1. First publication of eIDAS guidelines for AI (July 18, 2025)
Brussels has finally set an operational framework following the adoption of the AI Act. These guidelines directly echo Article 25 of the eIDAS Regulation: any electronic evidence, including that generated by AI, can no longer be dismissed solely on technological grounds.
2. Targeted definition of the models concerned
Obligations apply to general-purpose systems generating text, code, or video that exceed 10^23 FLOPs of computing power. In practice, any model crossing this threshold may be classified as “systemic.”
3. Transitional phase for existing models
Models placed on the market before August 2, 2025, benefit from a two-year compliance period (until 2027), but this tolerance is strictly limited: non-compliance may result in retroactive sanctions.
4. Differentiated responsibilities based on role
Only actors who substantially modify a high-risk general-purpose model will be subject to the same obligations as the model’s developers. End users remain exempt, unless otherwise stipulated by contract.
5. Training transparency: a problematic gap
The requirements are vague: what constitutes a “sufficiently detailed” summary of training sources? For legal professionals, this lack of clarity carries a real risk of litigation, particularly in terms of intellectual property.
⚖ Realistic takeaway:
Europe is crafting a careful balance between technological innovation and rights protection. This initial set of guidelines provides welcome clarity, but it also leaves significant questions unanswered, notably on legal liability, copyright in generation chains, and model insurance.
And this is just the beginning: my next post will focus on the “Digital Fairness Check”, the mechanism set to complement this AI framework… without overloading your neurons just yet 😉
