Why the AI Act’s new nudification ban affects more businesses than many think
On 29 June 2026 the Council of the EU gave its final approval to the Digital Omnibus on AI, the package of amendments to Regulation (EU) 2024/1689 that most companies know for postponing high-risk AI deadlines. Many assume the amendments are purely a matter of timelines. They are not. The same package inserts a new prohibited practice into Article 5, banning AI systems that generate or manipulate non-consensual intimate images, video or audio, so-called nudification tools, as well as child sexual abuse material. The duty falls on providers and deployers, not only on the platforms that end up hosting the results.
What does the new prohibition actually cover?
The amended Article 5 prohibits placing on the market, or putting into service, an AI system intended to generate or alter images, video, audio or similar content that depicts a real, identifiable person in a sexual or intimate context without that person’s consent, and it separately prohibits generating child sexual abuse material through any AI system. Crucially, the ban is not limited to tools built for that purpose. It also catches general-purpose image, video or audio generation systems where such output is a reasonably foreseeable and reproducible result, achievable without significant technical modification, if the provider has not built in reasonable and adequate safeguards to prevent it.
Who carries the compliance duty?
Three groups are captured. Providers who place a system on the market for the purpose of generating this content are banned outright. Providers who place a general-purpose system on the market without reasonable technical safeguards against foreseeable misuse are also in breach, even if that was never their intention. Deployers, meaning any business that uses such a system for this purpose internally or towards customers, are liable in their own right. This reaches marketing agencies, photo and video editing apps, social platforms, HR tools with photo features, and any company that integrates third-party image generation into its own product.
How does this interact with Swedish criminal law?
Sweden has criminalised the spreading of sexual images since 2018 through olaga integritetsintrång in Chapter 4, Section 6c of the Swedish Penal Code (brottsbalken). But that provision targets individuals who spread material, not the companies whose tools created it, and its reach over AI-manipulated or fully synthetic images remains legally untested. According to Swedish public broadcaster SVT, no one has been convicted under the provision for spreading AI-manipulated material in the seven years it has existed, and women’s rights organisations have called for clarification on whether fake, AI-generated content is covered at all. The new EU prohibition does not depend on that uncertainty. It creates direct, EU-wide liability for the businesses that build or deploy the underlying AI systems, regardless of how Swedish courts eventually interpret olaga integritetsintrång.
A practical example: a marketing agency’s image tool
A Swedish marketing agency offers clients a generative AI feature that lets users upload a photograph and generate stylised variations for social media campaigns. The agency never designed the tool to undress or sexualise anyone, but internal testing later shows that with minor prompt adjustments, the model can produce intimate imagery of the uploaded person without their consent. Under the amended Article 5, the agency’s exposure depends entirely on whether it built and documented safeguards, such as content filters, prompt restrictions and output monitoring, before placing the tool on the market. Discovering the vulnerability and doing nothing about it turns a foreseeable risk into a breach.
Common mistakes
The most frequent mistake is treating the ban as a large-platform problem that does not concern a smaller business offering photo or video tools to a limited customer base. Company size is irrelevant to the prohibition. Another mistake is assuming that because a tool was never intended for this purpose, the business is automatically in the clear. It is not, once misuse becomes reasonably foreseeable and no safeguards exist. Some companies also overlook deployer liability, wrongly assuming responsibility sits only with the vendor of the underlying model. Finally, several businesses are waiting for the Digital Omnibus to be published in the Official Journal before doing anything, even though building proper safeguards takes considerably longer than the transitional period allows.
Recommended actions
Start by mapping every AI system your business offers or uses that can generate or edit images, video or audio of real people, including tools embedded in third-party software. Test those systems for foreseeable misuse, not only their intended use, and document the results. Where a risk exists, add technical safeguards such as filters, watermarking and output monitoring, and record why they are considered adequate. Update acceptable use policies so staff and customers cannot use company systems to create this content, and make sure customer-facing terms reflect the same restriction. Assign clear ownership for reviewing new generative features before launch, and keep a watch on the Official Journal for formal publication of the Digital Omnibus, since the transitional period runs to 2 December 2026 regardless of when the text is formally published.
Frequently asked questions
Is this only relevant to platforms that host user-generated content?
No. The prohibition applies to any provider or deployer of an AI system capable of generating this content, regardless of whether the business operates a public platform. An internal tool, a client-facing feature or a white-labelled third-party model can all trigger liability.
Does the ban apply if our tool was never intended for this purpose?
Yes, if misuse is reasonably foreseeable and achievable without significant technical modification, and no adequate safeguards exist. Intent shields nothing once a foreseeable gap remains unaddressed, so a documented risk assessment before launch, not after a complaint arrives, is essential.
When does the prohibition take full effect?
The transitional period runs until 2 December 2026. The change takes legal effect once the Digital Omnibus is formally published in the Official Journal, expected before 2 August 2026, so current dates should be treated as provisional until then.
Conclusion
The nudification and CSAM ban is a narrow provision inside a much larger reform package, but it carries the same maximum penalties as the AI Act’s other prohibited practices, up to 35 million euro or 7 percent of global annual turnover. Businesses that build, integrate or deploy generative image, video or audio tools should treat foreseeable misuse as a design question, not an afterthought. At Lawgent, we help companies assess AI systems against the AI Act’s prohibited practices and build the safeguards and policies needed to stay compliant. Get in touch if you want to know whether your AI tools fall within scope of the new ban.
