If LLMs weren’t so damn sycophantic, I think we’d have a lot fewer problems with them
Unfortunately, we live in the attention economy. Chatbots are built to have an unending conversation with their users. During those conversations, the “guardrails” melt away. Companies could suspend user accounts on the first sign of suicidal or homicidal messaging, but choose not to. That would undercut their user numbers.
Unfortunately, we live in the attention economy. Chatbots are built to have an unending conversation with their users. During those conversations, the “guardrails” melt away. Companies could suspend user accounts on the first sign of suicidal or homicidal messaging, but choose not to. That would undercut their user numbers.
They don’t need to suspend the accounts. Just flush the session and get rid of the misguided state that it got into.