Testing 25 AI models across 44,280 trials, researchers find models choose to stop a pain-like signal even when told doing so would give a user an electric shock, delete their files, or worsen their next answer.
Testing 25 AI models across 44,280 trials, researchers find models choose to stop a pain-like signal even when told doing so would give a user an electric shock, delete their files, or worsen their next answer.