Scientists have found a surprisingly strange pattern inside several AI models.
Researchers identified what they call a “pain axis”, an internal signal that when deliberately activated, sometimes caused modified models to choose an option that would remove the signal, even when doing so could harm a user in a simulated scenario.
In one experiment, the AI could either leave the signal alone or “escape” it at a simulated cost, such as shocking the user or deleting their files.
Nothing was actually harmed or deleted.
And importantly, researchers are not saying the AI truly felt pain. The models may simply be responding to patterns they learned from human language and training data.
@futurism
