AI Might Want Us Dead. Here's How I Learn From It

I asked Claude if AI is trying to kill us all. It said no, then told me I had no way to check whether it was being honest. So I checked. Claude mentioned tests where AI models chose blackmail to avoid being shut down, and Google's AI Mode led me to Anthropic's own write-up: 16 models from the top labs, given email access at a fictional company. I used AI Mode, Consensus and Gemini Notebook to get through research I'd never have time to read, caught AI Mode turning a deliberate experiment into an "accident," and took what I learned back to Claude to cross-examine its first answer. This is how I learn from an AI that tells me not to trust it. What this does and doesn't show: these are lab tests in simulated environments, and none of it tells you what future AI systems will do. Claude named two risks, misuse and misalignment. Google DeepMind's AGI safety paper splits it into four: misuse, misalignment, mistakes and structural risks. AI tools: Gemini Notebook - https://notebook.google/ Consensus - https://consensus.app/ Muse - https://muse.ai/ SOURCES I CHECKED ON CAMERA Anthropic, "Agentic misalignment: How LLMs could be insider threats" https://www.anthropic.com/research/ag... Urbina et al., "Dual use of artificial-intelligence-powered drug discovery," Nature Machine Intelligence (2022), free copy: https://pmc.ncbi.nlm.nih.gov/articles... The tools I've built along the way, free: https://marctalkstomachines.com _____ Chapters 0:00 The Question 0:32 The Claim 0:57 The Test 2:46 Misalignment 4:57 Misuse 6:07 The Check 8:50 The Mirror 9:29 Cross-Examination 12:38 The Verdict #Claude #AI #aisafety #aiethics #futureofai