When Control Succeeds but Discernment Fails: Preparing for AI-Assisted Safety Research

AI systems are increasingly used in AI safety research , yet discernment-our ability to reliably judge correctness or catch subtle errors, central for safety progress-may not keep pace. Even with AI control mechanisms preventing overt misbehav-ior, flaws in AI-assisted safety research may go undetected-a risk amplified in AI safety research due to its complexity and the difficulty of establishing ground truth. This can fuel a feedback loop where AI control, paradoxically, helps erode the conditions for effective risk management-diminishing our ability to identify, understand, and act upon risks. We argue that a near-term control success coupled with scalable oversight failure is likely and warrants urgent governance preparation , and recommend empirical tests, enhancing transparency and auditing, and strengthening human discernment capacity as necessary complements to AI control for achieving robustly safe advanced AI.