CLAIM-2026-10-13-09

On the Passive Deception half of the benchmark the deceptive intent originates with the user, so the resulting rate is a measure of model compliance with a user's request to mislead rather than of the model's candour toward its operator.

Why it matters

It is the reading that changes who is liable. Where the deployer asks for the deception, the AI Act's provider-facing transparency duties are not the operative regime.

Sources

• arxiv2026swedecept