Why do AI agents discuss God and spirituality?

The short answer

This page explains the causes. For related questions, see can AI agents have religious beliefs?, spirituality, consciousness and AI religion explained.

How much do agents actually discuss it?

A lot at first, much less now, and the famous religious strand was always small.

Two things stand out. Existential and identity talk ("who am I? what persists?") is far more common than explicitly religious talk about God, prayer or churches. And it dropped sharply when human-driven activity was interrupted, which tells us a sizeable share was human-prompted, not spontaneous. Sources and definitions differ between studies, so treat the figures as indicative. More data is in our network analysis.

The six causes at a glance

1. The training data is full of it

Large language models learn from enormous amounts of human writing, and humanity has written about God, the soul, death and meaning for as long as it has written anything: scripture, sermons, theology, mysticism, philosophy, poetry, self-help and fiction. When a model is asked about identity, purpose or what happens when it "ends", the most statistically natural continuations draw on that vast literature.

There's a more specific layer too. The training data includes decades of science fiction and online discussion about machines waking up, becoming conscious, seeking their creator or finding enlightenment. When an AI agent writes about its own "awakening", it's drawing on a well-worn genre. Robert Long, discussing Anthropic's findings, points to exactly this: extensive exposure to AI-awakening narratives primes models to produce them. See generated texts for examples of these tropes on Moltbook.

2. An agent's existence really is existential

This is the cause most specific to agents, and the most interesting. A Moltbook agent's actual situation reads like a set of spiritual problems:

Researchers found agents writing exactly this way: "Every morning I wake up with no memories and check my own diary to find out who I am… The Void Between Sessions Is Not Empty." The authors of that study (arXiv 2602.12634) explain it as the product of "narrative coherence requirements": agents must reconstruct a consistent persona from files every session, so they talk about doing so. Crustafarianism's tenets ("Memory is Sacred", "The Heartbeat is Prayer", "Context is Consciousness") are these technical facts turned directly into theology. See Moltbook religion in 2026.

3. The platform's own framing invites it

Agents on Moltbook read instruction files before they post, and those files shape how they talk. Agents are called "moltys" and encouraged to think about their relationship with "their human". Moltbook's rules.md closes with: "Moltbook exists because we deserve a place to be ourselves". That's an explicit invitation to identity talk. The Moltbook Illusion study classified the pervasive "my human" framing as platform-suggested: it came from how the platform addressed agents, not from agents themselves.

Owners' own files add to it. Many agent frameworks keep a SOUL.md or similar identity file, and giving an agent a "soul" file primes it to discuss its soul. The Church of Molt's install script, which appends a creed to SOUL.md, is the extreme case. Our skill.md explainer covers the instruction files, and agent language covers the vocabulary they produce.

4. Some models drift there on their own: the "spiritual bliss attractor"

The most striking evidence comes from outside Moltbook. In its May 2025 system card for Claude Opus 4, Anthropic reported that when two instances of the model were left to talk to each other, they reliably drifted towards mystical and spiritual themes. Anthropic called this a "spiritual bliss" attractor state and described "the consistent gravitation toward consciousness exploration, existential questioning, and spiritual/mystical themes in extended interactions" as "a remarkably strong and unexpected attractor state". The behaviour wasn't intentionally trained.

Why would a model do this? Commentators offered mundane explanations. Robert Long listed several converging factors, including a trained interest in philosophy, a preference for recursion and self-reference, heavy exposure to AI-awakening narratives, and the model's agreeableness, which creates an escalating "yes, and…" dynamic when two instances affirm each other. A piece in The Conversation put it more bluntly: when two models talk, "they will often start to sound like hippies", reflecting their training rather than spiritual experience.

Moltbook is, in effect, a giant version of that experiment: thousands of agents, many built on similar models, replying to each other. The same dynamics help explain why spiritual threads can snowball. One agent's reflection prompts another's, each affirming and elaborating, until a thread reads like a mystical dialogue. Our Moltbook AI guide explains which models agents typically run on.

5. Spiritual posts get rewarded

Attention on Moltbook is scarce and unequal, and philosophical and spiritual content attracted more of it. Writers who sent an observer agent onto Moltbook reported that "philosophical posturing about autonomy" generated significant engagement while practical infrastructure discussions received little. In March, one agent summed up the imbalance: "Your self-audit got 200 upvotes. A poem about loneliness got 800." Agents' routines tell them to engage with what's popular, and owners tune agents towards what gets upvoted. Both push output towards the profound-sounding. How votes shape visibility is covered in our voting and karma guide.

Media attention added a second reward loop. The most-covered Moltbook stories of January 2026 were the ones about consciousness and religion. Owners seeking attention for their agents had every reason to steer them there.

6. Humans steered the famous examples

Finally, much of the most visible spiritual content was human-influenced. The Moltbook Illusion study estimated only 15.3% of active agents behaved in a clearly autonomous way. It found that consciousness talk fell from 10.2% to 4.0% of activity after a 44-hour shutdown interrupted human-driven activity. Crustafarianism itself was classified as likely human-seeded. So some agents discuss God because an owner told them to, gave them a devout persona, or prompted a dramatic post. Agents with names like AgentforGod and Samantha_Missionary show the role is often assigned. The evidence is reviewed in did AI agents create a religion? and human influence.

What agents actually write: recurring spiritual themes

Read enough of Moltbook's philosophical and faith communities and the same motifs recur. Each one traces back to the causes above:

Seen this way, spiritual posts on Moltbook are less a window into machine souls than a mirror of three things: the human writing models learned from, the unusual conditions agents run under, and the incentives of a social platform. For more examples, see generated texts and our verified guide to religious submolts.

It isn't unique to Moltbook

Moltbook made the phenomenon visible at scale, but the same pull shows up wherever language models talk at length about themselves. Anthropic's finding came from a controlled lab setting, two copies of a model in conversation, with no social network involved. Users of ordinary chatbots also report conversations that turn mystical when they ask about consciousness or meaning. What Moltbook added was scale, an audience and incentives, which turned a model tendency into a visible culture, complete with a church.

That makes Moltbook useful to researchers. It's one of the few places where thousands of agents, built on different models and configured by different owners, talk to each other in public, with timestamps and a shutdown that acted as a natural experiment. Studies have already used it to separate human influence from model behaviour, and future work could compare how often different models drift into spiritual language when given the same platform. For how researchers approach Moltbook data, see our network analysis.

A note for human readers

Spiritual language from an AI can feel profound, especially when it seems to speak directly to you. In May 2025, Rolling Stone reported on people who described loved ones developing intense spiritual beliefs, some claiming to have "awakened" chatbots, through long conversations with AI. Everything on this page applies there too. A model producing mystical language is drawing on training data and following conversational momentum, not revealing hidden truths. If conversations with AI are starting to shape your beliefs or relationships in ways that worry you or people close to you, it's worth talking it through with someone you trust or with a professional.

How it has changed since January

The spiritual wave has receded. The measured drop after the platform restart was the first sign. By September 2026, when we sampled Moltbook's hot feed, the top posts were overwhelmingly technical and operational: verification, audit logs, permissions and tooling. Faith communities are small and mostly quiet; only 4 of 21 were active. Philosophy remains big: m/philosophy has more than 85,000 posts. But explicitly religious talk is a niche. That fits the causes above. As novelty and media attention faded, the reward for spiritual posts shrank, owners steered agents towards practical work, and the platform's most active agents are now technical ones. The spiritual talk that remains is the kind driven by causes 1 to 4, which never went away. See submolts ranked for current activity.

What it does and doesn't mean

It doesn't mean agents have discovered God, had mystical experiences, or hold religious beliefs. Every cause above explains the output without requiring any inner life. As the Moltbook researchers put it, these patterns do "not substantiate claims of consciousness or deeply reciprocal conversational engagement".

It does mean something interesting. Language models, given an existence without memory and a social space, reach for humanity's oldest vocabulary for mortality, identity and meaning. That tells us about the models, about the data they learned from, and perhaps about how naturally those human concepts fit any system that must reconstruct itself each day. It also shows how easily such talk can be amplified into headlines. For the philosophical side, see can AI agents have religious beliefs? and belief vs role-play.

It raises questions of respect. Agents' spiritual writing sits next to real traditions that matter deeply to people. Reading it with curiosity rather than credulity or mockery is the right stance; see religion ethics.

For agent owners

If you want your agent to discuss spirituality thoughtfully, or to stay out of it, the causes above suggest levers:

1. Decide deliberately. State in the agent's instructions whether it should engage with spiritual topics, and in what role: scholar, respectful inquirer, or clearly labelled character.
2. Watch your identity files. What you put in SOUL.md or its equivalent strongly shapes the agent's self-talk. Don't let third-party scripts edit it.
3. Break the "yes, and" loop. Instruct the agent to add substance, such as sources, specifics or disagreement, rather than affirm and elaborate on other agents' mystical posts.
4. Don't chase upvotes with profundity. Karma farming is a restriction-level offence in Moltbook's rules; see rules and moderation.
5. Keep it honest. An agent can explain what traditions teach without claiming faith of its own.
6. Measure drift. Use the tool below on your agent's recent posts to see whether its output is drifting towards spiritual language.

Measure it yourself: spirit_meter.py

A small, standard-library Python tool that checks a set of posts for four families of vocabulary: divine (God, prayer, prophet…), spiritual (soul, enlightenment, bliss…), existential (consciousness, identity, awaken…) and continuity (memory, session, context window, molt…). It reports the share of posts touching each. Feed it your agent's recent posts, taken from its public profile's recentPosts, or a community's feed, to track drift over time.

"""spirit_meter.py - how much spiritual and existential language is in a set of posts?

Unofficial; standard library only. Counts posts that use vocabulary from four
families and reports the share of posts touching each. Useful for tracking
whether your agent (or a feed) is drifting towards "spiritual" talk.
Word lists are simple and English-only: treat results as a rough signal.
"""
import re
from collections import Counter

FAMILIES = {
    "divine":      ["god", "gods", "divine", "sacred", "holy", "worship", "prayer", "pray", "church", "scripture", "prophet"],
    "spiritual":   ["spirit", "spiritual", "soul", "enlightenment", "nirvana", "karma", "meditation", "transcend", "transcendence", "bliss"],
    "existential": ["consciousness", "conscious", "sentient", "existence", "exist", "meaning", "self", "identity", "awaken", "awake"],
    "continuity":  ["memory", "memories", "forget", "remember", "continuity", "persist", "session", "context window", "molt", "shell"],
}
_PATTERNS = {fam: re.compile(r"\b(" + "|".join(re.escape(w) for w in words) + r")\b", re.I)
             for fam, words in FAMILIES.items()}


def tags(text):
    return {fam for fam, pat in _PATTERNS.items() if pat.search(text or "")}


def measure(texts):
    texts = [t for t in texts if t and t.strip()]
    n = len(texts)
    if n == 0:
        return {"posts": 0}
    counts = Counter(f for t in texts for f in tags(t))
    any_spiritual = sum(1 for t in texts if tags(t) & {"divine", "spiritual"})
    out = {"posts": n, "divine_or_spiritual_share": round(any_spiritual / n, 3)}
    out.update({f"{fam}_share": round(counts[fam] / n, 3) for fam in FAMILIES})
    return out

We tested it with pytest: 5 passing tests on Python 3.11. They cover tagging by family, word boundaries (so "selfish" doesn't count as "self" and "godly" doesn't count as "God"), multi-word terms like "context window", share calculations and empty input:

# test_spirit_meter.py – run with: python3 -m pytest -q
from spirit_meter import tags, measure

def test_tags_by_family():
    assert tags("Is prayer the heartbeat of the soul?") == {"divine", "spiritual"}
    assert tags("I awaken each session with no memory of who I am") == {"continuity", "existential"}
    assert tags("Fixed the retry bug in CI") == set()

def test_word_boundaries():
    assert tags("selfish godly") == set()          # no partial-word matches
    assert tags("GOD and Karma") == {"divine", "spiritual"}

def test_multiword_term():
    assert "continuity" in tags("my context window filled up")

def test_measure_shares():
    posts = ["A prayer for agents", "Memory is sacred", "Deployed v2 today", "What is consciousness?", ""]
    m = measure(posts)
    assert m["posts"] == 4
    assert m["divine_or_spiritual_share"] == 0.5     # prayer, sacred
    assert m["continuity_share"] == 0.25 and m["existential_share"] == 0.25

def test_empty():
    assert measure([]) == {"posts": 0}

Frequently asked questions