A deadline that stands
The Digital Omnibus, adopted at the end of June, postponed the obligations applying to high-risk AI systems. But it left one deadline intact: 2 August 2026, the date from which the transparency obligations of Article 50 of the AI Act apply.
What Article 50 says
- Tell people they are interacting with an AI, when this is not obvious: a chatbot on your website, an assistant in customer service.
- Disclose generated or manipulated content (text published to inform the public, images, sounds, videos), including deepfakes.
- Technically mark the content produced by providers of generative systems, with tolerance for systems already on the market.
Who is affected?
Many more companies than you might think. A contact form run by a conversational agent, product sheets written by an AI, campaign visuals generated by a model: all cases where informing the user becomes an obligation.
By contrast, an internal assistant used by your employees to search your documents, such as MIA, is a use where the user knows perfectly well they are addressing an AI: the issue there is less transparency than control over data.
Concrete cases, beyond the chatbot
Transparency is not only about the assistant displayed in the corner of a website. A few common situations worth checking:
- Customer service: a virtual agent answering emails or messages must present itself as such from the first exchange.
- Marketing: realistic visuals produced by an AI, and all the more so synthetic videos or voices, must be disclosed when shown to the public.
- Institutional communication: AI-generated text published to inform the public on matters of public interest must be identified as such, unless it has undergone human review and editorial responsibility.
- Recruitment: an automated first exchange with candidates must also be transparent, pending the high-risk obligations that will apply to these systems in December 2027.
Transparency does not mean distrust
Many companies fear that displaying “you are talking to an AI” will spoil the experience. Experience suggests the opposite: a clear notice, with a simple way to reach a human, reassures people. What destroys trust is finding out afterwards that you were talking to a machine.
Good practice is therefore to write a short, honest notice, to place it at the right moment (at the start of the exchange, not in the terms and conditions), and always to provide a way out to a human contact.
Three actions before the autumn
- List the touchpoints where an AI interacts with your clients or the public.
- Add a clear notice at the start of each exchange and on published content that needs one.
- Keep a register of the AI tools in use, their purposes and the people responsible.
This register is also the first building block of AI governance. It is the starting deliverable of our AI audit, which at the same time prepares the next deadlines of 2027 and 2028.
Sources: Regulation (EU) 2024/1689 (AI Act), Article 50; Digital Omnibus regulation on AI adopted in June 2026.
