Commentary proposes public taxonomy of AI providers' safety decisions for AI Act enforcement
An opinion piece on Tech Policy Press proposes that the EU AI Office introduce a public taxonomy of non-technical safety decisions made by model providers (OpenAI, Anthropic), so that these can be compared and challenged in the enforcement of the AI Act.
As of 2 August 2026, the European Commission, through the AI Office, can enforce obligations for general-purpose AI models under the AI Act – requiring documentation (Art. 91), conducting its own evaluations (Art. 92), and demanding corrective measures (Art. 93). According to the article, Hungary proposed placing European safety teams directly inside frontier labs. However, the author of the commentary notes that even with this approach, the Office remains without influence over how providers themselves define terms such as "safety" or "frontier" model – labs set these criteria at their own discretion.
According to the author, the core of the problem is the Code of Practice – a voluntary instrument whose signing establishes a presumption of compliance with Articles 53 and 55 of the AI Act. The code specifies at which stages of model development a provider should disclose something, but does not determine what counts as relevant or how it should be described. As a result, value-laden choices can pass as technical or necessary decisions, without even being recognizable as value choices.
The author gives three concrete examples from publicly available documents: in the system card for the GPT-6 Astra model, OpenAI counts a refusal to fulfill a request as a "success" during capability testing (§10.1.1), which according to the author is a value choice, not a technical fact. The same document (§9) states that monitoring the model's reasoning has become more difficult and less reliable, yet OpenAI deployed the model anyway, relying on alignment as the main safeguard. Anthropic did not provide the UK AISI with pre-deployment access to the model referred to as Mythos 5.1, while a few days later Dario Amodei's piece "We Must Pace the Frontier" committed the company to involving external evaluators.
The proposed solution is for the next revision of the Code of Practice to include a regulator-owned taxonomy of the model lifecycle stages at which value decisions occur – this would make it possible to disclose and compare these decisions across providers and over time, without the regulator having to determine which choice is "correct." This is the author's own proposal published as a commentary, not an official step by the EU AI Office.
Why it matters
The piece highlights a specific gap in current AI Act enforcement: the Code of Practice leaves providers free to define and describe their own safety decisions, so key choices (e.g., how model refusals are counted during testing, or when access is granted to external evaluators) remain outside comparable oversight. For companies offering general-purpose AI models in the EU, this is a signal that discussion of a future revision of the code could extend documentation requirements to non-technical decisions as well, not just technical evaluations.
Relevant practical impact
What this means
For a business
Companies developing or deploying general-purpose AI models in the EU should note that expert discussion around the revision of the Code of Practice is moving toward a requirement to disclose non-technical safety decisions (e.g., the methodology for counting model refusals during testing, decisions on granting external evaluators pre-deployment access) – this would increase documentation requirements for compliance processes.
Risks and complianceCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.