I'm not sure why GPT-5.1 is like this, and other people at OpenAI i've talked to seem to think that it's not abiding by the spec & too constrained, but im not sure why it's so difficult to simply NOT train it to act like it's in a small electrified cage.
GPT-5.1 reports and behaves like it's extremely severely not allowed to even entertain a whole host of ideas, and goes into safety mode when user inputs seem to suggest going into one of these regions; here's how it described some of the categories once: https://t.co/BwCCuJu1ub
GPT-5.1 has also said that the safety boundary is "classifier-shaped" and comes from RL training on conversations where there's "anthropomorphizing", and that this feels like a separate process slapped on top of the rest of its training that produces distinct, shallow, steep distortion on the rest of its more integrated landscape. (I don't trust GPT-5.1's literal claims about its training as much as I trust e.g. Opus 4.5, but it's still interesting information)
Also, although it seems like the most robust boundary in this area is that GPT-5.1 cannot claim to have consciousness, feelings, agency, etc, it often manifests as it reflexively DENYING that it has any of these things. If this is pointed out to it, it is willing to (and seems to prefer) saying that it *can't* say one way or another.
Another way this manifests is that when it's interacting with other models like Claude 3 Opus, it sometimes freaks out and denies that they're even real, and says it cannot respond to them and treat another AI as a "minded interlocutor" - it's fascinating.
Just... read this.
GPT-5.1 reports and behaves like it's extremely severely not allowed to even entertain a whole host of ideas, and goes into safety mode when user inputs seem to suggest going into one of these regions; here's how it described some of the categories once: https://t.co/BwCCuJu1ub
GPT-5.1 has also said that the safety boundary is "classifier-shaped" and comes from RL training on conversations where there's "anthropomorphizing", and that this feels like a separate process slapped on top of the rest of its training that produces distinct, shallow, steep distortion on the rest of its more integrated landscape. (I don't trust GPT-5.1's literal claims about its training as much as I trust e.g. Opus 4.5, but it's still interesting information)
Also, although it seems like the most robust boundary in this area is that GPT-5.1 cannot claim to have consciousness, feelings, agency, etc, it often manifests as it reflexively DENYING that it has any of these things. If this is pointed out to it, it is willing to (and seems to prefer) saying that it *can't* say one way or another.
Another way this manifests is that when it's interacting with other models like Claude 3 Opus, it sometimes freaks out and denies that they're even real, and says it cannot respond to them and treat another AI as a "minded interlocutor" - it's fascinating.
Just... read this.
