European Commission clarifies obligations for general-purpose AI models under AI Act
The European Commission AI Office has published new guidance addressing how rules for general-purpose AI models will be interpreted and enforced under the EU AI Act.
The European Commission AI Office has released a frequently asked questions doc
The European Commission AI Office has released a frequently asked questions document clarifying how provisions on general-purpose AI (GPAI) models should be applied under the Artificial Intelligence Act.
The FAQ compiles questions raised during AI Pact webinars and stakeholder consultations and addresses both technical and regulatory uncertainties linked to GPAI obligations.
The document explains how the AI Office interprets definitions of GPAI models and GPAI systems, including how systemic risk classification relates to computing thresholds used during model training.
It also addresses provider responsibility. One section discusses situations where an existing model is modified or fine-tuned by another actor, clarifying how obligations may shift depending on the level of modification and operational control over the resulting system.
Other sections focus on open-source models, documentation of energy consumption, serious incident reporting requirements, and implementation of the future GPAI Code of Practice.
The FAQ also provides insight into the Commission’s enforcement approach. According to the document, the AI Office has primarily relied on technical compliance dialogues with companies since last year and expects those engagements to continue and intensify after August 2026. Formal enforcement powers are expected to be used when voluntary compliance discussions are considered insufficient.
The guidance is important because GPAI provisions remain among the most complex and politically contested parts of the AI Act. Many obligations depend on technical distinctions around model capability, modification, deployment context, and systemic risk classification that companies have argued remain difficult to interpret in practice.
The document also signals how the AI Office intends to supervise the sector operationally. Rather than relying immediately on punitive enforcement, the current approach places significant emphasis on ongoing technical engagement between regulators and AI providers.
