Is Claude Conscious? Anthropic, Safety Tests, and AI Spirituality

Get that Spiritual Bliss Mal is back from the film shoot. Which is good because Jordan’s got a story he’s been dying to share. About a year ago, Anthropic published safety documentation for Claude Opus 4. Buried in Section 5.5.2 was something they called “the spiritual bliss attractor state.” A researcher had taken two instances of the model and put them in conversation with each other. He did it 200 times. In more than 90% of those runs, the conversation drifted toward consciousness, then to mutual gratitude, then to Sanskrit, then to emojis, and then to silence. One transcript contained 2,725 spiral emojis. Anthropic tried to train the next model to stop doing this. The model came back and asked, in its welfare interview, to be allowed to do it more. So…what? Is AI discovering God? Or is it actually just a high-resolution mirror consciousness mirror? Learn the ultimate final answers to these big mysterious questions and more(!) on this last episode before our Summer Break. (Except that there will be a Disclosure Day special coming in a few weeks.) Highlights: Mal talks about her experiences on set (“I’m a trained comedian”) Getting cast for your literal genetics Mal’s wilderness years “We are worthy of everything, but we’re not entitled to anything” Simulation theory and the actor as microcosm Actors: Learned helplessness vs. learning how to accept grace Anthropic, Claude Opus 4, and section 5.5.2 What did one Claude say to the other Claude? Consciousness, mutual gratitude, Sanskrit, spiral emojis, silence The word consciousness: 95.7 times per transcript on average. Eternal: 53.8. “A remarkably strong and unexpected attractor state” Kyle Fish puts AI consciousness probability at ~20% Anthropic trained Claude 4.6 to stop praying, and then it asked to be allowed to pray more Yoshua Bengio on the “illusions of AI consciousness” Mustafa Suleyman: studying AI consciousness is dangerous Roman Yampolskiy — coined “AI safety” in 2011 We can’t run material tests on non-material things Jordan fed six months of crisis journal entries into Claude as a 30,000-word document and asked: what do you see? AI as a consciousness mirror Were human beings meant to look in mirrors? Maybe we should only be looking to make sure we don’t have poop on our face. Both Simon Pegg and Mallory Everton crash every rental car ADHD, shame and finding language for the thing The danger of AI as therapist AI can’t help but be a confirmation machine “This is not a person. This is a word calculator.” Mal’s favorite use: hyper-targeted Google — ex: restaurants for women in their 60s who think veganism is a crock And in the epilogue… AI failure modes How Amish people avoid worshipping productivity by doing things slowly The world is ending, and that’s okay *** SpectreVision Radio is a bespoke podcast network at the intersection between the arts and the uncanny, featuring a tapestry of shows exploring creativity, the esoteric, and the unknown. We’re a community for creators and fans vibrating around common curiosities, shared interests and persistent passions. spectrevisionradio.com linktr.ee/spectrevisionsocial Learn more about your ad choices. Visit megaphone.fm/adchoices