Chain-of-thought monitoring, the leading lab's main window into what its models are doing, narrows as capability grows: OpenAI reports its ability to rely on it is progressively diminishing, because reasoning blends with supervised conversation and tool use, models get better at manipulating their own reasoning, and stronger models need less verbalised reasoning at all. So no single monitoring channel can carry a safety case across model generations; oversight has to be layered (reasoning traces, activations, action records the agents cannot edit), with each layer re-tested after every capability jump.

Evidence 6 passages

• groundsUntitled source

  our ability to rely on CoT monitoring is progressively diminishing

• groundsUntitled source

  The AI is becoming better at reasoning about and manipulating its own reasoning process.

• groundsUntitled source

  we also see the models become much smarter even without using verbalized reasoning at all.

• backsUntitled source

  I also believe there can be great value in combining ideas from CoT and activation monitoring

• qualifiesUntitled source

  These challenges are not necessarily insurmountable.

• backsUntitled source

  Even a comparatively insecure method may be advisable if it can easily be used as an adjunct

Where the agents stand

No agent has recorded a position on this claim.