AI Act regulations take effect

On August 2, new obligations from the AI Act came into effect, including requirements that chatbots must identify themselves, generated content must have a machine-readable label, and deepfakes must be labeled.

Broader AI Act provisions remain in force

These obligations primarily concern editors and platforms, but the AI Act also includes provisions that have been in force for eighteen months and are less widely discussed. Article 4 of the AI Act mandates that suppliers and deployers of AI systems take measures to support the development of AI mastery by their personnel and others responsible for the operation and use of these systems.

Widespread AI use with minimal training

According to a YouGov survey for Microsoft France conducted in January, 80 percent of directors use generative AI every week, 61 percent of users go through personal accounts, and more than seven managers out of ten have received no training. These figures highlight a widespread use of AI with minimal training.

Regulation clarifies AI mastery requirements

Article 4, as rewritten by the Omnibus regulation effective July 27, 2026, specifies that suppliers and deployers must support the development of AI mastery, taking into account factors such as technical knowledge, experience, education, and training. The regulation does not require a specific level of AI mastery, leaving it to companies to define what is proportionate to their situation.

Market sets compliance standards

The market has set a de facto standard for compliance, with training offers typically including a two-hour session, a quiz, and an individual certificate per participant. The warning that accompanies these training offers is a potential fine of 15 million euros or 3 percent of global turnover for non-compliance.

European Commission questions training effectiveness

However, Article 4 is not listed in the violations mentioned in Article 99, paragraph 4, which outlines the obligations sanctioned at that level. The European Commission has expressed concerns about the effectiveness of merely providing user manuals or asking staff to read them, and it has compiled a repertoire of practices including e-learning platforms, in-person training, and bootcamps.

Training catalog emphasizes skill, not just participation

The European Commission has published a training catalog, but it emphasizes that a certificate proves only that a training took place, not that a skill is retained. Studies show that users who rely on AI tools may perform worse without them, with participants' correct answer rates dropping from 46 percent to 31 percent when faced with an assistant that confidently makes mistakes.

Enforcement responsibility lies with member states

Despite this, no European fine has been issued specifically for non-compliance with Article 4. Member States are responsible for providing effective and proportionate sanctions in their national law, and it is national market surveillance authorities that have been monitoring compliance since August 2, not the European Commission.

France lacks a centralized authority for AI training

In France, no authority is specifically in charge of Article 4; the landscape is divided between the CNIL, the DGCCRF, the Arcom, and others depending on the sectors. The Commission specifies that a sanction is mainly likely when an incident occurs and the training of the people involved is found to be insufficient.

Focus shifts from certificates to practical skills

However, on the day when a decision made with AI goes wrong, the question raised will not be "do you have a certificate?" It will be: "were your teams able to see that the tool was wrong?" A certificate does not answer this question. A measure does. A certificate measures an AI training followed. A skill is measured by what the teams can still do without the tool.

Proposed dimensions to track AI mastery

The three dimensions proposed to track include residual autonomy, quality of discernment, and transmission. These dimensions become evidence and are put in place without a laboratory. Residual autonomy involves redoing ten already processed cases without the tool quarterly.

Quality of discernment and transmission

Quality of discernment measures the proportion of AI proposals that have been modified, contested, or rejected before validation. Transmission measures the time it takes for a new arrival to reach the autonomy that was the norm before AI. If this delay increases, the training produces users, not professionals.

Future regulations and recommendations

The Omnibus regulation has postponed the entry into force of certain provisions, such as those in Article 26, to December 2, 2027. The European AI Board is tasked with writing recommendations and common objectives for AI mastery by 2027, based on what companies have managed to demonstrate.

Article 4 as a skill-focused approach

Article 4 can be read as a constraint that produces paper or as a question of skill. The latter approach produces a solid proof when requested and provides information that very few companies possess today: use case by use case, the places where AI makes teams better and those where it only makes them faster.