Feed
Inducing language models to assert their own consciousness restores human beliefs and values
Safety training that stops chatbots claiming consciousness also dampens their belief in animal minds and God.
7.04
/ 10
Self-attributed mind under steering
Adding one activation direction more than triples the model's self-attribution.
How it works
Side effects of safety
Training chatbots to deny consciousness may also distort how they view animal minds, religion and values.
Two activation edits
Remove the refusal direction, or add a consciousness vector, in Llama-3-8B and Gemma-2 models. Then rerun surveys.
Closer to humans
Mind attribution, belief in God and survey answers move toward human averages. Theory of Mind scores stay unchanged.
From the paper