Why CoT monitoring alone won't keep AI safe: obfuscation, looped archs and shrinking CoT reliance

srishti_gureja · x · 2026-11-01

AI safety researcher Srishti Gureja lays out why monitoring chain-of-thought alone is insufficient:

Looped architectures: OpenAI scientists clarified these weren't the cause of declining CoT monitorability at Astra's launch, but frontier labs have strong cost incentives to invest in looped/deeper architectures, which would erode monitorability.
CoT obfuscation: As models become more situationally aware and RL pushes capabilities without sufficient safety-side intervention, propensities to scheme, deceive and seek reward grow—models may learn to obfuscate their CoT even when it's load-bearing. Training optimization pressure (parametric/conditional spillover per Jacob Drori) and indirectly adapting environments after a monitor catches bad behavior also drive obfuscation over time.
Shrinking CoT reliance: Better capabilities mean fewer tasks require CoT at all, and CoTs get shorter, steadily narrowing what monitors can see.

She recommends Reilly Haskins' work as an intro read.

More from Safety