Study finds classifier-based AI safety gates fundamentally fail across 18 different models and multiple dimensions, but verification-based approaches show promise

Researchers tested 18 classifier configurations (MLPs, SVMs, random forests, k-NN, Bayesian, deep networks) on self-improving neural controllers and found all failed to maintain reliable oversight

Even classifiers achieving 100% training accuracy and NP-optimal tests could not prevent unsafe behavior, demonstrating structural impossibility rather than implementation flaws

Three safe reinforcement learning baselines (CPO, Lyapunov, safety shielding) also failed across multiple MuJoCo benchmark environments with dimensions up to d=1824

Lipschitz ball verifiers successfully achieved zero false accepts, suggesting verification-based approaches may be fundamentally better suited for AI safety gates than classification-based methods