2 August 2026 was the date many Italian companies had circled in red. In recent weeks the picture changed: Regulation (EU) 2026/1744 — the Digital Omnibus on AI, published in the Official Journal of the European Union on 24 July 2026 and in force from 27 July — deferred most obligations for high-risk systems. But it did not touch Article 50.
If your company runs a chatbot, generates text, images or audio with AI, or publishes synthetic content, obligations apply within days.
What applies on 2 August 2026 (and what slipped)
| Obligation | Application date |
|---|---|
| Annex III high-risk systems (Art. 6(2)) | 2 December 2027 |
| AI embedded in regulated products, Annex I (Art. 6(1)) | 2 August 2028 |
| Transparency — Article 50 | 2 August 2026 (unchanged) |
Two operational points:
- Generative AI systems already placed on the market before 2 August 2026 have until 2 December 2026 to comply with the machine-readable marking requirement in Article 50(2) only.
- Regulation 2026/1744 also introduces two new prohibitions in Article 5, applicable from 2 December 2026, covering the generation of intimate material without explicit consent and child sexual abuse material.
The message is therefore the opposite of "we have more time": the part of the AI Act that touches the largest number of companies — transparency — arrives on schedule.
The four obligations in Article 50
Article 50 distinguishes between those who provide the system (provider) and those who use it in the course of their activity (deployer).
1. Disclose that the user is interacting with an AI — Art. 50(1)
Providers of systems intended to interact directly with natural persons must ensure users are informed of this, at the latest at the time of the first interaction. The obligation falls away only where this is obvious to a reasonably well-informed, observant and circumspect person. The Commission's guidelines clarify that AI agents fall within this provision.
2. Mark synthetic outputs — Art. 50(2)
Providers of systems generating synthetic audio, images, video or text — including general-purpose AI systems — must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. In practice: watermarks, metadata, content provenance tooling.
3. Inform about emotion recognition and biometrics — Art. 50(3)
Deployers of emotion recognition or biometric categorisation systems must inform the people exposed to them, without prejudice to GDPR obligations. Before that, check whether the use falls under a prohibition in Article 5.
4. Disclose deepfakes and public-interest text — Art. 50(4)
Deployers of a system that generates or manipulates content constituting a deepfake must disclose that the content has been artificially generated or manipulated. The same applies to text published to inform the public on matters of public interest.
The exemptions — and the two most common mistakes
Exemptions exist, but they are narrower than they are often described:
- Assistive function for standard editing: marking does not apply where the AI performs a merely assistive function for standard editing and does not substantially alter the input data. Fixing grammar qualifies; rewriting or generating text does not.
- Artistic, creative, satirical or fictional works: the obligation is reduced to disclosing the existence of the generated content in a way that does not hamper the display or enjoyment of the work.
- Editorial control: for public-interest text the obligation does not apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility.
- Law enforcement: systems authorised by law to detect, prevent, investigate or prosecute criminal offences.
The two recurring mistakes:
- "The provider handles it." Article 50(2) sits with the provider, but Article 50(4) is yours: if you publish a deepfake or generated text on matters of public interest, the disclosure is on you.
- "A line in the terms of service is enough." The information must be given in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure, and must meet accessibility requirements (Art. 50(5)).
European fines and the Italian deepfake offence
On the administrative side, breaching Article 50 falls under Article 99(4) of the AI Act: up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. For SMEs, including start-ups, the lower of the two figures applies instead (Art. 99(6)).
In Italy a criminal dimension is added. Law 132/2025, in force since 10 October 2025, introduced Article 612-quater of the Criminal Code — "Unlawful dissemination of content generated or altered with artificial intelligence systems" — punishing with one to five years' imprisonment anyone who causes unjust harm by transferring, publishing or otherwise disseminating, without the consent of the person concerned, images, video or voices falsified or altered by AI and capable of deceiving as to their authenticity. The offence is prosecuted upon complaint, save for the cases of ex officio prosecution set out in the provision.
Operational checklist
- Map the touchpoints: chatbots and agents, generators of text, images, audio or video, emotion recognition or biometric categorisation systems.
- Qualify the role for each system: are you provider or deployer? The obligations differ, and a company is often both, on different systems.
- Check your vendors: ask in writing how they implement Article 50(2) marking, and put it in the contract.
- Design the notices: the chatbot disclaimer, the visible label on synthetic content, the notice for emotion recognition and biometrics.
- Define the editorial process for what you publish: who reviews, who signs off, who takes editorial responsibility. That is precisely what triggers the exemption.
- Document your choices and the exemptions you rely on: in a review, what counts is what you can demonstrate, not what you did without recording it.
- Consider the Code of Practice on transparency of AI-generated content, published in June 2026 and assessed as adequate by the Commission and the European AI Board. Signing is voluntary, but it is the fastest route to demonstrating compliance.
It is worth remembering that the Commission's guidelines on Article 50, adopted on 20 July 2026, are not binding — but they are the yardstick authorities will use to read the provision.
Where to start
Transparency is the AI Act obligation easiest to underestimate and most visible from the outside: a chatbot without a disclaimer or an unlabelled synthetic image can be spotted by anyone, with no inspection required. A gap assessment focused on Article 50 — which systems, which role, which notices, which evidence — takes a few days and closes the distance between what you do and what you can prove.