If your company has put a chatbot, assistant or content generator into production on top of someone else's model, 2 December 2026 probably concerns you more than you think. On that date the transition period for content marking under Article 50(2) of the AI Act (Regulation (EU) 2024/1689) ends, and two new bans introduced by Regulation (EU) 2026/1744, the Digital Omnibus, start to apply.
This is not a deadline for large model providers only: the duty follows the system, not the model.
What applies on 2 December 2026
| Change | Who it affects | Maximum fine |
|---|---|---|
| Marking and detectability of outputs (Art. 50(2)) | Providers of generative systems placed on the market before 2 August 2026 | €15 million or 3% of worldwide annual turnover |
| Ban on generating non-consensual intimate material and child sexual abuse material (Art. 5) | Providers and deployers, within the limits set out below | €35 million or 7% of worldwide annual turnover |
For SMEs and start-ups the lower of the fixed amount and the percentage applies (Art. 99(6) AI Act). In Italy, supervision generally falls to ACN, under Legislative Decree 179/2026.
Three things 2 December does not change:
- The other Article 50 duties, such as telling people they are talking to an AI and disclosing deepfakes, have applied since 2 August 2026, including to existing systems.
- A system placed on the market from 2 August 2026 has no transition period.
- High-risk systems remain on a separate timetable: 2 December 2027 and 2 August 2028.
Are you the provider? The test if you use an LLM
Article 50 concerns AI systems, not models. Under Article 3(3), a provider is whoever develops a system, or has one developed, and places it on the market or puts it into service under their own name or trademark. Using someone else's model through an API does not take that role away.
| Case | Likely role | Marking (Art. 50(2)) |
|---|---|---|
| You use a commercial assistant as it is, for internal work | Deployer | On the assistant's provider |
| You sell an app under your brand that generates text or images via an API | System provider | On you |
| You run an assistant on your website that you built on an LLM | Provider and deployer | On you |
| You have an internal generator, restricted to a closed perimeter | Provider, if you put it into service under your name | To be assessed: see the exceptions below |
You can rely on marking from the model provider or on a third-party tool, but outsourcing the technology does not outsource the duty: you must check the solution and be able to show that it is adequate.
What marking really means
Article 50(2) asks for two things: outputs marked in a machine-readable format and detectable as artificially generated or manipulated. Solutions must be effective, interoperable, robust and reliable, as far as technically feasible. The Commission's guidelines of 20 July 2026 are not binding, but national authorities are expected to rely on them. They specify that cost can weigh in a proportionality assessment, but does not exempt anyone on its own.
The Code of Practice, which is voluntary and has been assessed as adequate by the Commission and the AI Board, sets out how to do it in practice:
- Images, audio, video and text in a file (PDF, Word, HTML): signed metadata plus an imperceptible watermark.
- Free-form text (such as a chat reply): an imperceptible watermark, because metadata is not workable. The Code does not require it for very short text.
- Detection: a tool that lets people verify where content came from, with understandable results.
A visible "made with AI" notice does not satisfy Article 50(2). The reverse also holds: the provider's technical marking is not enough for someone publishing a deepfake, who must disclose it visibly under Article 50(4).
Signing the Code (about 190 organisations had done so by the end of July 2026) does not bring immunity, but it offers a simpler way to demonstrate compliance. Non-signatories must show adequacy by other means and, the Commission warns, may receive more requests for information.
When marking is not needed
The exclusions are specific, and should be documented if you rely on them:
- Standard editing: grammar and spelling correction, format conversion. The guidelines include translation here. Summaries and substantive rewrites still have to be marked.
- Out-of-scope outputs: source code, configuration files, very short outputs such as single words or labels, machine-to-machine communication.
- AI agents: intermediate steps and outputs not meant for people. The duty may apply to what the user sees.
- Closed B2B uses: the guidelines allow some relief for internal or industrial uses only with a controlled perimeter, no external circulation and safeguards against foreseeable misuse. It does not apply to public or consumer-facing systems. An internal chatbot is therefore not excluded just because it is internal.
The two new bans: who should worry
From 2 December 2026, Article 5 prohibits systems that generate or manipulate:
- realistic images, video or audio of the intimate parts of an identifiable person, or showing them in sexually explicit activities, without their free, specific, informed, unambiguous and explicit consent;
- child sexual abuse material.
The two roles are treated differently:
- Providers are covered if this is the intended purpose, or if the output is reasonably foreseeable and reproducible without substantial modification and the system lacks adequate technical measures. This therefore also reaches those offering image generation or editing for entirely lawful purposes, but without filters.
- Deployers are covered only if they use the system for that purpose.
Outside the ban are, for example, virtual clothing try-ons with no intimate parts exposed or with the person's consent, and edits that do not increase exposure, such as adjusting contrast or background. The ban does not rule out criminal prosecution: in Italy, Article 612-quater of the Criminal Code on deepfakes also remains in place.
Checklist for the coming weeks
- List the generative systems you offer or use, with their market placement date: before or after 2 August 2026?
- Assign a role to each: your brand or someone else's, external market or internal use.
- Ask the model provider in writing what it marks, with which technique, whether the mark survives your pipeline and which detection tool it offers.
- Check the pipeline: compression, conversion and post-processing can strip metadata and watermarks.
- Choose the route: sign the Code or document an alternative solution.
- Record the exclusions you rely on and why they apply.
- Test the safeguards if the system handles images, video or audio: input and output filters, refusal of prohibited requests, abuse detection.
Where to start
The most useful question is also the simplest: on which systems are you a provider without knowing it? An independent gap assessment focused on Articles 50 and 5 clarifies roles, technical gaps and the evidence to keep, and tells you what to close before 2 December.