EU AI Act on 2 August 2026: exactly what changes

On 2 August 2026, the EU AI Act's Article 50 transparency obligations become applicable: chatbots must disclose that they are AI, AI-generated or manipulated content must be marked, and deepfakes must be labeled. The 2026 digital omnibus (an EU package that shifted several deadlines) deferred the high-risk regime to 2027/2028. Article 50 it did not defer. This date stands.

The omnibus caused enough confusion that many companies now believe "the AI Act got postponed." Parts of it did. The part most likely to apply to an ordinary business running a chatbot or publishing AI-generated content did not. Here is the precise picture.

What Article 50 requires from 2 August 2026

Article 50 of Regulation (EU) 2024/1689 sets transparency duties for specific AI uses:

• Chatbot disclosure. People interacting with an AI system (a support chatbot, a voice agent) must be informed they are talking to AI, unless it is obvious to a reasonably well-informed, observant and circumspect person given the circumstances and context of use. The Commission's non-binding Guidelines read that exception restrictively and expect the provider to be able to demonstrate it.
• Marking AI-generated content. Providers of systems that generate synthetic audio, image, video, or text must ensure outputs are marked as artificially generated or manipulated, in a machine-readable and detectable way (the "watermarking" duty, meaning software can tell the content came from a machine).
• Deepfake labeling. If you use an AI system to generate or manipulate image, audio, or video content that resembles existing people, objects, places, entities, or events and would falsely appear authentic or truthful, you must disclose that the content is artificially generated or manipulated. A related duty covers AI-generated text published to inform the public on matters of public interest, unless the text went through human review or editorial control and a natural or legal person holds editorial responsibility for it.
• Emotion recognition and biometric categorisation systems (software that reads feelings or sorts people by physical traits) require informing the people exposed to them.

These duties fall on both providers (who build or ship the system) and deployers (who use it in their business), and the split runs paragraph by paragraph. Article 50(1), the chatbot notice, and Article 50(2), the marking of synthetic output, are provider duties. Article 50(3), emotion recognition and biometric categorisation, and Article 50(4), deepfakes, are deployer duties.

Most SMEs are deployers, so the deepfake disclosure is yours outright. The chatbot notice is legally your vendor's to build in, unless you developed it, or had it developed, and put it into service under your own name or trademark, even for your own use, or you modified a vendor's system, for example with new training data, and put it into service under your own name. In either case you are the provider. Either way you are the one your customer meets, and an unlabelled assistant on your site is your problem in practice whoever owes the duty. Check that it is switched on, and get the vendor's answer in writing.

For the practical "how" of marking and disclosure, the European Commission published Guidelines on the Article 50 transparency obligations on 20 July 2026. The formal Communication, in substantively the same text, is dated 5 November 2026 (C(2026) 7702 final). They are non-binding. They complement the voluntary Code of Practice on Transparency of AI-Generated Content published on 10 June 2026. Note what the Code covers: Article 50(7) ties codes of practice to the detection, marking and labelling of artificially generated or manipulated content, so adhering to it can help evidence the marking, deepfake and public-interest text duties, not the chatbot notice. And the omnibus is explicit that such codes "have limited legal effect, and in particular do not grant a presumption of conformity". Useful, not a shield.

What the omnibus changed, and what it left alone

The 2026 digital omnibus, published in the Official Journal as Regulation (EU) 2026/1744 on 24 July 2026 and in force since 27 July 2026, reshuffled the AI Act timeline:

• High-risk obligations are deferred. Annex III stand-alone high-risk use cases (HR screening, credit scoring, and similar) now apply from 2 December 2027. Annex I embedded high-risk products (AI in regulated machinery, medical devices, etc.) apply from 2 August 2028. If your exposure was "we might be high-risk," you have real runway. Use it for preparation, not procrastination.
• Article 50 is NOT deferred. The transparency duties apply from 2 August 2026 as scheduled.
• One technical grace period exists: the machine-readable watermarking sub-duty gets a transition to 2 December 2026 for systems already on the market before 2 August 2026. Note what this covers: the technical marking mechanism, not the duty to disclose. A chatbot still has to say it is a chatbot from 2 August.

Already in force: Article 4 AI literacy

One obligation predates all of this and keeps getting forgotten: Article 4 (AI literacy) has applied since 2 February 2025. The omnibus rewrote it rather than removing it, and the rewrite is worth reading closely. Providers and deployers must now "take measures to support the development of AI literacy" of their staff and of others operating AI systems on their behalf, judged against those people's technical knowledge, experience, education and training, and the context of use. The article then says plainly that the duty "does not require providers or deployers to guarantee any specific level of AI literacy of any individual".

That second sentence is the reassuring part, and it is the one most summaries drop. You are not being asked to certify anyone's competence. You are being asked to have done something deliberate about it. In practice that means you can show something: a short policy, role-appropriate training, a record of who was trained on what.

Readiness checklist for 2 August

Work through this in an afternoon. Most items are documentation, not engineering:

1. Inventory your AI touchpoints. Where do customers interact with AI (chat, voice, email triage)? Where do you publish AI-generated or AI-assisted content (product images, marketing video, articles)?
2. Add chatbot disclosure. A clear line at the start of the interaction ("You are chatting with an AI assistant") is likely to suffice in most cases; riskier contexts, including complaints handling, are likely to need periodic reminders (Guidelines, para 40). Do not rely on "it's obvious." That is a judgment call an authority may not share, and the Guidelines limit the exception to cases where "there is almost no doubt left" about the nature of the interaction.
3. Label synthetic media you publish. Deepfake-style content (realistic people, places, events) needs a visible disclosure. Text on matters of public interest that went through human review or editorial control, with a natural or legal person holding editorial responsibility, is exempt from the text disclosure. Unreviewed machine output is not.
4. Ask your AI vendors about marking. The machine-readable watermarking duty sits primarily with providers. Ask yours how their outputs are marked and note the answer. If a system was on the market before 2 August 2026, the technical marking has a grace window to 2 December 2026.
5. Write down your Article 4 literacy measures. A one-page policy plus a training record is a defensible floor for a small deployer. Nothing written down is not.
6. Classify what you run. Confirm none of your systems fall into Annex III high-risk categories. If any do, your deadline is December 2027. Distant, but the gap analysis is worth starting now.
7. Keep evidence. Screenshots of disclosures, the vendor answers, the literacy policy, the classification memo. If anyone asks, the file is the answer.

What 2 August actually is

For most businesses, not a cliff. A documentation exercise with a hard date. The companies heading for a bad autumn are the ones who read "the omnibus deferred things" as "nothing applies," and are running undisclosed chatbots and unlabeled synthetic media into an obligation that is live.

FAQ

Q: Did the omnibus postpone the AI Act?

A: It deferred the high-risk regime (Annex III to 2 December 2027, Annex I to 2 August 2028). Article 50 transparency duties apply on 2 August 2026 as originally scheduled, and Article 4 literacy has applied since 2 February 2025.

Q: We use a third-party chatbot vendor. Whose duty is the disclosure?

A: Both sides have roles: the provider must design the system so people can be informed; as the deployer you are the one facing your customers. Making sure the disclosure actually appears in your implementation is the safe operating assumption.

Q: Does an AI-assisted blog post need a label?

A: AI-generated text published to inform the public on matters of public interest requires disclosure, unless it has undergone human review or editorial control and a natural or legal person holds editorial responsibility. The Guidelines list advertising copy and product descriptions as outside the text duty, unless they include claims on matters such as health, consumer safety or sustainability (para 131). Auto-published machine output on public-interest topics is the case that needs a label.

Q: What is the penalty exposure?

A: We do not attach fine figures to Art. 50. Serious numbers require citing the exact penalty tier for the exact conduct, and most figures circulating online mix the tiers. The defensible statement is shorter: the duties apply, enforcement exists, and the fix costs an afternoon. The precise national enforcement setup is still being stood up in many member states, which is exactly why a clean evidence file now is the cheap option.

Sources

• Regulation (EU) 2024/1689 (EU AI Act), full text incl. Articles 4 and 50: eur-lex.europa.eu
• European Commission, AI Act overview and implementation timeline: digital-strategy.ec.europa.eu
• European Commission, Guidelines on the transparency obligations under Article 50, published 20 July 2026; formal Communication C(2026) 7702 final, 5 November 2026 (non-binding): ec.europa.eu
• European Commission, digital omnibus (simplification package): digital-strategy.ec.europa.eu