Compliance

AI Act: when do you have to disclose the use of AI?

Deepfakes, chatbots, generated text, influencer images: what the AI Act and French law require you to disclose since 2 August 2026, and what needs no label.

The DPO Connect team 18 August 2026 6 min read

Since 2 August 2026, the transparency obligations of the AI Act (Regulation (EU) 2024/1689) have applied in full. In practice, some content generated or altered by AI must now carry an explicit disclosure, or the publisher faces penalties. But far from everything is covered. Here is what the law requires, and what it leaves you free to do.

What the AI Act says

Article 50 of the Regulation follows a simple logic: as soon as content or an interaction produced by AI is likely to mislead someone about what they see, read or hear, transparency becomes mandatory. Four situations account for most of the obligations.

The legal disclosure obligations

Deepfakes and manipulated images

Clear, visible labelling is mandatory for content generated or manipulated by AI that resembles real people, places, objects or events but is not genuine: deepfakes within the meaning of Article 3, point 60, of the Regulation.

The label must be noticeable from the first exposure to the content: a doctored image of a public figure, a video depicting an event that never happened, an audio recording imitating a real voice must all disclose that they are artificial.

Direct interaction: chatbots and assistants

You must inform users when they are interacting with an AI system (chatbot, voice assistant, conversational agent), unless this is obvious from the context to a reasonably well-informed person.

A customer support chatbot must therefore introduce itself as such from the start of the conversation. This is the choice we made for Lea, the AI assistant built into DPO Connect: users know at all times that they are talking to an AI.

Text published to inform the public

Articles or texts generated by AI and published to inform the public on matters of public interest must disclose the use of AI.

The exception matters: no disclosure is required when the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for it. A newsroom that uses AI as a production tool, then reviews and stands behind its articles, has nothing to disclose.

Influencers and advertising: the French rule

In France, Law No. 2023-451 of 9 June 2023 on commercial influence adds a specific obligation: any image generated or altered by AI that shows a face or a body must carry the words ‘Images virtuelles’ (virtual images), and ‘Images retouchées’ (retouched images) for conventional retouching.

This national rule applies on top of the AI Act, on every platform, and is enforced by France’s DGCCRF.

What is not mandatory

Personal or internal use

No disclosure is required for uses that involve no misleading public distribution: writing internal emails, generating code, producing drafts, meeting summaries or working documents. The AI Act protects the public, not your internal processes.

Artistic or satirical works

Works of art, fiction, parody and satire benefit from a lighter regime: where the content is evidently part of a creative work, transparency can be limited to a disclosure that does not hamper the display or enjoyment of the work, provided the public is not at risk of being misled about the reality of what is depicted.

In practice: 4 habits to stay compliant

Map your uses of AI. List who in the organisation generates what (text, images, chatbots) and for whom.
Classify each use: public or internal? Only public distribution that is likely to mislead triggers an obligation.
Standardise your disclosures. A clear, consistent wording (‘AI-generated content’, ‘Images virtuelles’) works better than case-by-case labels.
Document editorial review. For texts reviewed by a person, keep a record of the approval: that is what exempts you from the disclosure.

These habits follow the same logic as GDPR: list, qualify, document, prove. If you already manage your GDPR compliance in DPO Connect, mapping your AI uses fits naturally into your roadmap and your evidence vault.

Sources: Regulation (EU) 2024/1689 (AI Act), Articles 3, 50 and 99; French Law No. 2023-451 of 9 June 2023 regulating commercial influence; European Commission guidelines on the application of Article 50 (2026).

Frequently asked questions

Since when have the AI Act transparency obligations applied?

The transparency obligations in Article 50 of the AI Act have applied since 2 August 2026, the date of general application of Regulation (EU) 2024/1689. Since that date, three disclosures have been mandatory. Deepfakes must be labelled as artificial content. Chatbots must identify themselves as such. AI-generated text published to inform the public must state its origin. The Regulation was phased in. It entered into force on 1 August 2024. Practices posing an unacceptable risk have been banned since February 2025. Obligations for general-purpose AI models followed in August 2025. The rest of the text, including transparency, has applied since August 2026. Organisations that publish AI-generated content must therefore comply now. The text provides for no additional grace period.

Do you have to disclose that a text was written with AI?

Yes, a disclosure is mandatory when AI-generated text is published to inform the public on matters of public interest, under Article 50 of the AI Act. There is an important exception. No disclosure is required if the text has undergone human review. A person or organisation must also hold editorial responsibility for it. In practice, a news article generated and published without review must disclose the use of AI. The same article, reviewed, corrected and owned by an editorial team, is exempt. Internal uses trigger no obligation: emails, drafts, code, working documents. The decisive test is the risk of misleading the public about where the information comes from. This rule has applied since 2 August 2026.

What is a deepfake under the AI Act?

A deepfake is AI-generated or manipulated image, audio or video content that resembles real people, places or events and could falsely appear to be authentic. This definition is set out in Article 3, point 60, of Regulation (EU) 2024/1689. Anyone who publishes such content must clearly disclose that it is artificial. The label must be visible from the first exposure to the content. It cannot be tucked away in a legal notice or a secondary description. Examples include a video depicting a fictitious event, or audio imitating a real voice. Evidently artistic or satirical works benefit from a lighter regime. The disclosure is still required, but it must not hamper the enjoyment of the work.

What label must an AI-generated influencer image carry?

In France, any influencer image showing a face or body generated or altered by AI must carry the words ‘Images virtuelles’ (virtual images). This obligation comes from French Law No. 2023-451 of 9 June 2023 regulating commercial influence. The same law requires the words ‘Images retouchées’ (retouched images) for images that have simply been edited. It covers all commercial communication on social media, whatever the platform. This French rule comes on top of the European transparency obligations of the AI Act. The label must be clear, legible and identifiable on the image or video throughout its display. France's DGCCRF has been actively checking these labels since 2024. A missing label is treated as a misleading commercial practice. It carries criminal penalties: up to two years' imprisonment and a €300,000 fine.

What are the penalties for failing to disclose AI use?

Breaching Article 50 of the AI Act can lead to a fine of up to €15 million or 3% of total worldwide annual turnover. The higher of the two amounts applies, under Article 99 of the Regulation. These fines have applied since 2 August 2026 to providers and deployers of AI systems alike. In France, influencers face additional penalties. Omitting the ‘Images virtuelles’ label counts as a misleading commercial practice. It is punishable by two years' imprisonment and a €300,000 fine. Reputational risk adds to the legal risk. Being identified as a publisher of undisclosed AI content erodes public trust for a long time. Documenting your disclosures and your editorial review remains the best protection in the event of an inspection.