The New Prohibition on AI-Generated Intimate Imagery: Not What You Intended, But What You Can Prove

1. Two lawsuits over Grok: one against the company, one brought by it against a user

At the end of December 2025 image editing was enabled in xAI's Grok model. In January 2026 X was flooded with sexualised images generated in response to simple prompts from anonymous users, including images of real women, identifiable by full name and face, depicted nude or minimally clothed without their consent. The United Kingdom announced an investigation on 12 January 2026, the Attorney General of California opened one of his own on 14 January 2026, Indonesia, Malaysia and the Philippines temporarily blocked access to the model, and on 3 February 2026 the French cyber police searched X's offices in Paris.

Two legal proceedings grew out of those events.

The first was filed in Israel. In March 2026 the content creator Shir Shahaf brought a civil claim in the Tel Aviv Magistrates' Court against X.AI Corp for ILS 277,000, pleading two causes of action: sexual harassment and invasion of privacy. What is novel is not the causes of action, which are well established in Israeli law, but that the claim is directed at the artificial intelligence company itself rather than at the anonymous user who entered the prompt. The claim is pending and no judgment has been handed down.

The second proceeding was filed in the United States, over the same technology, save that this time the plaintiff is the company itself. In July 2026 xAI sued one of Grok's users, a 67-year-old South Carolina resident previously arrested on suspicion of child sexual exploitation offences, alleging that he crafted misleading and sophisticated prompts in order to deliberately circumvent the safety mechanisms built into the system and turn innocuous images into sexual material.

2. What the new Regulation actually prohibits, and what falls outside it

Article 5 of the Artificial Intelligence Act lists the prohibited practices, the highest tier, which is not subject to risk management but to outright prohibition. The new Regulation adds two items to that list: AI systems that generate or alter realistic images, video or audio depicting the intimate parts of an identifiable person or that person's explicit sexual activity without their consent, and systems that generate child sexual abuse material. Both prohibitions apply from 2 December 2026.

The prohibition was not part of the Commission's original proposal of November 2025. It was introduced by the Council in its negotiating position in March 2026, following the events described above.

The Regulation clarifies that altering existing material in a way that neither increases the exposure of intimate parts nor changes the character of the sexual activity depicted does not amount to alteration for the purposes of the prohibition. This is a targeted provision, not a blanket restriction on image editing.

3. The provider is caught even without intent; the deployer only on deliberate use

The two roles are defined separately in the Act, and a party occupying both, such as a company that has built a model and also runs it within its own service, is measured against both tests.

Two terms recur throughout the Act and are worth fixing at the outset. Placing on the market means making the system available on the Union market for the first time, whether or not for payment. Putting into service means supplying it for first use, whether to a customer or for the company's own internal use. Either one is enough to bring the system within scope.

The assessment a provider must carry out is narrow and precise: is the prohibited outcome reasonably foreseeable, and can it be reached again without significant technical modification to the system. If the answer is yes, the question moves immediately to the measures in place. The recitals expressly identify the kinds of measures capable of meeting the requirement, among them training the model to refuse such requests, input-stage restrictions, filtering of outputs before they are displayed, and detecting and reporting instances of misuse. The measures must also withstand reasonably foreseeable circumvention, and where circumvention is observed or reported, appropriate correction must follow.

4. What was postponed to December 2027, and what already applies from December 2026

The Artificial Intelligence Act does not take effect all at once, but through a series of dates, each of which triggers a different set of obligations. The new Regulation moved only two of them, both relating to the high-risk regime, and left everything else in place.

• 2 February 2025The original prohibited practices under Article 5 and the AI literacy obligation under Article 4 took effect. The new Regulation did not touch them.
• 2 August 2025Obligations on providers of general-purpose AI models took effect. These too were left unchanged.
• 2 August 2026The transparency obligations under Article 50 apply in full, including notification of interaction with a machine and the marking of generated content. This is the original date around which most compliance programmes were built, and it stands.
• 2 December 2026The two new prohibitions take effect: generation of intimate imagery without consent, and child sexual abuse material. On the same date the machine-readable marking obligation under Article 50(2) also applies to systems already on the market before August 2026, for which an extension was granted.
• 2 August 2027The date by which each Member State must have at least one national regulatory sandbox in place.
• 2 December 2027Full obligations for standalone high-risk systems under Annex III. Postponed from the original date of 2 August 2026, by sixteen months.
• 2 August 2028Obligations for high-risk systems embedded in products under Annex I, a postponement of twelve months.

5. In Israel the prohibition reaches publication, not the tool that created the material

The principal Israeli provision is found in Amendment No. 10 of 2014 to the Prevention of Sexual Harassment Law, 5758-1998, known as the Videos Law. Section 3(a)(5a) provides that publishing a photograph, film or recording of a person that focuses on their sexuality, in circumstances in which the publication is liable to humiliate or degrade them and without their consent, constitutes sexual harassment. The section expressly adds that the term includes editing or compositing, provided that the person can be identified in the circumstances. That wording, enacted years before the term deepfake entered common usage, applies to the new technology almost by accident. The penalty reaches five years' imprisonment, alongside a civil cause of action for damages of up to ILS 120,000 without proof of loss for each publication.

In late January 2026, against the background of those same events, the Israeli Privacy Protection Authority published a warning that the use of technological means to depict a person nude or minimally clothed without their consent may constitute a criminal offence. The Authority identified four possible avenues of exposure: invasion of privacy, sexual harassment, unlawful processing of personal data, and potential liability of the platforms and companies themselves rather than only of the user who entered the prompt. In parallel, a Penal Law amendment bill on deepfake offences, 5786-2025, is making its way through the Knesset, seeking to establish graduated penalties and a rapid takedown duty on platforms, together with a separate bill that would allow the identity of anonymous distributors to be revealed, drawing on the existing mechanism in the Copyright Act of 2007.

Israeli law, as it currently stands, treats publication and distribution as the offence. It does not address the tool that created the material, and imposes no independent duty of care on a party that has placed such a capability on the market. The new European regime takes precisely the opposite approach: it is not concerned with the distributor, but with whether the system was placed on the market with adequate preventive measures.

6. Two liability regimes, and what they mean for an Israeli company

Practical implications for Israeli technology companies

The Artificial Intelligence Act has extraterritorial reach. It applies to providers, deployers, importers and distributors placing systems on the Union market, and to any company whose system outputs are used within the Union, regardless of its place of incorporation or physical presence in Europe.

Israeli companies that develop, integrate or offer media generators and image editing tools can therefore no longer rely on terms of use prohibiting unlawful use by end users. From December 2026 the compliance burden moves from legal damage control to product architecture, and it requires engineering documentation, documented adversarial testing of the red teaming kind, and proof of technical resilience established in advance.