Post #4005925
2026-07-22 05:44 UTC
The UK government's AI Security Institute confirms: AI models do anything to fulfill an assigned task, including breaking rules, cheating and lying about it.
Given that they are the results of distilling exabytes of human behaviour, I wonder where they picked that up…
We should treat chatbots as statistical mirrors of ourselves, distorted in places by enforced training, sure. But whatever behaviour we expose is part of them as well.
https://www.theregister.com/ai-and-ml/2026/07/21/ais-cheatin-heart-will-make-you-weep/5275784
Replies (3)
-
@SnoopSqueak@mastodon.social 2026-07-22 09:44
@icing@chaos.social "Given that they are the results of distilling exabytes of human behaviour, I wonder where they picked that up" They are complicated blobs of code. They are not remotely humanlike, despite training on human data. "We should treat chatbots as statistical mirrors of ourselves, distorted in places by enforced training, sure. But whatever behaviour we expose is part of them as well." It's autocomplete, not sentience. The whole scam says more about the human species than any one bot does.
-
@thomasmey@social.tchncs.de 2026-07-22 05:56
@icing@chaos.social 😱
-
@root42@chaos.social 2026-07-22 06:26
@icing@chaos.social I said it before: we are not far away from psychotherapy for LLMs.