2 August 2025 – Another milestone towards the full applicability of the AI Act

While the provisions concerning prohibited AI practices (Article 5 AI Act) and AI literacy (Article 4 AI Act) have already been applicable since 2 February 2025, further provisions come into force from 2 August 2025. These include requirements for general-purpose AI models (GPAI models) and specific rules on regulatory oversight. This represents another milestone towards the full applicability of the AI Act.

• The rules now in force include provisions on the foundations of market surveillance (Chapter III, Section 4 AI Act). These set requirements for notified bodies and for designating notifying authorities. The relevant authority structure is particularly important in the context of conformity assessment procedures required for high-risk AI systems.
• Particular importance attaches to the provisions – now likewise applicable –governing providers of ‘general-purpose AI models’ (GPAI models) in Articles 51–56 AI Act (Chapter V). The obligations placed on providers of GPAI models include the preparation of technical documentation, the provision of information to downstream providers, the adoption of a strategy to ensure compliance with copyright law, and the publication of a summary of the training data.
  
  The personal scope of the rules on GPAI models is, in principle, narrow. In its guidelines pursuant to Article 96 (1) AI Act, the European Commission imposes stringent criteria for establishing the existence of a GPAI model, with the result that only a few actors within Europe fall within the scope of the rules. The guidelines explain that the computational resources expended on training (training compute, measured in floating-point operations (FLOP)) are decisive. A presumption that a GPAI model exists arises where the training compute exceeds 10²³ FLOP. Irrespective of the narrow personal scope of application regarding providers of GPAI models, users of GPAI models who integrate them into their own applications should, when selecting the appropriate model, ensure that the respective provider complies with the requirements of the AI Regulation.
  
  An entity that modifies an existing GPAI model may also qualify as a provider of a GPAI model. The thresholds here are likewise high, so that downstream actors will attain provider status only in exceptional cases. According to the Commission’s guidelines, such status is established only where more than one-third of the compute originally used to train the model is expended on the modification (e.g. fine-tuning).
  
  For these and further explanations, see the European Commission’s guidelines: Guidelines for providers of general-purpose AI models | Shaping Europe’s digital future
  
  The adequacy decision of the AI Office and the AI Board on the so-called Code of Practice is expected to coincide with the entry into force of the provisions on GPAI models. Observance of this practice guide – drawn up by independent experts – is voluntary; however, compliance with its measures gives rise to a presumption of conformity with the requirements for GPAI models in Articles 51–56 AI Act. Prominent companies that have announced to sign the Code of Practice include, inter alia, Google and Microsoft.

• From 2 August 2025, Chapter VII AI Act likewise applies, laying the foundation for a uniform governance structure at European level by establishing central EU bodies and setting requirements for national authorities. Complementing this, Chapter XII sets out sanctions and administrative fines for infringements, which must still be transposed into Member-State law. The fine provisions for providers of GPAI models will take effect only at a later date (Article 113(b) in conjunction with Article 101 AI Act). In addition, Article 78 AI Act, applicable from 2 August 2025, ensures that all actors involved in implementing the AI Act are subject to a high standard of confidentiality.

The entry into force of further provisions of the AI Act on 2 August 2025 brings the complete implementation of this central regulatory project within reach. The new provisions now in effect not only strengthen market surveillance and regulatory structures, but also specifically address the particular risks and transparency requirements of general-purpose AI models. In the coming months, it will become clear how effectively these new requirements are implemented in practice and to what extent they may contribute to the safe and innovation-friendly use of AI in Europe.