A disclosure date, not the arrival of every rule
Europe reaches another AI Act deadline on Sunday. This one will be visible to ordinary users.
From 2 August, the European Commission's AI Office and authorities in EU member states begin enforcing the law. At the same time, Article 50 transparency rules start applying to chatbots, generative systems, deepfakes and some other uses of AI.
The simplest part is the conversation itself. If a person is interacting directly with an AI system, the provider must make that clear. A chatbot should not quietly present itself as a human operator.
The date matters, but it is easy to overstate. This is not every part of the AI Act switching on together. Rules for many high-risk systems have been pushed into 2027 and 2028.
What starts now is narrower: disclosure, marking, oversight and a more practical route for complaints.
Two different labels sit inside the same rule
The law separates what providers build into an AI system from what deployers tell the public when they use it.
Providers of generative AI must make generated or manipulated audio, images, video and text detectable in a machine-readable form. The Commission says those marks should be effective, interoperable, robust and reliable as far as technically feasible.
Deployers have a more visible duty. They must disclose when people are exposed to emotion recognition or biometric categorisation. They must label deepfakes. They must also disclose AI-generated or altered text about matters of public interest when it has not gone through human review and editorial responsibility.
That last exception matters for publishers. Human review does not erase every ethical question around generated text, but it changes the specific Article 50 labelling duty described in the Commission's guidance.
A machine-readable mark and a label for a person are not the same thing. One is designed for detection systems. The other is there so a reader, listener or viewer can understand what they are seeing.
The industry code is optional; the law is not
About 190 organisations had signed the EU's Code of Practice on Transparency of AI-generated Content by the end of July. The list includes large model companies, media and detection businesses, retailers, public bodies and many smaller firms.
The code gives signatories a recognised way to show how they comply. Providers can follow its marking measures, while deployers can use its approach to labelling deepfakes and certain public-interest text.
Joining the code is voluntary. Meeting the legal obligations is not.
A company can choose another method, but it will need to show that its alternative is equivalently adequate. The Commission says those choices may then be assessed individually by different market-surveillance authorities.
That creates a practical reason to sign even for companies that dislike voluntary codes: one shared route is easier to explain across 27 national systems than 27 separate arguments about an in-house solution.
Enforcement starts before practice is settled
The AI Office can request technical documents, evaluate models, demand corrective measures and issue fines in the areas it supervises. National authorities handle much of the wider market surveillance.
The Commission has also opened complaint routes, including a general AI Act complaints tool, a whistleblower channel and a channel for downstream providers using general-purpose models.
The machinery exists. The harder questions now move into ordinary cases.
A visible label may disappear when content is cropped or reposted. A machine-readable mark may survive one editing tool and fail after another. A voice bot may disclose itself once at the start of a long call, leaving open how prominent that notice needs to be later.
The guidelines offer examples and definitions, but consistent enforcement will take decisions, tests and probably disputes. A common rule still has to survive many products and many national authorities.
What is confirmed, intended and still open
Confirmed: Article 50 transparency obligations apply from 2 August 2026. Interactive AI must identify itself, generative systems need machine-readable content marking, and deployers face disclosure duties for deepfakes and several other uses.
The Commission's position: the rules should reduce deception, help people calibrate trust and give businesses a predictable path to compliance. The voluntary code is its preferred practical route, not a replacement for the law.
Still open: how consistently labels will appear, how durable technical marks will be after ordinary editing, how national authorities will judge alternative methods and how quickly early complaints produce public decisions.
Also confirmed: the delayed high-risk rules are on a different clock. Sensitive uses such as employment, education and migration are due from 2 December 2027. AI embedded in regulated products follows on 2 August 2028.
Sunday's change is not the final AI Act moment. It is the point where a clear part of the law becomes something users can actually notice — and authorities can begin to test.
Sources
- European Commission — AI Act enforcement from 2 AugustPrimary Commission release published 31 July 2026. Source for the enforcement date, user-facing disclosure duties, machine-readable marking and new complaint channels.
- European Commission — Article 50 transparency guidelinesPrimary implementation guidance used to distinguish provider and deployer duties, the human-review exception and the authorities responsible for enforcement.
- European Commission — Transparency code and signatoriesPrimary code page used to verify that participation is voluntary, the legal obligations remain mandatory and about 190 organisations had signed by the end of July.
- European Commission — AI Act framework and timelinePrimary policy overview used to verify enforcement powers and the revised December 2027 and August 2028 dates for high-risk systems.
