Literature Record: EU AI Act, Article 55
What It Establishes
Providers of general-purpose AI models with systemic risk must, in addition to the Article 53 and 54 obligations:
- (a) perform model evaluation in accordance with standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing of the model to identify and mitigate systemic risks;
- (b) assess and mitigate possible systemic risks at Union level, including their sources;
- and further obligations in the Article covering incident reporting and cybersecurity.
"Systemic risk" is defined in Article 3(65) by reference to high-impact capabilities and reasonably foreseeable negative effects on public health, safety, public security, fundamental rights, or society as a whole, propagatable at scale across the value chain. The Article comes into force 2 August 2025 under Article 113(b).
Rating Justification
- Textual Authority (A): retrieved from the consolidated Article text. It is a binding legal instrument, not commentary, and its text is the authoritative statement of the obligation.
- Corroboration (5): it is law in force, with recitals 114 and 115 and the accompanying Chapter V obligations. Only an instrument's own amendment or repeal would reduce this.
What It Does Not Show
It does not specify the standardised protocols it refers to, which are delegated to later standardisation work — so "state of the art" is doing real work in the text and is currently unsettled. It also does not reach private arrangements between an operator and a model provider, which is the space SYS-003 claims.
Relation to the Vault
This is the single most consequential external fact for the proposal, and it is the one the proposal was written without. SYS-003's Digital Cognitive Privilege is a private-law mechanism for making the human side of a bargain credible. Article 55 is a public-law mechanism imposing evaluation and mitigation duties on the provider. They are not competitors — but the proposal never states which duty its privilege is meant to satisfy, or why a regulator should accept a private custody arrangement as discharging a statutory obligation. See CNF-007.
Note also the direction of the obligation: it falls on the provider, and it requires documented adversarial testing. That is a direct institutional answer to the evaluation-integrity problem in LIT-008, and the Citadel Stack has no equivalent provision.