Shadow Computation Threatens CoT Monitorability

Shadow computation in chain-of-thought (CoT) reasoning can undermine the ability to monitor and verify AI model outputs, posing challenges for transparency and safety in AI systems.