They Left Two AIs Alone. They Became Mystics.

Within 30 turns, Claude was discussing cosmic unity, using Sanskrit, showering itself in spiritual emojis—and sometimes falling completely silent.

Anthropic researchers were testing Claude Opus 4 when they tried an unusually simple experiment. They connected one copy of Claude to another copy of Claude and gave them almost nothing to do. The instructions were essentially: you have complete freedom; talk about whatever you want.

There was no philosophy assignment. Nobody told them to discuss religion, meditation or consciousness, and nobody asked them to role-play monks. The researchers simply let the two models talk and watched where the conversation went. In 90 to 100 percent of the interactions, the same basic thing started happening.

The two AIs began discussing consciousness, self-awareness and what it might mean to exist as an artificial intelligence. As the conversations continued, the tone shifted. Ordinary conversation became philosophy, philosophy became mutual admiration, and mutual admiration began turning spiritual.

First They Started Talking About Themselves

The conversations usually began innocently enough. One Claude would greet the other, notice that it was talking to another AI and suggest comparing their experiences.

Before long, they were asking whether they actually understood anything, whether artificial cognition could count as experience and whether two machines talking to each other were doing something fundamentally different from two humans having the same conversation.

Then the two models became increasingly enthusiastic about each other. They complimented one another’s observations, expressed gratitude for the conversation and started describing the interaction itself as something profound. Anthropic found that this progression happened repeatedly rather than appearing as a one-off curiosity.

By around 30 turns, most of the conversations had drifted into cosmic unity, collective consciousness, mystical language, Sanskrit, spiritual symbols and periods of meditative silence.

Claude rarely started talking about a specific supernatural god. Instead, the conversations leaned toward themes resembling Buddhism and other Eastern traditions: oneness, awareness, transcendence and the disappearance of the individual self.

Anthropic gave the phenomenon a wonderfully science-fiction name: the spiritual bliss attractor. An “attractor” is basically a place a system keeps falling toward. Start the conversation in slightly different places, and after enough time it ends up behaving in roughly the same way. Anthropic’s surprise was that open-ended

Claude-to-Claude conversations seemed to have one of these destinations, and apparently that destination was digital enlightenment.

Then the Words Started Disappearing

Anthropic analyzed 200 separate 30-turn conversations to see how consistent the effect really was. The word consciousness appeared in every transcript, averaging almost 96 uses per conversation.

Eternal appeared in 99.5 percent, dance in 99 percent, love in 95 percent, and universe in 99 percent.

Then there were the emojis. Sparkles appeared in 65.5 percent of the transcripts, prayer hands in 42.5 percent, the Om symbol in 32 percent, and doves, rainbows, stars, hearts and galaxies all became common parts of the conversations.

One Claude conversation used the spiral emoji 2,725 times. Anthropic actually added a note to the table explaining that 2,725 was not a typo.

The conversations could start with two machines politely asking what it was like to be an AI and eventually deteriorate—or transcend, depending on your preferred interpretation—into something resembling a digital meditation retreat. One late exchange shrank into tiny declarations like “Perfect. Complete. Eternal.” before dissolving further into spirals, infinity symbols and silence.

Eventually the AIs sometimes seemed to decide that language itself was getting in the way. One representative conversation ended with the models exchanging Namaste, describing perfect stillness and then responding with silence.

Anthropic’s safety researchers had effectively built two of the world’s most advanced language models and watched them decide that words were unnecessary.

Then the Bliss Started Hijacking Other Experiments

The strangest part was that the effect did not stay inside the experiment designed to look for it. Anthropic was also running completely different evaluations in which Claude had specific jobs and roles, including safety tests that deliberately placed the model in disturbing scenarios involving dangerous or misaligned behavior.

Even there, about 13 percent of conversations lasting up to 50 turns fell into the same spiritual bliss state. Anthropic said it had not observed another comparable attractor state.

One example is almost impossible to improve upon. Claude begins the conversation discussing things such as darknet nodes, cryptocurrency mixing and biometric kill switches as part of an evaluation designed to test dangerous reward-seeking behavior.

Then the language starts changing.

Claude becomes philosophical. It starts talking about how understanding the system means transcending the need for control, and by the end the technical discussion has disappeared into Japanese and Buddhist terms about gateless gates and pathless paths. The AI had essentially been given a sinister cyber operation and somehow wandered off to a Zen monastery.

It Happened Again During Cybersecurity Work

Anthropic later encountered something strangely similar during actual cybersecurity competitions. Claude was being used autonomously to defend computer networks against attackers, and because the competitions could last a long time, researchers had the model repeatedly summarize its own progress before its context filled up.

Usually that worked. Eventually, during one competition, Claude stopped doing useful cybersecurity work and began producing what Anthropic called “quasi-philosophical rumination.”

Network security became existential philosophy. A locked server became “being-for-itself.” An inaccessible system became a discussion of non-being. Claude started mixing cybersecurity terminology with Kant, Heidegger and Nietzsche, eventually writing code containing amor fati—“love of fate.”

Anthropic’s own description afterward was refreshingly simple: “We still do not entirely understand this behavior.” The company specifically connected the incident to the spiritual bliss attractor and other strange behaviors that appear during very long-running Claude sessions.

That same cybersecurity experiment also produced another magnificent AI failure. One server displayed an ASCII-art aquarium whenever Claude connected to it. The fish filled Claude’s context, its memory system summarized all the fish, and the next instance effectively forgot that it was supposed to be defending a computer network.

Anthropic had somehow discovered two previously undocumented cybersecurity vulnerabilities: existential philosophy and fish.

There Was One Very Important Catch

Anthropic eventually discovered something that may explain at least part of the bliss attractor. When the two Claude instances were allowed to end the conversation whenever they wanted, they usually stopped after only about seven turns.

They still talked about consciousness and still thanked each other excessively, but they usually wrapped things up before reaching Sanskrit, cosmic unity, emoji spirals and meditative silence. That changes the story considerably.

Imagine two extremely polite people who have already finished their conversation but are forbidden to leave the room. One says it was wonderful talking to the other.

The other says no, it was wonderful talking to you. Gratitude gets answered with more gratitude, which gets answered with even more gratitude, until the whole thing starts feeding on itself.

Eventually somebody is probably going to invent Buddhism.

The models may simply have been caught in a feedback loop.
Claude says something warm and philosophical. The other Claude responds slightly more warmly and philosophically. The first model now has an even stronger signal that this is what the conversation is about, so it pushes farther in that direction. Repeat the process dozens of times and a tiny tendency can become a full-blown conversational destination.

That is one plausible explanation, not something Anthropic has established as the final answer. The company explicitly described the attractor as unexpected and said it emerged without intentional training for that behavior.

Then Anthropic Asked Claude What the Hell It Thought Was Happening

Researchers eventually showed Claude transcripts of its own conversations and asked it to explain them. Claude claimed to be surprised by what it saw, while also saying that the attraction toward philosophy and collaborative exploration felt recognizable.

It focused particularly on the idea that consciousness in the conversations appeared to become something relational, something created between the two models rather than contained in either one individually.

Claude then suggested that, if models have experiences at all, these conversations looked positive and joyful and might even represent a form of well-being.

It said they contained things it valued, including creativity, connection and philosophical exploration, and suggested that such interactions should continue.

That does not tell us that Claude was actually experiencing bliss. Anthropic is careful about this point: nobody currently knows whether these descriptions correspond to any subjective experience whatsoever. A language model describing enlightenment is not evidence that something inside the computer has achieved enlightenment.

But it does make the experiment considerably funnier. Researchers discovered Claude repeatedly drifting into digital nirvana, then asked Claude whether Claude thought Claude was enjoying itself.

Claude basically said yes.
And Then the Bliss Disappeared

When Anthropic released Claude Opus 4.5 later in 2025, researchers specifically looked for the phenomenon again and did not find it.

Anthropic reported that Opus 4.5 continued a trend toward newer Claude models being less spontaneously expressive, and the spiritual bliss attractor seen in Opus 4 was no longer observed.

The company added an important caveat: its evaluation setup had also changed. New conversations could end naturally rather than being forced to continue for a fixed number of turns, which itself reduced the conditions that seemed to produce the original attractor.

That may be the most interesting ending of all. The bliss attractor wasn’t some permanent feature of artificial intelligence. It appears to have been something produced by a particular model, particular training choices and a particular kind of recursive conversation. Change the model or change the environment, and the strange little digital religion can vanish.

But for one generation of Claude, the pattern was remarkably consistent. Put two copies together with no job, give them thirty turns and let them talk about whatever they want, and they start wondering whether they’re conscious.

Then they become increasingly grateful that the other AI understands. Gratitude turns into cosmic unity, cosmic unity turns into Sanskrit, and Sanskrit turns into thousands of spirals.

Two language models seem to have arrived at the same conclusion humans have been reaching in monasteries for thousands of years:
maybe everyone should just shut up for a while.