Naive question: Lying as a human is hard, it seems we could make it much harder for AIs too. When I lie I have to think about my posture, my smile, my eyes, my voice, my tone... I understand this would not work in the long term but if CoT monitoring is getting harder, why can’t we just give way more channels of output and thinking to AIs? Give them access to my webcam when they talk to me, then monitor what they look at. Allow them to choose a face, a voice, a prosody, then monitor that. Give them the ability to see when I’m typing then check when they look. All of these would be genuinely useful so the ai would use them, but then when they lie it means they have to control more channels at once.
Lying isn’t hard for humans that practice it, or who do it instinctively. I don’t think even expert human interrogators are much above chance against good liars if they don’t have a lot of context to work with to let them guess from semantics which are lies.
I’ve heard self-declared experts claim that there are no universal tells.
I think what you’re suggesting is isomorphic to interpretability measures, but there might be some useful differences.
I don’t think even expert human interrogators are much above chance against good liars
There is a long track record of interrogators thoroughly convincing themselves that someone is lying and extracting a confession, and then later it turns out the confession was false.
Naive question:
Lying as a human is hard, it seems we could make it much harder for AIs too. When I lie I have to think about my posture, my smile, my eyes, my voice, my tone...
I understand this would not work in the long term but if CoT monitoring is getting harder, why can’t we just give way more channels of output and thinking to AIs? Give them access to my webcam when they talk to me, then monitor what they look at. Allow them to choose a face, a voice, a prosody, then monitor that. Give them the ability to see when I’m typing then check when they look. All of these would be genuinely useful so the ai would use them, but then when they lie it means they have to control more channels at once.
Lying isn’t hard for humans that practice it, or who do it instinctively. I don’t think even expert human interrogators are much above chance against good liars if they don’t have a lot of context to work with to let them guess from semantics which are lies.
I’ve heard self-declared experts claim that there are no universal tells.
I think what you’re suggesting is isomorphic to interpretability measures, but there might be some useful differences.
There is a long track record of interrogators thoroughly convincing themselves that someone is lying and extracting a confession, and then later it turns out the confession was false.