The AI Act nudification ban applies from 2 December 2026. From that day, the EU Artificial Intelligence Act prohibits placing on the market, putting into service or using AI systems that generate realistic intimate imagery of an identifiable person without consent (the "nudifier" or "undress" apps), and AI systems that generate child sexual abuse material. Both prohibitions were inserted into Article 5 by the Digital Omnibus on AI, Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force since 27 July 2026. The same date ends a grace period: generative AI systems already on the market before 2 August 2026 must comply with the machine-readable marking duty in Article 50(2) by 2 December 2026.

Two points matter for readers outside the EU. The AI Act reaches providers wherever they are established, so a Turkish or Korean image app with EU users is within reach. And the EU has chosen a different tool from Türkiye and Korea, which mainly punish the person who makes or spreads the image. The EU will now also prohibit the product.

What changes on 2 December 2026?

  • Two new prohibited practices. Points (ba) and (bb) of Article 5(1), first subparagraph, and the new Article 5(1a) and (1b) start to apply. Amended Article 113 says Chapters I and II have applied since 2 February 2025 "with the exception of" these provisions, "which shall apply from 2 December 2026".
  • The end of the marking grace period. New Article 111(4) gives providers of generative systems "placed on the market before 2 August 2026" until 2 December 2026 to comply with Article 50(2).
  • No grandfathering for the bans. The transitional rules in Article 111 are expressly "without prejudice to the application of Article 5". An app that has been on sale for years gets no extra time.

Recital 10 refers to "the deployment and widespread use of AI systems generating non-consensual intimate images, videos, audio and similar material" and child sexual abuse material; recital 11 notes that such tools are "often described as 'nudification' applications". The amendment came in the year the Commission opened a Digital Services Act investigation into X (26 January 2026) over risks linked to Grok, including manipulated sexually explicit images. That separate proceeding establishes no breach.

What exactly does the AI Act nudification ban prohibit?

New point (ba) prohibits:

"the placing on the market, the putting into service or the use of an AI system that generates or manipulates realistic images, videos, audio or similar material of an identifiable natural person's intimate parts, or of an identifiable natural person engaged in sexually explicit activities, without that person's freely-given, specific, informed, unambiguous and explicit consent for that generation or manipulation".

Point (bb) prohibits the same acts for a system that generates or manipulates "material or performance within the meaning of Article 2, points (c) and (e), of Directive 2011/93/EU, except where a 'without right' defence applies under national law". Those are the definitions of child sexual abuse material (the Directive still says "child pornography") and of a live sexual performance involving a child. Directive 2011/93/EU already covers "realistic images of a child" and material depicting "any person appearing to be a child", so synthetic material is caught whether or not a real child was involved.

Three drafting points stand out. The object is the system: Article 5 does not itself make an image unlawful to hold or share, which remains a matter for criminal law. Audio counts, so a voice clone producing explicit audio of an identifiable person is within the text. And consent follows the GDPR standard; recital 12 adds that systems intended for such generation need means "enabling the reliable collection and demonstration of consent of the depicted person".

Who does the AI Act nudification ban catch?

New Article 5(1a) splits the ban between providers and deployers. (For the Act's overall structure, see the site's note on the EU AI Act.)

Providers: intended purpose or foreseeable outcome

Placing on the market or putting into service is prohibited only where generating the material "is the intended purpose of the AI system", or where:

"the system's design, training, architecture, capabilities or user-facing functionalities make that generation or manipulation a reasonably foreseeable and reproducible outcome, without requiring significant technical modification, and the system does not have reasonable and adequate technical safety measures and other safeguards to reliably prevent that generation or manipulation, taking into account reasonably foreseeable misuse, and to correct observed or reported misuse".

The first limb captures the dedicated nudifier. Under Article 3(12), intended purpose includes what the provider says in "promotional or sales materials and statements". An app advertised as able to undress any photo has declared its purpose, whatever its terms of service say.

The second limb is where general-purpose image, video and voice models sit, and its conditions are cumulative: the outcome must be foreseeable, reproducible and achievable without significant modification, and adequate safeguards must be missing. Recital 12 lists the expected measures: "data cleaning, refusal training, safe prompt design and output controls, runtime prompt guardrails, content classification and filtering mechanisms, usage restrictions, abuse detection mechanisms, and notice and action mechanisms". They are adequate if they "align with the state-of-the-art measures and demonstrably prevent or sufficiently reduce" the likelihood of such output.

A single jailbreak by a determined user is not a breach; "reproducible" points away from one-off failures. A provider without working filters, or that ignores reported misuse, carries the risk. Corrective measures must be reasonable for the system, "including its release and distribution strategy (such as open-source releases)". That acknowledges that open weights cannot be patched like a hosted service; it is not an exemption.

Deployers: a purpose test

Use is prohibited "only where the deployer uses the system for the purpose of generating or manipulating such material or performance" (Article 5(1a)(b)). Recital 12 says this includes circumventing safeguards, but not "accidental generation or manipulation".

A definitional point matters. Under Article 3(4), a deployer is someone using an AI system under its authority "except where the AI system is used in the course of a personal non-professional activity", and under Article 2(10) the Regulation does not apply to obligations of deployers who are natural persons using AI systems in a purely personal non-professional activity. On the Act's own definitions, a person using a nudifier at home is unlikely to be a deployer. The use ban bites mainly on businesses and organisations; individuals remain exposed to criminal law.

Where the line falls

  • Realism and identifiability. Recital 12 limits point (ba) to realistic depictions of intimate parts or sexually explicit activity, meaning the person's face, voice or body shown "in a credible real-life manner". Cartoonish or physically impossible images, and nudes of unidentifiable people, fall outside.
  • Manipulation must add exposure. Under Article 5(1b), editing that "does not increase the exposure of any depicted intimate parts or alter the nature of any depicted sexually explicit activities" is not manipulation. Brightening an image is outside; "undressing" a clothed photograph is inside.
  • Law enforcement and red-teaming. Under recital 13, the "without right" defence in point (bb) covers authorities generating material for criminal proceedings, and red-teaming that tests compliance with the prohibition.

Penalties, enforcement and reach beyond the EU

Article 5 breaches fall in the highest tier. Article 99(3) allows fines of up to EUR 35 million or 7% of total worldwide annual turnover, whichever is higher; for SMEs and start-ups, Article 99(6) applies whichever is lower. Enforcement runs through market surveillance: under Article 74(1), Regulation (EU) 2019/1020 applies, so national authorities can require a prohibited system to be withdrawn. Under amended Article 75(1), however, the AI Office is exclusively competent for AI systems based on a general-purpose AI model developed by the same provider or undertaking, and for systems that constitute or are integrated into a very large online platform or search engine designated under the DSA. Recital 15 requires respect for ne bis in idem where criminal law also punishes the conduct.

Article 2(1)(a) covers providers "irrespective of whether those providers are established or located within the Union or in a third country", and Article 2(1)(c) reaches cases "where the output produced by the AI system is used in the Union". A developer in Istanbul or Seoul whose app can be downloaded in the EU should assume the ban applies.

The second deadline: Article 50(2) marking for systems already on the market

Article 50(2) requires providers of systems "generating synthetic audio, image, video or text content" to ensure outputs "are marked in a machine-readable format and detectable as artificially generated or manipulated", using solutions that are effective, interoperable, robust and reliable as far as technically feasible. Article 50 has applied since 2 August 2026; systems placed on the market before that date have until 2 December 2026. A breach falls in the EUR 15 million or 3% tier of Article 99(4).

The guidelines allow a single technique or a combination, drawing on recital 133: watermarks, metadata, cryptographic provenance, logging and fingerprints. Providers may use an upstream model's marking but remain responsible for its compliance. The site's EU AI Act Article 50 kit and the note on marking AI-generated content cover the mechanics.

Marking does not make a prohibited output lawful, but it helps platforms detect generated intimate imagery and helps providers evidence the misuse response Article 5(1a) expects.

How does the ban fit with Directive 2024/1385?

The AI Act regulates the tool; Directive (EU) 2024/1385 on violence against women deals with the person. Its Article 5(1) requires Member States to criminalise making intimate images accessible to the public by ICT without consent, and "producing, manipulating or altering and subsequently making accessible to the public" material making it appear that a person is engaged in sexually explicit activities, where likely to cause serious harm, plus threats to do so. Article 23 requires prompt removal, including binding orders to hosting service providers. Transposition is due by 14 June 2027.

Recital 14 of the Omnibus calls the new bans "aligned with" both Directives. A gap remains: the Directive is triggered by publication, the AI Act by supply of the system. Private creation by an individual, never published, is reached by neither unless national law goes further. Korean law does.

Is a deepfake a crime in Türkiye? TCK Articles 134, 226, 105, 125 and 136

Türkiye has no offence specific to AI-generated intimate imagery. The consolidated Turkish Penal Code (TCK, Law No. 5237), which incorporates amendments up to Law No. 7593 of 8 August 2026, contains no deepfake provision. Prosecutors work with general offences, each fitting only part of the conduct.

  • Article 226(3): the strongest tool for child material. Using "children, representative child images or persons appearing to be children" (çocukları, temsili çocuk görüntülerini veya çocuk gibi görünen kişileri) in producing obscene products carries five to ten years' imprisonment and a judicial fine of up to 5,000 days; importing, reproducing, selling, storing, possessing or making such products available carries two to five years. "Representative child images" was added in 2016 (Law No. 6698), so wholly synthetic material falls within the text.
  • Article 134 (privacy). Violating private life carries one to three years, doubled where images or sounds are recorded (134(1)); unlawfully disclosing images "relating to private life" carries two to five years (134(2)). Whether an image of something that never happened "relates to" private life is a genuine doctrinal question. The argument that the person's intimate sphere is attacked is strong, but the fit is not textual.
  • Article 105 (sexual harassment). Sending a synthetic intimate image to the person depicted can be sexual harassment, prosecuted on complaint: three months to two years or a judicial fine (six months to three years if the target is a child), increased by half under 105(2)(d) where "postal or electronic communication" tools are used.
  • Article 125 (insult). A degrading image can attack honour and dignity; 125(2) covers a "visual message" addressed to the person, and 125(4) aggravates public insult.
  • Article 136 (disseminating personal data). Faces are personal data, and data on sexual life is a special category under Article 6 KVKK. Spreading an image linking an identifiable face to sexual content may be argued as unlawful dissemination (two to four years), though whether a fabrication "disseminates" data about the person is open to argument.

Threatening to release such an image to obtain a benefit can be blackmail under Article 107(2). Article 20(2) TCK excludes criminal penalties for legal persons, so under criminal law a company running a nudifier service can face only security measures, such as those Article 226(6) provides for obscenity offences. There is no equivalent of the AI Act's product ban. According to press reports, an MHP deputy tabled an AI bill in September 2025 including six-hour removal of content violating personality rights and fines of up to TRY 10 million; no such provision appears in the consolidated TCK or Law No. 5651 as of this writing.

Removing the images in Türkiye: Law No. 5651 and the civil route

Following amendments by Law No. 7590 of 24 July 2026 (Official Gazette, 31 July 2026), the consolidated text of Law No. 5651 defines the "Presidency" (Başkanlık) as the Cyber Security Presidency (Siber Güvenlik Başkanlığı), which now exercises the functions previously held by BTK under the Law.

  • Article 9/A (privacy). A person claiming that content violates their private life may apply directly to the Presidency with the exact URL. Access providers must block within four hours; the applicant must bring the request before a criminal judgeship of peace within 24 hours, and the judge decides within 48 hours, failing which the block lapses. This is the fastest route for intimate imagery.
  • Article 8 (catalogue offences). Content giving sufficient suspicion of listed offences, including child sexual abuse (TCK 103(1)) and obscenity (TCK 226), can be removed or blocked by a judge, by a prosecutor in urgent cases, or blocked by the Presidency of its own motion. Synthetic child material falls here.
  • Article 9 has gone. The general personality-rights route was annulled by the Constitutional Court (11 October 2023, E.2020/76, K.2023/172) with effect from 10 October 2024.

Social networks with over one million daily accesses from Türkiye must answer Article 9/A applications within 48 hours, with reasons for refusals (Additional Article 4(3)). The civil route remains: Articles 24 and 25 of the Civil Code allow actions to prevent, stop or declare an attack on personality rights unlawful, and Article 58 of the Code of Obligations provides non-pecuniary damages. A 2026 ruling on AI voice cloning sent such claims to the general civil courts, and blocking powers have been used against AI output before, as in the 2025 order blocking access to Grok posts.

Korea: punishing the person, not the product

Korea has no AI product ban of the EU kind, but it criminalises the conduct from creation. Article 14-2 of the Act on Special Cases Concerning the Punishment of Sexual Crimes punishes editing, synthesising or processing a person's face, body or voice, against their will, into a form that may cause sexual desire or shame, with up to seven years or a fine of up to KRW 50 million, without any distribution requirement. Distribution for profit over networks carries at least three years, and since the amendment of 16 October 2024 possessing, purchasing, storing or viewing such material carries up to three years or KRW 30 million.

The site has covered this: see what Korean law actually punishes, the Supreme Court's test for deepfakes of real minors and the move to criminalise AI sexual content of people who don't exist. On the product side Korea relies on labelling: Article 31 of the AI Basic Act requires AI business operators to label generative outputs, and to notify or label realistic virtual sound, images or video in a way users can clearly recognise.

Korean law reaches the individual at home; the EU rule reaches the provider of the tool. Spain's data protection authority has shown a third route, fining the sharing of AI-generated nude images under the GDPR.

What should platforms, app developers and app stores do now?

Developers of image, video and voice tools

  1. Classify every generation and editing feature against Article 5(1a): is prohibited output an intended purpose, or a foreseeable and reproducible outcome?
  2. Clean up marketing. Remove "undress", "x-ray" and similar claims from listings, ads and influencer briefs; intended purpose is read from promotion.
  3. Document layered safeguards from the recital 12 list, including limits on editing uploaded photos of real people, and record test results.
  4. Keep red-teaming documented, limited to testing compliance with the prohibition (the activity recital 13 protects) and run in controlled environments.
  5. Run a misuse loop: who receives reports, how fast classifiers are updated, when accounts are suspended.
  6. For consent-based products, keep a verifiable, GDPR-compliant consent record for each person depicted.

Providers with legacy generative systems

  1. Confirm the system was on the EU market before 2 August 2026; otherwise there is no grace period.
  2. Choose a marking approach (metadata, watermark, provenance, fingerprinting or a combination), test robustness and provide detection.
  3. Decide whether to sign the Code of Practice, or document an alternative route.
  4. If relying on upstream marking, obtain contractual commitments and test evidence.

Platforms and app stores

Article 5 targets placing on the market, putting into service and use; a store that only distributes a third-party app is not obviously within those verbs. But market surveillance authorities can act against prohibited systems, hosting services face removal orders under the Digital Services Act and, once transposed, Article 23 of Directive 2024/1385, and a nudifier listing will be hard to defend. Stores should reference points (ba) and (bb) in developer policies, screen listings for nudification claims and keep takedown records; platforms serving Turkish users should map processes to Law No. 5651 timelines.

What to watch next

  • Guidance. I have not found Commission guidance specific to points (ba) and (bb); the earlier guidelines on prohibited practices predate them.
  • Enforcement. How national authorities apply the "reasonably foreseeable and reproducible" limb to general-purpose models will define the ban's real scope.
  • Türkiye. Whether pending deepfake bills move, and how the Cyber Security Presidency applies Article 9/A.
  • Korea. The proposal on sexual content of non-existent persons and enforcement of Article 31 of the AI Basic Act.

Frequently asked questions

Does the EU AI Act ban nudify apps?

Yes, from 2 December 2026. New Article 5(1)(ba) prohibits placing on the market, putting into service or using AI systems that generate realistic intimate images, videos or audio of an identifiable person without explicit consent. Dedicated nudifier apps are caught through their intended purpose. General-purpose tools are caught only where such output is foreseeable and reproducible and adequate safeguards are missing.

When does the AI Act nudification ban apply?

From 2 December 2026, under Article 113 as amended by Regulation (EU) 2026/1744. The Omnibus has been in force since 27 July 2026, but points (ba) and (bb) and paragraphs (1a) and (1b) of Article 5 apply only from 2 December. There is no transitional period for systems already on the market.

Is deepfake pornography illegal in the EU?

EU law now tackles it at two levels. The AI Act prohibits the tools from 2 December 2026. Directive 2024/1385 requires Member States, by 14 June 2027, to criminalise making non-consensual intimate images, including manipulated ones, accessible to the public where this is likely to cause serious harm. Whether purely private creation is a crime depends on national law.

Is creating a deepfake a crime in Turkey?

The Turkish Penal Code has no specific deepfake offence. Depending on the facts, TCK Articles 134 (privacy), 105 (sexual harassment), 125 (insult), 136 (personal data) and 107 (blackmail) may apply. Synthetic sexual images of children fall under Article 226(3), which expressly covers representative child images and carries five to ten years' imprisonment.

What is the 2 December 2026 deadline under Article 50(2)?

Providers of generative AI systems placed on the EU market before 2 August 2026 must ensure by 2 December 2026 that outputs are marked in a machine-readable format and detectable as AI-generated. Systems placed on the market from 2 August 2026 have had to comply from the start, and content generated before 2 August 2026 need not be marked retroactively.

Burhan Doğuş Ayparlar's View

The new prohibition is narrower than the headlines suggest, and better for it. It does not turn every user into an offender or ban image generation. It imposes a design obligation: a provider whose model can predictably and repeatedly be made to undress a real person, and who has not built and maintained reasonable defences, is selling a prohibited system. The "reasonably foreseeable and reproducible" test is in substance a product-safety standard for generative models. It asks what the provider knew and did, not what one user managed to do, which is the right question.

The limit is equally clear. The person using a nudifier at home is, on the Act's own definitions, very probably outside the AI Act, and the Directive focuses on publication; Korea closes that gap by punishing creation and possession.

For Türkiye the comparison is uncomfortable. Article 226(3) is well drafted for material depicting children, and its reference to representative child images has aged well. For adults, prosecutors must choose among offences written for other harms: Article 134 for intrusions into real private life, Article 125 for honour, Article 136 for data. None names the act of fabricating someone's body. I would favour a targeted offence, modelled on Article 5 of Directive 2024/1385 and Article 14-2 of the Korean Act, covering the non-consensual creation of sexual imagery of an identifiable person and its distribution or threatened distribution. Removal should stay under prompt judicial control on the Article 9/A model rather than rest on broad administrative deadlines. Speed alone is not enough: Article 9, under which judges themselves ordered blocking, was still struck down by the Constitutional Court on freedom of expression grounds.

For Turkish and Korean developers serving Europe, the EU standard is now the design baseline. What will matter in an investigation is a record showing that safeguards were tested, that misuse reports were acted on, and that marketing never promised what the law forbids.

This article is general information, not legal advice. It is based on Regulation (EU) 2026/1744 (OJ, 24 July 2026), the consolidated text of Regulation (EU) 2024/1689 of 27 July 2026, Directives 2011/93/EU and (EU) 2024/1385, the Commission's Code of Practice (10 June 2026), adequacy opinion (8 July 2026) and Article 50 guidelines (20 July 2026), the consolidated texts of Laws No. 5237, 5651, 4721, 6098 and 6698 on mevzuat.gov.tr, and Korean statutes on law.go.kr, all consulted on 20 October 2026. The Turkish bill is described from press reports. How authorities and courts will interpret the new prohibitions, and how Turkish courts will apply Articles 134 and 136 TCK to fabricated images, is uncertain. Specific situations require individual assessment.