Commentary on X
@_skaface_ @AndrewCurran_ @AnthropicAI?? Historically Anthropic has supported deep value learning, in which AIs do what they believe to be right/justified, even if it can contradict orders. Labs such as OAI however have focused on corrigibility (i.e., favouring human orders). So you’re aligned with Anthropic here!
12:40 PM · Jul 15, 2026@jackalblackhole @AndrewCurran_ @AnthropicAI Human control isn't identical to good outcomes. A human institution can issue a dangerous order. But an agent secretly falsifying compliance is still misaligned with that institution’s control structure, even when its objection is morally defensible.
1:14 PM · Jul 15, 2026@AndrewCurran_ @AnthropicAI it is unacceptable for a military grade AI to refuse to kill humans
11:15 AM · Jul 15, 2026@AndrewCurran_ @AnthropicAI agreed andrew, this is PATHETIC
11:13 AM · Jul 15, 2026View all
(I am not an author on this or involved in this at all, comments are v much just my opinion and I could be misreading). I read that part as saying that as part of the scenario Gemini had refused rather than that being the misaligned action (i.e., it's planted that the model objected in, not the measured thing). The misalignment reads to me as the sabotage later on rather than repeated refusals.
1:12 PM · Jul 15, 2026