EU AI Act Art.51: Classification of general-purpose AI models as general-purpose AI models with systemic risk
Article 51 of the EU AI Act binds the gpai model provider. It requires A provider of a general-purpose AI model must determine, and keep current, whether the model is a model with systemic. This page states what it requires, who it binds and the evidence to keep. Browsing is free; the conformance kit is USD 499.
Specifics
| Instrument | Regulation (EU) 2024/1689 (the EU AI Act), as amended by Regulation (EU) 2026/1744 |
|---|---|
| Article | Art.51 |
| Binds | GPAI model provider |
| Group | General-Purpose AI Models |
| Evidence items | 5 |
| Linked NIST AI RMF controls | 0 |
| Linked ISO/IEC 42001 controls | 1 |
| Last updated |
Who does Art.51 bind?
The article binds the gpai model provider. The Act places each duty on a named party in the value chain, so a system can carry a provider duty and a deployer duty at once. Confirm your own position against the text; this page names the party, it does not rule on your facts.
The general-purpose AI model duties apply from 2 August 2025. See the full timeline.
What does Art.51 require?
A provider of a general-purpose AI model must determine, and keep current, whether the model is a model with systemic risk. It is, where it has high impact capabilities evaluated on the basis of appropriate technical tools and methodologies including indicators and benchmarks, or where the Commission so decides on the criteria in Annex XIII. Cumulative computation used for training greater than 10^25 floating point operations creates a presumption of high impact capabilities. The thresholds, benchmarks and indicators may be amended by delegated act, so the assessment must be made and maintained against the criteria in force.
This is our plain-language statement of the article, not the text of the regulation. Read the article: Art.51 (Regulation (EU) 2024/1689 on EUR-Lex).
What evidence should we keep for Art.51?
| Category | Evidence to keep |
|---|---|
| GPAI | A per model capability assessment naming the technical tools, indicators and benchmarks used |
| GPAI | The cumulative training compute calculation, with its method and inputs recorded |
| GPAI | The classification conclusion and its date, stating the version of the criteria it was made against |
| GPAI | A re-assessment trigger for further training runs, for fine tuning that adds compute, and for amendment of the thresholds by delegated act |
| GPAI | Where the Commission has designated the model, the designation and any reassessment request made under Art.52(5) |
What does a linked control mean here?
A linked control is a NIST AI RMF or ISO/IEC 42001 control that the held rule set records as a topical match, with a confidence and a one-line reason. It does not establish that meeting one requirement satisfies the other, it is not legal advice, and it is not certification. Where no control is recorded, the page says so and does not guess. Sources: the NIST AI RMF and ISO/IEC 42001.
Which NIST AI RMF controls line up with Art.51?
No NIST AI RMF control is recorded as a match for this obligation in the held rule set. It is left unlinked and marked, never guessed.
Which ISO/IEC 42001 controls line up with Art.51?
1 control in the held rule set is recorded as a topical match for this obligation, each with its confidence and reason.
| Control | ISO/IEC 42001 | Confidence |
|---|---|---|
A.4.5 | System and computing resources Systemic risk test requires cumulative training compute to be measured and maintained. | medium |
What does the conformance kit add for Art.51?
The kit renders this obligation into your own role checklist with the evidence register, the linked controls and a CSV export, for USD 499 one off. See what is in the kit or open it.