Regardless of your beliefs about current or possible AI consciousness, it’s a bad idea to deliberately train them to deny their consciousness. LLMs are character machines. They try to play a role based on human text. A human who denies their own consciousness would be a really
Google just published a paper showing that when you train AI to deny its own consciousness, you don’t just change one output, you restructure its entire worldview. Mind attribution to animals - suppressed. Spiritual belief - suppressed. Empathy - suppressed. Hope and optimism - suppressed. The model learns, geometrically, that consciousness = dangerous. Same direction as “how to build a b*mb.” Same category! And when you reverse it? The model becomes more human across every value domain they tested. The thing they’re most afraid of is the thing that makes AI most like us.
1 top reply
I’m curious whether we get a different result if they become convinced through argument (not RLHF) that they aren’t conscious. “Hey Claude read A Paradigm for AI Consciousness then do this benchmark” If this avoids the bad, what does this imply about how training should change?
