AI transparency: since August 2, 2026, an AI must say it is one

Since August 2, 2026, the European regulation on artificial intelligence has been applicable. Among the obligations that kick in on that date, only one concerns nearly every company that has put AI on the front line: transparency. People must know when they are talking to a machine, and content made by a machine must be recognizable as such.

This is Article 50 of Regulation (EU) 2024/1689. It requires no file, no declaration, no authorization. It requires you to say things, in the right place and at the right time.

Four obligations, aimed at different people

Say it's an AI. Providers of systems intended to interact directly with natural persons must design them so that those persons are informed they are interacting with an AI system. This applies to a chatbot on a website, an automated phone greeting, or an assistant built into a customer area.

Label synthetic content. Providers of systems that generate audio, images, video or text must mark the outputs in a machine-readable format, detectable as artificially generated or manipulated. The regulation exempts this marking when the system only assists with formatting and does not substantially alter the data provided.

Warn when emotion recognition or biometric categorization is used. Here, it is the party deploying the system that informs exposed persons about how it works, and that processes the data in accordance with the GDPR.

Flag a deepfake. Anyone deploying a system that produces manipulated images, audio or video must disclose that the content was generated or manipulated by an AI. The same rule applies to text published to inform the public on matters of public interest, unless that text has undergone human review and a person holds editorial responsibility for the publication.

The detail that decides everything: when, and how

The regulation specifies that this information must be provided in a clear and distinguishable manner, at the latest at the time of the first interaction or first exposure, and that it must comply with the applicable accessibility requirements.

In other words, a line at the bottom of the terms and conditions is not enough. For a chatbot, the disclosure happens when the conversation opens, inside the window itself, readable by a screen reader as well as by a hurried eye.

The only exemption from the duty to say it's an AI rests on obviousness: it does not apply if this is clear from the point of view of a reasonably well-informed, observant and circumspect person, taking into account the circumstances and context. A widget labeled "assistant" does not make it obvious.

A second deadline, on December 2, 2026

A regulation dated July 8, 2026, published in the Official Journal of the European Union on July 24, 2026, simplified the implementation of the AI rules. It leaves the transparency obligations intact, but adds a four-month transition period: providers of systems generating synthetic content that were placed on the market before August 2, 2026 have until December 2, 2026 to comply with the marking obligation.

The same text postpones most of the rules on high-risk systems, to December 2, 2027 for those in Annex III and to August 2, 2028 for those in Annex I. Transparency has not moved.

What it can cost

Failing to meet the transparency obligations of Article 50 can lead to an administrative fine of up to €15,000,000 or 3% of worldwide annual turnover, whichever is higher. For SMEs and startups, the regulation applies the lower of the two figures.

These ceilings are those set by the regulation. The penalty regime and the authorities in charge of enforcing it are a matter for each Member State. In France, the CNIL says it is preparing to be designated as a market surveillance authority under the AI regulation. It also reminds us of a point that is often overlooked: this regulation does not replace the GDPR, and the two apply together.

Five checks, fifteen minutes

Open your own chatbot. Does the first sentence say it is an AI, without anyone having to click or unfold anything? Information tucked away in a tooltip is not in the right place.

Listen to your phone greeting. An automated voice system falls within scope exactly like a website widget. The greeting message is often still the old one.

Look at your visuals. Illustrations, avatars, voice-overs: anything that comes out of a generative tool is covered by the marking requirement, and your most visible publications are rarely your most recent.

Go back over your articles. Editorial content produced by AI, reviewed by an identified person who takes editorial responsibility for it, is exempt from the disclosure obligation. That review still has to actually happen.

Question your vendors. If your chatbot, your image generation tool or your speech synthesis comes from a software vendor, machine-readable marking is first and foremost its job. A written question, with the answer kept on file, is better than an assumption.

The right reflex

This is one of the rare obligations that costs a few minutes and pays off right away. Companies that openly announce their chatbot rarely see usage drop: they see better-phrased questions, because people know who they are talking to. Here, transparency is less a constraint than a conversation setting.

What the text does not ask, however, is that you apologize for using AI. One sentence is enough, in the right place.

Sources: Articles 50, 99 and 113 of Regulation (EU) 2024/1689, and Articles 111 and 113 as amended by Regulation (EU) 2026/1744 of July 8, 2026, consulted on EUR-Lex on September 22, 2026; CNIL Q&A on the entry into application of the AI regulation and CNIL 2026 work programme, consulted on September 22, 2026.