The European Union will start enforcing Article 50 transparency obligations under its AI Act on Aug. 2, 2026 across all 27 member states. According to The Register, the rule can be reduced to a simple standard: AI systems must identify themselves as AI.
Article 50 sets out two main requirements. First, people must know who they are interacting with. For AI systems that engage directly with humans, including chatbots, virtual assistants, and AI agents that can complete workflows on their own, providers must make clear that the other side is a machine. Second, users and software must be able to identify synthetic output. Audio, images, video, and text that are generated or modified by AI must carry machine-readable markings showing that the material is synthetic.
European Commission Executive Vice-President Henna Virkkunen said, while explaining the related guidance, that the measures “support providers and deployers in fulfilling their obligations under the AI Act, while at the same time ensuring citizens are informed when they are interacting with AI.” On the same day, the Commission also obtained enforcement powers over general-purpose AI models.
Disclosure must happen at the first interaction
Article 50(5) fixes the timing of the disclosure requirement. The information must be provided “in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure.” In practical terms, that means a company cannot treat a line buried in terms of service as sufficient notice.
There are only two exemptions. One applies when it is already obvious to a reasonably well-informed natural person that the system is AI. The other applies to systems used by law enforcement authorities to detect, prevent, investigate, or prosecute criminal offences. The first exemption may sound broad, but the burden of proof remains with the provider. Relying on the assumption that users should have figured it out is not a stable compliance approach.
There is also a carve-out on the content-marking side. Assistive editing tools such as spelling and grammar correction are excluded if they do not materially alter the original input.
Three categories of obligations under Article 50
Read across the text, the operational requirements fall into three groups.
- Interaction disclosure under Article 50(1): Providers of chatbots, virtual assistants, and AI agents must make sure users know at the first interaction that they are dealing with AI.
- Synthetic content marking under Article 50(2): Providers of generative AI systems must add machine-readable markers to audio, images, video, and text outputs so they can be detected as artificially generated or manipulated.
- Disclosure for deepfakes and public-interest text under Article 50(4): Deployers must disclose deepfake content. AI-generated text published to inform the public on matters of public interest must also be disclosed, unless it has undergone human review and a person has taken editorial responsibility. Artistic, creative, satirical, and fictional works are subject to a lighter form of disclosure.
Penalties can reach €15 million or 3% of annual global turnover
The fine structure is set out in Article 99(4). Violations of Article 50 can lead to penalties of up to €15 million or 3% of a company’s total worldwide annual turnover in the previous financial year, whichever amount is higher.
The framework is softer for smaller entities. Small and medium-sized enterprises and startups get a lower-threshold buffer based on whichever amount is lower. EU institutions themselves are capped at €750,000. The bodies that will issue penalties are the market surveillance authorities in each member state, not Brussels directly.
As the report noted, €15 million may be a manageable fine for a large platform, but for a 20-person startup it could be existential.
Non-EU providers are covered if the output is used in Europe
One of the easiest mistakes for teams in Taiwan and elsewhere in Asia is to misread the scope. Article 50 does not turn on where a company is incorporated. It turns on where the output goes.
For generative AI providers in third countries, the marking obligation still applies if their system outputs are used within the EU. The practical test is not whether a business has a European subsidiary. It is whether it has European users.
Voluntary code is already in motion ahead of enforcement
The timeline is moving. The European Commission finalized its Code of Practice on Transparency for General-Purpose AI-Generated Content in June. The code is voluntary, but signing it can be used as supporting evidence of compliance with Article 50(2) and 50(4).
The code was assessed as adequate on July 8 and 9. Companies seeking to join the first group of signatories must submit their forms by 18:00 Central European Time on July 22.
The shift here is straightforward. The EU is turning an industry norm — that AI should say it is AI — into an enforceable legal obligation with penalties attached. After Aug. 2, saying a company did not realize it fell within scope will not serve as an effective defense, according to the report.
Key questions
When does Article 50 of the EU AI Act become enforceable?
It becomes enforceable on Aug. 2, 2026 in all 27 member states, with national market surveillance authorities responsible for imposing penalties. On the same day, the European Commission also gains enforcement powers over general-purpose AI models.
Does all AI-generated content need to be labeled?
As a rule, AI-generated images, audio, video, and text must carry machine-readable markers, and deepfakes must be clearly disclosed. Exceptions include assistive tools such as spelling and grammar correction that do not materially alter the original text, as well as AI-generated text that has undergone human review and carries human editorial responsibility.