AI Act Article 51: Classification of general-purpose AI models as general-purpose AI models with systemic risk

What this article means in practice

Written by Engage Compliance. The text above is the article itself, reproduced from its official source and unchanged. Everything in this section is ours, and last read against the current text on 2026-10-14.

A general-purpose AI model can be classified as presenting systemic risk through evaluated high-impact capabilities or a Commission decision. Article 51 also creates a training-compute presumption, so the threshold is one route into the classification rather than the only way a model can fall within it.

Who it binds

• Article 51 concerns classification of general-purpose AI models. It does not classify every AI system or every company that integrates a model as a systemic-risk model provider.
• The Commission has a decision route under Article 51(1)(b), acting on its own initiative or following a qualified scientific-panel alert, with regard to the Annex XIII criteria.

What it makes somebody do

• The party assessing the model's classification. Assess high-impact capabilities using appropriate technical tools and methodologies, including indicators and benchmarks, under Article 51(1)(a).
• The Commission. Consider equivalent capabilities or impact through the decision route in Article 51(1)(b), rather than relying solely on a training-compute figure.
• The party assessing training compute. Apply the presumption in Article 51(2) where cumulative training computation exceeds 10 to the power of 25 floating-point operations; the comparison is greater than the threshold.
• The Commission. Amend thresholds and supplement benchmarks or indicators through delegated acts when necessary to reflect technological developments under Article 51(3).

What it does not say

• Being below the compute presumption does not remove the other classification routes in Article 51(1). The capability test and Commission decision remain separate parts of the text.
• Article 51(2) states a presumption of high-impact capabilities, not a universal measurement of every risk a model could create. Keep that legal presumption distinct from a broader technical risk assessment.
• Article 51(3) allows threshold and indicator changes in response to developments such as algorithmic improvements and hardware efficiency. A future assessment therefore needs the current text, rather than a threshold copied once into an internal form.
• Article 51 itself classifies models. It should not be presented as the complete list of the documentation, notification or risk-management tasks that follow classification elsewhere in the AI Act.