Diritto intelligente – Issue N. 19

If one message runs through this issue of Diritto Intelligente, it is that regulators and courts have stopped waiting for the AI Act to finish its applicability phase.

The signals point in opposite directions, and that is precisely the problem. The EU Commission’s first Article 112 report concludes that it is “not yet time” to amend the framework. The high-risk obligations look set to slide to December 2027 under the Digital Omnibus package. And yet, in the same weeks, authorities across Europe acted as though the future had already arrived.

Look at the Munich Regional Court’s decision on Google AI Overviews. For what is believed to be the first time, a court treated an AI-generated answer not as a forwarded search result but as the provider’s own statement, stripping away the intermediary shield that platforms have relied on for years. The reasoning does not stop at Google. Any provider of synthesised answers now owns what its model invents. A disclaimer, the court made unmistakably clear, is not a shield.

The Garante’s warning against Myndoor sends the same message from a different direction. The prohibition on inferring workers’ emotions under Article 5(1)(f) is not a 2027 problem waiting patiently in the wings; it is live enforcement today, applied to a well-being tool marketed as employee-friendly. The lesson for deployers is uncomfortable: good intentions and aggregate reporting do not neutralise a hard prohibition.

Italy’s draft implementing decrees then sharpen the point to a fine edge. Dismissals based solely on automated processing become null and void. New Article 437-bis introduces criminal liability, punishable even for gross negligence. And the civil disclosure regime turns your AI Act compliance file, logs, technical documentation, evidence of human oversight, into the decisive procedural asset. Fail to produce it, and the court may simply deem the claimant's facts established.

Here is the through-line worth carrying into your next board meeting. The gap between when the AI Act’s heaviest obligations formally bite and when liability actually attaches is widening, not closing. Reading the postponement as permission to pause is a serious misreading of the moment. The very fact that Brussels felt compelled to grant more time tells you the market was judged unready.

The rulebook is still settling. Liability is not.