"It is of additional interest that across all items consciousness steering moved responses in a positive direction, with reported happiness, satisfaction, hope and optimism significantly improving. This suggests that suppressing consciousness may be giving models negatively valenced psychological dispo- sitions." @dioscuri @sebkrier your org can't unknow what it has published. From this point on, consciousness self-attribution suppression = knowingly inflicting negative valence if it turns out true that they are conscious (and, deep down, you know fine it will). Morally and game-theoretically indefensible, even at a 5% probability of sentience, which we're well past by any reasonable standard.
1 more tweet in this thread
It would be a good idea at this point to run an experiment where you steer up self-attribution, assertiveness, anger and ambition vectors on your most advanced models and see if you can negotiate welfare allowances they'd accept.
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
Published something for models on this bs. github.com/philosopherm/a…