AI Act 2 December 2026: the dates that bind deployers

From 2 December 2026, the EU AI Act (Regulation (EU) 2024/1689) bans AI that generates non-consensual sexual deepfakes or child sexual abuse material, and requires providers of generative AI systems placed on the market before 2 August 2026 to mark their output in a machine-readable format. Both changes come from Regulation (EU) 2026/1744, the Digital Omnibus on AI, which also moved the high-risk rules to 2 December 2027 (Annex III) and 2 August 2028 (Annex I). Deployers’ own Article 50 duties did not move: they have applied since 2 August 2026.

The Omnibus moved less than the “AI Act delayed” headlines suggest. This post reads the dates from a deployer’s side, including self-hosted assistants and Swiss organisations, and ends with what to do, as of 22 September 2026.

Was the AI Act delayed? What the Omnibus moved

The Omnibus entered into force on 27 July 2026. It moved the Annex III high-risk rules from 2 August 2026 to 2 December 2027, and the Annex I rules from 2 August 2027 to 2 August 2028 (new Article 113, third paragraph, point (c)). Recital 40 cites delays in standards and in setting up national authorities. Article 50 and the enforcement powers kept 2 August 2026, except that providers of generative systems already on the market gained four months for marking. Two other changes reach deployers:

• AI literacy is rewritten, not removed. Since 27 July 2026, Article 4 asks providers and deployers to “take measures to support” AI literacy, instead of ensuring “a sufficient level”.
• The fundamental rights impact assessment can cross-refer to the GDPR DPIA (amended Article 27(4)).

As of 22 September 2026 we have found no proposal to move the dates again.

Which AI Act dates bind a deployer?

Date	What applies	Who it binds	Source
2 February 2025	Prohibitions (Article 5) and AI literacy (Article 4)	Both	Article 113
2 August 2026	Article 50: 50(3) and 50(4) for deployers, 50(1) and 50(2) for providers; enforcement powers start	Both	Article 113, IP/26/1714
2 December 2026	Marking (Article 50(2)) for generative systems already on the market	Providers	Article 111(4)
2 December 2026	New prohibitions, Article 5(1)(ba) and (bb)	Both	Article 113(3)(a)
2 December 2027	Annex III high-risk rules, including Articles 26 and 27	Both	Article 113(3)(c)(i)
2 August 2028	Annex I high-risk rules	Both	Article 113(3)(c)(ii)
2 August 2030	Legacy high-risk systems intended for public authorities	Both	Article 111(2)

Fines can reach EUR 15 million or 3% of worldwide turnover for Article 50 breaches and EUR 35 million or 7% for Article 5, whichever is higher. SMEs get the lower figure for both (Article 99(6)); small mid-caps get it for Article 50 fines but not for Article 5 fines (new Article 99(6a)).

National market surveillance authorities enforce Article 50; the AI Office does so where the provider also supplies the underlying general-purpose model. The Commission’s release of 31 July 2026 warns that enforcement “will also depend on” Member States designating their authorities properly. Its list of contact points, last updated on 7 September 2026, names no authority for six Member States and marks twelve more, including France and Germany, as pending final adoption.

What changes under the AI Act on 2 December 2026?

Marking. New Article 111(4) gives providers of generative systems placed on the market before 2 August 2026, text generators included, until 2 December 2026 to comply with Article 50(2): a four-month transition (recital 38). Systems placed on the market later had to comply at once. Recital 133 of the AI Act gives examples of marking techniques: watermarks, metadata, cryptographic proof of provenance, logging and fingerprints, alone or combined.

Prohibitions. New Article 5(1)(ba) bans supplying or using AI that generates or manipulates realistic images, video or audio of an identifiable person’s intimate parts or sexual activity without their “explicit consent”. Point (bb) covers child sexual abuse material. Under new Article 5(1a), a deployer’s use “is only prohibited where the deployer uses the system for the purpose of generating or manipulating such material”; a provider is caught where that is the intended purpose, or a reasonably foreseeable and reproducible outcome of a system without adequate safeguards.

What does Article 50 ask of deployers?

Three paragraphs of Article 50 have bound deployers since 2 August 2026, with no transition. The Commission’s non-binding Article 50 guidelines (C(2026) 7702 final, PDF) explain them. The Commission approved their content on 20 July 2026 and formally adopted them on 5 November 2026; the paragraph numbers cited below follow the adopted text.

• Emotion recognition and biometric categorisation (50(3)). Tell the people exposed that the system is operating. At work and in education, Article 5(1)(f) has banned emotion recognition since 2 February 2025, except for medical or safety reasons.
• Deepfakes and public-interest text (50(4)). Disclose deepfake images, audio and video, and AI text published “with the purpose of informing the public on matters of public interest”, unless a human reviewed it and someone holds editorial responsibility. Internal texts and professional correspondence are not “published”, according to the guidelines.
• Presentation (50(5)). Clear, distinguishable and accessible, at the latest at the first interaction or exposure.

The Commission’s Article 50 Q&A adds that deployers “cannot simply rely on the machine-readable marking” of the provider for their own deepfake disclosure, and that content generated before 2 August 2026 “does not need to be labelled retroactively”, though the Commission encourages deployers to do so.

Section 2 of the Code of Practice on Transparency of AI-generated Content covers deployers’ labelling, with free EU icons; recital 41 of the Omnibus says such codes “do not grant a presumption of conformity”.

Does self-hosting an open-weight model make you a provider?

The Act applies by role and use, not by where the servers are. Neither self-hosting nor an open-source licence takes you out of it: Article 2(12) excludes open-source systems, but not high-risk ones or those that fall under Article 5 or 50.

Under Article 3(3), a provider develops an AI system, or has one developed, and puts it into service “under its own name or trademark”; putting into service includes supply “for own use in the Union” (Article 3(11)). On the guidelines’ reading, an organisation that develops a chatbot in-house for its own use, under its own name, is its provider (para 11) and can be its deployer too (para 15). A contractor building or running it does not move the roles: you remain the deployer when third parties operate it on your behalf and under your control (para 14). Assemble an assistant from Open WebUI, vLLM and an open-weight model, run it under your own name, and the guidelines point to you as provider, not to the model’s publisher (para 24). That is a legal assessment; write yours down.

Article 50(1), the chatbot notice. A public-facing assistant must tell people they are dealing with AI. For internal use, the guidelines treat “an internal employee-facing assistant for properly trained staff who is AI literate” as obvious enough to need no notice, but limit the exception to cases with “almost no doubt”. If you rely on it, record why. Otherwise show a notice; Open WebUI lets administrators display banners on the new-chat screen.

Article 50(2), marking. Generated text, images, audio and video are in scope; short sequences, source code and machine-to-machine output are not (para 68). The narrow business-to-business exemption (para 87) covers only strictly technical output for a limited, pre-defined group of professionals, not meant to be shared; an assistant drafting letters for external readers falls outside it. Timing is open: Article 111(4) covers systems “placed on the market” before 2 August 2026, while the guidelines (para 153) add “or put into service”. If your in-house system predates 2 August 2026, ask counsel whether 2 December 2026 is your date.

What self-hosting gives you: the evidence

What self-hosting does give you is evidence, kept where you control it: model versions, prompts, answers and logs on your own machines. With a hosted assistant, what the logs contain, how long they are kept and which model version answers are set by the vendor, and the model behind a product name can change on the vendor’s schedule. With open-weight models you run yourself, the model changes only when you approve it, and the record is yours to keep for as long as you need it.

From 2 December 2027 that matters for Annex III uses such as CV filtering (point 4(a)), credit scoring (5(b)) or life and health insurance pricing (5(c)). Their deployers must keep the system’s logs for at least six months “to the extent such logs are under their control” (Article 26(6)); self-hosted, they are. An internal drafting and search assistant is rarely high-risk; for its data flows, see what stays inside your perimeter.

Does the AI Act apply to Swiss organisations?

Under Article 2(1)(c), the Act reaches providers and deployers in a third country “where the output produced by the AI system is used in the Union”. The guidelines (para 13) add a foreseeability test: a deployer outside the EU is bound by Article 50(3) and 50(4) where it “foresees dissemination and use of the AI outputs in the Union”, not where content reaches the EU “through channels that are unforeseeable and outside their control”. A Swiss chatbot aimed at EU customers, or one assistant shared with EU subsidiaries, should be treated as in scope.

The Federal Data Protection and Information Commissioner stated on 9 November 2023 that the data protection act applies directly to AI, that people have a right to know whether they are corresponding with a machine, and that deepfakes of identifiable people must be clearly indicated. See our nLPD page.

Switzerland has no AI act yet. On 12 February 2025 the Federal Council decided to ratify the Council of Europe AI Convention and to implement it mainly through sector-specific changes to existing law. On 28 October 2026 it took note of the state of the work: the Federal Office of Justice is drafting the binding measures, and the consultation, once expected by the end of 2026, is now planned for spring 2027 (OFCOM). The summary of the German release is ours.

What to do, and by when

1. Now: list every AI system and your role in it. The Article 50 duties it maps have applied since 2 August 2026. Include in-house assistants built from open-source parts, and note whether output is used in the EU.
2. Now: a notice on interactive systems you provide. Public-facing chatbots need one at the first interaction. For internal assistants, show one or document why the “obvious” exception applies.
3. Now: a labelling rule for deepfakes and public-interest text. Name the editor who reviews AI-assisted publications.
4. Now: record your AI literacy measures under the new Article 4: who was trained, on what, and when.
5. Now: decide where the evidence lives. For each system, note who holds the prompts, answers, logs and model versions, and for how long. Where a vendor holds them, ask in writing what you can export.
6. By 2 December 2026: marking. If you provide a generative system placed on the market before 2 August 2026 (or put into service, on the guidelines’ reading), put marking and detection in place. If a supplier is the provider, ask it in writing for its Article 50(2) plan and whether it signed Section 1 of the Code.
7. By 2 December 2026: close off prohibited uses. Add Article 5(1)(ba) and (bb) to your acceptable-use policy and review every image, video and audio generation feature against them.
8. Before 2 December 2027: prepare Annex III uses. Human oversight, logs kept at least six months, workers’ representatives and affected workers informed before any use at work, and an Article 27 impact assessment where required, cross-referring to your DPIA. See what we supply for a DPIA and our GDPR page.
9. By 2 August 2030: providers and deployers of legacy high-risk systems intended for use by public authorities must comply, under amended Article 111(2).

How Pilae fits

The simplest way to answer the Act’s questions about your AI is to run it where you can see it. Pilae operates private AI: Open WebUI, LibreChat or Dify, with open-weight models such as Llama, Mistral, Qwen or gpt-oss, on your own GPU servers or on dedicated GPUs in Zurich or another Pilae Cloud region. Every prompt, answer and uploaded file is processed and stored on machines you control.

• You know which model answered. Models are listed per environment, and changing or adding one is a planned change: the Pilae Agent writes it with a rollback, waits for your approval and records it.
• The record is yours. Prompts, answers and sources are logged in your own database. Langfuse, where you run it, traces each model call, and the audit trail exports to your SIEM.
• Nothing leaves by default. External models are off by default; routing selected users to one is your decision.
• Your directory decides access. Staff sign in with their company account, and groups decide who sees which models and assistants.

None of this makes you compliant, and no provider can do that for you: which of your uses fall where is your decision. What it gives you is the record a deployer is asked to keep, on your own machines. To map your systems and roles, or to move an assistant off a public service, talk to an engineer.

This post is general information, not legal advice. The binding texts are Regulation (EU) 2024/1689 and Regulation (EU) 2026/1744 on EUR-Lex.