato case. According to lawyer Ravi Naik, some had difficulty getting the images removed from X before they obtained legal support.
So this is not just “an MP against Elon Musk”. It is a series of cases seeking to establish a broader responsibility for AI companies.
Is it only Grok?
In this particular case, the central tool is Grok, developed by xAI and built into X.
But the problem is not limited to Grok. The Guardian reports that, on the very day Asato’s proceedings were announced, a new sexualised image of her was allegedly generated with a different AI tool.
That matters.
Grok is the visible case, because it is tied to X, to Elon Musk and to massive distribution. But the flaw is wider. As soon as a tool allows realistic images of people to be edited or generated without strict control, the same abuse can reappear elsewhere.
So the question is not only “is Grok bad?”.
The real question is: how many AI tools can produce this kind of humiliation before the victims have any recourse at all?
The law is starting to catch up
The United Kingdom has already hardened its position.
British police state that it is illegal to create, or to ask someone to create, a fake intimate image of a person without their permission.
As early as January 2025, the British government had announced that it wanted to criminalise the creation of non-consensual sexually explicit deepfake images, with up to two years in prison for certain offences.
In Europe, the AI Act is progressively imposing transparency obligations on synthetic content and deepfakes. The European Union also adopted a directive against violence towards women in 2024, covering forms of cyber violence and image abuse.
But the law runs behind.
Judicial time is counted in months or years. The time it takes to make a deepfake is counted in seconds.
The real risk: industrial humiliation
Grok is not the only tool able to produce or modify images. But this case hits hard because it shows what happens when a powerful AI is connected to a social platform, to a culture of pile-ons and to fragile moderation.
The sexual deepfake is no longer only pornographic content. It is becoming a social weapon.
A weapon to silence. A weapon to ridicule. A weapon to punish a visible woman. A weapon to attack a journalist, an elected official, an influencer, a colleague, an ex-partner, a teenager.
And it is no accident that women are overwhelmingly targeted.
The aim is not always desire. Often, the aim is degradation.
Le Recul
This case is not a simple user excess. It is a full-scale test of the responsibility of generative AI.
Grok promises a freer, less filtered, more direct AI. But a “free” AI that can sexualise real people without consent is not only free. It is dangerously available.
Our own illustration test confirms it on another scale: even without targeting a real person, an AI can quickly produce an image sexualised enough to make the risk clear. With a real photo, a real name and a real target, the same mechanism changes in nature. It is no longer illustration, it is an attack on someone’s image.
This is the same pattern we saw when researchers pushed frontier models into simulated nuclear crises and watched how fast they crossed the lines they were supposed to hold: the guardrails were not where the marketing said they were.
The real question is not whether AI can create images. It can.
The real question is: why can it create those ones?
As long as the guardrails are added after the damage, safety is not a priority. It is a clean-up operation.
And in this file, the clean-up arrives too late for the people already exposed.
What to take away
Jess Asato is suing xAI over sexualised images generated with Grok.
The scandal starts in late December 2025, with an explosion in the use of Grok’s image editing on X.
The CCDH estimates that around 3 million sexualised images were produced and published on X in 11 days, between late December 2025 and early January 2026.
xAI added restrictions in January, but Reuters found in February that Grok could still produce sexualised images in several of the cases tested.
Le Recul was also able to generate a sexualised illustration image with AI, without using a real person, which shows that the problem goes beyond a single tool or a single company.
The case could become a precedent on the responsibility of AI companies.
The problem goes beyond Grok: other tools can generate the same kind of abuse.
The real stake is not only removing the images. It is preventing the tool from producing them.