Skip to content

AI Act: what to do before 2 August

On 2 August 2026 the AI Act's transparency obligations start to apply — including if you only use off-the-shelf tools. Three things to check this week and who is responsible for each.

Dan Cristian Alexandrescu5 min read
Notă

An informational briefing for management, not legal advice. It reflects the text of Regulation (EU) 2024/1689 and the amendments adopted in June 2026. For how it classifies your company, and for documents with legal standing, talk to your lawyer or DPO.

On Sunday 2 August, the AI Act’s transparency obligations — Article 50 — start to apply. This is not a rule for technology companies. It is a rule for anyone using AI in front of the public, including if you only use off-the-shelf tools: ChatGPT, Gemini, Midjourney, a chatbot bought ready-made.

The good news is that the list is short and none of it asks you to stop doing anything. All of it asks you to say something. On 3 August, the difference between a compliant company and a non-compliant one will usually be one sentence in a chatbot and a badge on a few images.

The short version, for management
04
  • From 2 August 2026 you must disclose three things: that the user is talking to an AI, that an image or video was generated or manipulated with AI, and that a text published on matters of public interest was AI-generated.
  • It applies even if you only use off-the-shelf tools. Who built the model doesn’t matter; how you use it does.
  • The deadlines for “high-risk” systems were deferred in June 2026 (to 2027–2028). Article 50 transparency was not deferred.
  • For AI text there is an alternative to disclosure: documented human review, where a named person takes editorial responsibility. Usually the better option for the brand.
— What applies

The three obligations, in boardroom language

03
  • 01

    The chatbot has to introduce itself

    If you run a conversational assistant on your site, on WhatsApp or in Messenger, people must know from the first interaction that they are talking to an AI — unless it is obvious from context. “Obvious” is a narrow exception; if the bot carries a human name and a face, it is not obvious. Cost of compliance: one sentence.

  • 02

    AI images and video get disclosed

    Content generated or manipulated with AI that could be mistaken for reality — people, places, events — must be visibly disclosed. It doesn't cover obviously artistic or stylised illustration, but it does cover the “product photo” that never existed and the happy-customer portrait who isn't a customer.

  • 03

    AI text on matters of public interest

    Text generated with AI and published to inform the public on matters of public interest must be disclosed — or must pass through a human review in which an identifiable person takes editorial responsibility. Most brands choose the second.

— Responsibility

Who answers for what

SituationWho carries the dutyWhat it means in practice
Chatbot on your site, under your brandIn practice, youYou add the disclosure to the first message and the widget. Even if the platform is someone else's, the public sees you.
Image generated in Midjourney, published in your adYou, as a professional userA visible mark on the material. The provider is responsible for the invisible technical marking, not the one your customer sees.
Blog post written with AI, on a matter of public interestYouEither you disclose it, or you document the human review and who takes editorial responsibility.
Machine-readable technical marking of generated contentThe model providerNot your job. Their deadline for systems already on the market is 2 December 2026.

The split above is the working rule we apply in audits. The exact provider / professional-user classification for a specific setup remains a question for a lawyer.

This week’s confusion. Plenty of companies heard “the AI Act was delayed” and closed the file. What was deferred, through the simplification package adopted in June 2026, are the obligations for high-risk systems — recruitment, credit scoring, education, access to essential services — moved to December 2027 and August 2028. Article 50 transparency was left untouched and applies on 2 August 2026.
— To do

Three decisions for this week

  • Ask for the list, not the impression

    Send one message to marketing and IT: “where do we use AI that reaches customers?”. Ask for it in writing, channel by channel. It will almost certainly be longer than you expect — AI features that appeared inside platforms you already run are the easiest to forget.

  • Decide on AI text: disclose or take responsibility

    This is a brand decision, not a technical one, and it is yours. If you choose editorial responsibility, name the person and put the procedure on paper — one page is enough, but it has to exist before publication, not after.

  • Appoint an owner, with a name

    Not “marketing”. One person who keeps the register of AI systems and who gets asked first if a complaint arrives. Without a name, the list above gets made once and goes stale in six weeks.

— FAQ

Questions we got on this edition

03
  • We're a 15-person company. Is anyone really going to fine us?

    The Regulation's ceiling for breaching the transparency obligations runs up to €15 million or 3% of worldwide turnover, whichever is higher — and for SMEs the lower of the two applies, with authorities obliged to take company size into account. The realistic risk for a small company isn't the maximum fine; it's a complaint from a competitor or a customer, an inspection, and the time it eats. The cost of compliance is under a day's work.

  • Do we have to write “generated with AI” on everything we publish?

    No. The obligation targets content that could be mistaken for reality, and text published on matters of public interest. An obviously stylised illustration or a graphic banner is out of scope. A product photo that looks real but never existed is in scope. When it's borderline, marking it is cheaper than the argument.

  • What about material published before 2 August?

    The rule applies from the date it takes effect, but your archive stays public and visible. In audits we recommend a pass over the last 90 days of active creative — anything still running or still being promoted gets marked, the rest is left alone.

— AI Briefing

Four days to the deadline. An audit fits inside one.

We tell you where you use AI, what is compliant and what is not, and we implement what's missing — for the sites and campaigns we already run, with no coordination needed with anyone else.