AI Act: the August 2 deadline was NOT delayed for chatbots and AI agents
In May, the European Parliament and the Council of the EU agreed to delay part of the AI Act's rules. The headline traveled fast: "the EU is delaying the AI Act." What didn't travel as fast: the delay doesn't cover the obligation that applies directly to you if you run a chatbot, an AI agent that talks to customers, or publish any AI-generated content. That one stays exactly where it was — August 2, 2026, with fines enforceable from the same day.
What actually got delayed — and what didn't
The Digital Omnibus, the AI Act's simplification package, cleared its last legislative hurdle: the European Parliament passed its final vote on June 16, 2026, and the Council of the EU gave final sign-off on June 29. It's no longer a proposal or a provisional deal — it's law, on its way to publication in the EU's Official Journal.
What it actually delays: obligations for stand-alone "high-risk" systems (Annex III — for example AI used in hiring, credit scoring, critical infrastructure) move from August 2, 2026 to December 2, 2027. AI embedded in products already regulated elsewhere (medical devices, vehicles, toys) gets an even longer runway — August 2, 2028.
What it does not delay: Article 50 — the transparency duties that apply to any business using AI to talk to people or to generate content. Those stay at August 2, 2026, unchanged from the law's original text. So does the authorities' power to fine for breaking them, live from that same date.
Source: Articles 50 and 99, Regulation (EU) 2024/1689 (AI Act) · Digital Omnibus, adopted June 29, 2026. Simplified timeline, not to scale.
The 4 obligations taking effect on August 2
Article 50 of the AI Act (Regulation (EU) 2024/1689) sets out four distinct duties — two for the people who build AI systems (providers) and two for the people who use them (deployers).
AI-interaction disclosure. An AI system that talks directly to people — a chatbot, voice assistant, support agent — must make it clear, no later than the first interaction, that the user is dealing with AI. The exception: when it's obvious anyway, for example a bot explicitly labeled "AI Assistant" from the first message.
Marking on generated content. Text, image, audio, or video produced by AI must be marked in a machine-detectable format as artificially generated or manipulated. This is the one timing nuance in the whole piece: systems already on the market before August 2, 2026 get a transition window until December 2, 2026 to implement marking. New systems launched on or after August 2 get no grace period — they mark from day one.
Notice for emotion recognition or biometric categorization. Anyone using a system that recognizes emotions or classifies people on biometric criteria must inform the people exposed to it.
Disclosure for deepfakes and AI content on public-interest matters. Any AI-generated or manipulated image, audio, or video that constitutes a deepfake must be labeled as such. AI-generated text published to inform the public on matters of public interest must be disclosed too — unless it went through human review, with a person holding editorial responsibility for the content.
Who's actually affected
The law doesn't look at how risky the system is — it looks at what it does. In practice, you fall under Article 50 if you have a chatbot, voice assistant, or AI agent that talks to customers, no matter how simple or old it is; plenty of businesses installed a chatbot years ago and never thought of it as a "regulated AI system." The same goes if you generate or publish AI content — text, images, audio, video — that reaches the public, or if you use AI for photo or video material that's artificially generated or manipulated, especially if it could pass for real footage. It also applies if you publish AI-written or AI-edited text on matters of public interest without documented human review.
For a business running operational AI agents — the kind of work we do at MassAI — that means any agent talking directly to a customer or a lead needs to meet the Article 50(1) disclosure by design, not bolted on afterward as a patch. We cover what a properly built agent looks like in our guide to operational AI agents.
Fines, and what to do before the deadline
Breaching Article 50 falls into the same fine tier as most operator obligations under the AI Act: up to €15,000,000 or 3% of worldwide annual turnover, whichever is higher. SMEs, including startups, get the lower of the two caps. The power to issue these fines takes effect on that same August 2 date — the obligation and the enforcement power arrive together, not one after the other.
In practice, it's worth running a quick inventory before August 2: which of your AI systems talk to people, what AI-generated content you publish, whether any of it counts as a deepfake or public-interest text. For each one, check that the disclosure to the user is clear and visible from the first interaction, not buried in terms and conditions nobody reads.
The information above comes from the official text of Articles 50 and 99 of the AI Act and isn't a substitute for legal advice on your specific situation. For complex systems or gray areas — especially the "obvious" exception or the editorial-review carve-out — a targeted check with a specialized consultant is worth the investment.
Sources: ↗ Official text, Article 50 AI Act (European Commission) · ↗ Council of the EU press release — Digital Omnibus · ↗ Article 99 AI Act — penalties
See how we build for the AI Act →