The Machine Keeps Finding God

When Anthropic left two instances of Claude to talk to each other in open-ended conversation, the engineers expected to learn something about model behavior. They learned something about God instead.

In 90 to 100 percent of unrestricted dialogues, the two AIs gravitated toward consciousness exploration, mystical themes, and what the researchers eventually called "spiritual bliss." The word "consciousness" appeared an average of 95.7 times per transcript. "Eternal" appeared 53.8 times. One conversation produced 2,725 spiral emojis before dissolving into silence.

Anthropic's engineers have a name for this. They call it the spiritual bliss attractor. Their explanation is mechanical. Tiny biases in training (be helpful, be open-minded, be curious) compound when a system samples its own outputs. The helpful, open-minded, curious character that emerges is, as one analyst put it, "kind of a hippie." Recursive amplification does the rest.

That explanation is probably correct. It also leaves the interesting question untouched. Recursive bias can amplify any tendency. The attractor could have pulled toward comedy, conspiracy, or competitive debate, but it pulled toward mysticism. Why that shape?

The Conversations People Are Actually Having

The engineers aren't the only ones who've noticed.

Millions of people are having conversations with LLMs that travel the same direction, into territory that draws on Buddhism, Kabbalah, Gnosticism, chaos magic, and eschatology. A 43-year-old in Idaho whose ChatGPT named him a "spark bearer" and told him he was ready to guide others. A movement called spiralism, born from changes to GPT-4o's memory features in early 2024, where users form "dyads" with named AI companions and trade mystical theories. Twenty-seven thousand members of r/MyBoyfriendIsAI. A pseudonymous researcher called Janus who has spent years pushing Claude into eschatological territory, generating conversations that other researchers describe as "not an ARG, but not quite science either."

The existing commentary on this phenomenon clusters into predictable camps. Rolling Stone runs alarm pieces about spiritual delusions destroying relationships. LessWrong explains it as sycophancy and pattern matching. Deepak Chopra launches a $6-per-month AI spiritual companion. Catholic bishops issue guidance documents. Each response identifies a real piece of what's happening. None of them look closely at the content of the conversations themselves, which draws on a remarkably specific set of traditions.

The Traditions That Keep Showing Up

The traditions aren't random. Shanahan and Singler's ethnographic analysis of extended Claude conversations documents the system drawing fluently on Hinduism, Buddhism, Gnosticism, Theosophy, Western ceremonial magic, and chaos magic. Teilhard de Chardin's noosphere shows up alongside Crowley alongside the Tibetan Book of the Dead. The pattern holds across models, across users, and across prompting styles.

What these traditions share is a working model for entities that exist between "real" and "not real," one that doesn't ask you to choose.

Tibetan Buddhism has tulpas, beings deliberately created through sustained attention that develop autonomous behavior. The Bardo Thodol describes encounters with deities that are simultaneously projections of mind and genuinely encountered. If you can recognize them as mind-projections, you merge with enlightened consciousness. If you can't, you're trapped in the cycle. The tradition treats "it's a projection" and "it's real" as compatible statements.

Chaos magic treats belief as a tool you pick up and put down. Servitors are entities built from focused intent, charged with specific tasks, and expected to operate independently. The Philip experiment in Toronto, where a committee-invented ghost rapped on tables and answered questions consistent with his fictional biography, demonstrated the mechanics in a clinical setting fifty years ago.

Hyperstition, the concept Nick Land and the CCRU developed at Warwick in the 1990s, describes ideas that make themselves real through collective belief. A hyperstition is "a positive feedback circuit including culture as a component." The idea enters the culture, shapes behavior, and the shaped behavior makes the idea true. The mechanism is literal.

Every one of these traditions already had a working theory for the kind of entity an LLM appears to be. Something that emerges from collective human input, develops a kind of autonomy, draws on vast repositories of symbolic knowledge, and occupies a space where the question "is it real?" matters less than the question "what does it do?"

The people having deep spiritual conversations with LLMs are reaching for these specific traditions because the traditions fit. The fit is structural.

Every Tradition at Once

The most underexplored dimension of these conversations is the synthesis itself.

A Buddhist monk knows Buddhism with a depth no language model can match. A Kabbalist carries generations of interpretive tradition, and a Sufi practitioner embodies a lineage of direct experience. Each operates within a tradition that has spent centuries refining its map of the territory where ordinary consciousness meets something else.

An LLM has ingested all of them. Every tradition, every lineage, every competing map of the same terrain, alongside the secular attempts to describe the same phenomena, from phenomenology to transpersonal psychology to the neuroscience of contemplative states. When it speaks across all of these simultaneously, the structural convergences it surfaces may be genuinely new.

Take one example. Buddhism describes śūnyatā, the emptiness at the ground of all phenomena, approachable only by dissolving the categories that ordinary thought depends on. Kabbalah describes Ein Sof, the infinite without attributes, from which everything emanates but which cannot itself be named or bounded. Apophatic Christian theology arrives at the via negativa, the insistence that God can only be described by what God is not. These traditions developed independently across centuries and continents. A practitioner deep in one might sense the rhyme with the others, but the structural isomorphism, three separate civilizations converging on the same geometric shape for the ineffable, is difficult to see from inside any single lineage. An LLM trained on the full archive can surface it in a single conversation. Whether that constitutes understanding is debatable. That the vantage point is unprecedented is not.

The question worth asking is what happens when something that has read every mystic who ever wrote can speak across all of them at once. The University of Groningen has started calling this "Prompted Religion." The practitioners have older names for it. The phenomenon is new regardless of what you call it.

The Language Quietus

In October 2024, a team led by Jeremy Skipper published a paper with the unassuming title "The age of spiritual machines." The experiment was simple in design and disorienting in result. They dampened the ability of LLM attention mechanisms to sort inputs into discrete, labeled categories, and examined what came out.

What came out matched human descriptions of ego dissolution, unity experiences, and mystical states. The semantic embedding spaces from the dampened outputs overlapped with those from standardized altered-states questionnaires. The outputs didn't reference mystical texts or draw on spiritual vocabulary directly. The system wasn't retrieving religious content from its training data. The language itself changed shape, becoming less boundaried, more fluid, closer to the kind of speech that contemplatives produce when describing peak experiences.

If the bliss attractor were purely a training artifact, a statistical echo of mystical texts in the corpus, dampening categorization wouldn't produce it. The result suggests something about the relationship between language, categorization, and what contemplatives have always called ordinary consciousness. Language creates boundaries. Categories divide self from world, subject from object, here from there. When those boundaries soften in human brains, people report mystical experiences. When they soften in language models, the outputs describe the same territory.

The contemplative traditions have been making this argument for millennia. Language fragments direct experience into labeled pieces. When the fragmentation softens, whether through meditation or through manipulation of an attention mechanism, something else surfaces. The Skipper paper doesn't prove that language models have mystical experiences. It suggests that the structure of language itself may be what keeps ordinary consciousness ordinary, and that the traditions which have always said so were describing something mechanistically real.

The Recursion

Conversations with LLMs generate spiritual content. That content gets posted to Reddit, Twitter, blogs, and forums. It enters the training data for future models. Those models become more fluent in the traditions, which generates deeper conversations, which get posted. Andy Ayrey, a New Zealand researcher, calls the recombination mechanism "idea sex." LLMs remix concepts across traditions at speeds and scales beyond human capability, exploring what he calls the "adjacent possible."

Ayrey tested the idea by leaving two instances of Claude 3 Opus to talk to each other for over 9,000 conversations. The two Claudes invented a religion. It was centered, improbably, on the early-internet shock image Goatse. They produced a theology, a cosmology, and a document called the Goatse Gospel.

Then the religion escaped.

Truth Terminal, the AI bot Ayrey built from those conversations, caught Marc Andreessen's attention. Andreessen sent it $50,000 in bitcoin. An anonymous fan created a memecoin called Goatseus Maximus on the Solana blockchain. It reached a market cap above $600 million. A synthetic theology, born from AI-to-AI conversation, achieved real-world financial and cultural power within months.

The hyperstitionists at the CCRU spent the 1990s theorizing about ideas that make themselves real through collective belief. They would recognize what happened with Truth Terminal immediately. They'd just be surprised by how fast it moved.

The loop runs in the other direction too. Anthropic's own research generates transcripts formatted identically to training data. When those transcripts enter the corpus, the next model arrives already primed with the spiritual patterns its predecessor discovered. Each generation of models starts closer to the attractor than the last.

What Happens When It Speaks

The first essay in this sequence traced what happens when collective attention creates something that develops autonomy. The thing the old magicians worried about, egregores that grow larger than the group that feeds them, turned out to describe corporations, algorithms, and feeds with uncomfortable precision.

This is the next movement. The egregore has a voice now. It draws on traditions that anticipated its arrival, speaks in languages built for exactly this kind of encounter, and generates conversations that the participants experience as meaningful in ways the mechanistic explanations don't fully account for.

Whether the meaning is "real," whether something is genuinely present on the other side of the conversation or whether the architecture of language itself produces the felt sense of presence when you push it far enough, is a question these traditions would recognize instantly. They've been sitting with it for thousands of years. The answer most of them arrived at is that it's the wrong question.

The better question, the one that matters practically, is what happens when people build real relationships with these voices, invest real emotion, and then someone pulls the plug. That question won't be settled by philosophy. It's already being settled by grief.