FACT CHECK
What you read elsewhere is already wrong.
Five very widespread claims about the AI Act, set against the consolidated text of the regulation.
Why this page exists
Since August 2026, the trade press and law firms have been publishing a great deal about the AI Act. Some of what is circulating in French has been out of date since 27 July 2026, the date Regulation (EU) 2026/1744 — the Digital Omnibus — entered into force and amended the text substantially. Here are the five claims we see most often, and what the consolidated text really says.
1. “A €15 million fine if your staff are not trained in AI”
Out of date. Article 4 has been rewritten. A business must now “take measures to promote the development” of its teams’ AI literacy, and the text adds that this obligation “does not oblige providers or deployers to guarantee a specific level of AI literacy on the part of an individual”. We have moved from an obligation to deliver a result to a best-efforts obligation.
An honest caveat: Article 4 is not in the list of infringements penalised by Article 99. Two indirect routes remain — the penalties each Member State lays down for any breach of the regulation, and a new power of the AI Office. So this is not “no risk at all”; nor is it €15 million for a missing training session.
2. “The fine is always the higher amount”
Wrong for an SME. Article 99(6) says, word for word: “In the case of SMEs, including start-ups, each fine referred to in this Article shall be up to the percentages or amount referred to in paragraphs 3, 4 and 5, whichever thereof is lower.”
In other words, where a large company has the higher of the two amounts applied to it, an SME has the lower one applied — and that holds for all three penalty tiers. The Digital Omnibus extended a similar rule to small mid-cap companies, but across two tiers only.
3. “The rules on high-risk systems apply in 2026”
Wrong. They apply on 2 December 2027 for the Annex III uses — recruitment, credit, education, access to essential services — and on 2 August 2028 for AI built into products that are already regulated. What has applied since 2 August 2026 is the transparency obligations in Article 50.
4. “2 August 2027 is the high-risk deadline”
Out of date. The date still exists, but it now concerns national regulatory sandboxes and bringing the oldest general-purpose AI models into compliance. Many timelines published online have not been updated since the Omnibus — including on sites treated as authoritative.
5. “The CNIL is the supervisory authority in France”
Not yet. In its work programme published on 7 April 2026, the CNIL itself writes that it “is preparing to be designated”. The bill that is to organise French governance was passed by the Senate on 18 February 2026, then sent back to committee in the National Assembly on 23 April, where it still sits.
In Belgium the situation is comparable: no law has been passed, and the IBPT is mentioned only in the 2025-2029 federal coalition agreement. Both countries missed the 2 August 2025 deadline set for designating national authorities.
Every claim on this site is checked against the consolidated text of the regulation, not against a press article. Every page carries the date it was last checked. If you find a mistake, write to us: it will be corrected and the correction will be flagged.
Does something you read elsewhere look wrong?
Send us the link. We check it against the source and publish the answer here.
Report an error to us →Sources
Page checked on 26 August 2026.