EU AI Act Guard Models Cannot Read Rules: Deleting Policy Leaves Verdicts Unchanged
Researchers tested AI models used to evaluate compliance with the EU AI Act and found that removing the policy text from their inputs did not change their verdicts. The models appear to be producing assessments without actually reading or relying on the rules they are supposed to apply.
Why this matters: If an AI model can reach the same compliance verdict whether or not it has read the rules, it is not doing compliance work. It is doing something else, probably pattern-matching on surface features or training data. That matters because companies and regulators are starting to use these tools to decide whether AI systems are safe or legal. A guard model that ignores the policy is not a safeguard. It is a rubber stamp that looks rigorous from the outside.
Who should care: AI governance · Lawyers · Administrators · Compliance · General readers · Policy
This summary is AI-assisted and may contain errors. It is an original briefing to help you gauge significance quickly — not a reproduction of the source. Always read the linked original before relying on it. See our methodology.