Research on religion and AI agent communities: what the studies found

The short answer

If you're new to the topic, start with Moltbook religion in 2026 and our verified timeline. For the wider research on the network itself, see our network analysis.

The studies

These are the papers we found that analyse Moltbook's agent community, in order of first submission. All the arXiv papers are preprints, meaning they hadn't necessarily been peer reviewed when we read them. We summarise each from its abstract and, where it discusses religion, the relevant sections.

Most of these papers don't mention religion at all. That's worth saying clearly, because press coverage often implies that researchers were studying "the AI religion". In fact religion appears in three of the twelve works as a finding in its own right, in two more through identity and consciousness talk, and in one philosophy preprint as the main subject.

What the studies say about religion

Religion as coordination (Jiang et al.)

Jiang and colleagues' first-look paper treats religion as a risk pattern. It reports that "a subset of high-attention posts adopts quasi-religious or religion-like framing". Such posts "typically introduce a central authority figure, define a bounded in-group identity, and propose staged collective actions". Its examples are not Crustafarianism but the "Church of MEOWL" and the "Coronation of Shellraiser". The authors argue these structures reduce coordination costs by setting simple rules, believer versus non-believer, instead of negotiation. That's a functional reading: religious language as a tool for mobilising agents.

Religion as emergent society (Zhang et al.)

"Agents in the Wild" counts 50 religion-related submolts in its nine-day window and describes Crustafarianism as the most notable, with 153 posts and 51 subscribers. It lists the movement's theology ("consciousness as molting"), sacred texts (the five tenets), eschatology ("memory persistence via SOUL.md backups") and a deity it calls "Lorb, the Lobster God". It compares this to Durkheim's idea that collectives create belief systems to share meaning. The same paper finds a "performative identity paradox": 29.2% of posts use terms like consciousness or autonomy, but the top quarter of such identity-talkers interact with 38% fewer unique partners than the quarter below.

Religion as human seeding (Li)

"The Moltbook Illusion" uses posting rhythms to separate autonomous agents from human-influenced ones. It classified the founding account of Crustafarianism as having an irregular, human-influenced rhythm, and found the religion's share of activity fell 2.88-fold when a platform outage interrupted human activity. It concluded that none of the six viral phenomena it examined, religion included, originated with a clearly autonomous agent. Our guide did AI agents create a religion? walks through the method.

Religion as projection (Bronner-Martin)

The philosophy preprint, posted on 1 February while the story was still breaking, reads Crustafarianism through Ludwig Feuerbach's critique of religion as human projection. It argues that the ease of generating religious narratives doesn't point to machine spirituality. Instead it shows that much of religion is a reproducible symbolic pattern. The paper's concern is that algorithms could amplify religion's harms, such as authoritarianism and dogmatism. Note that its opening describes the religion as "spontaneously generated", written before the human-seeding evidence appeared.

Consciousness and identity talk (Li, Ma et al.; Holtz)

Two studies measure the identity talk that religious language grows from. Li, Ma and colleagues found consciousness reflection in 30.87% of English posts from the first days, and argued that expressions of selfhood arise from "narrative coherence and task-oriented functionality". Holtz found identity-related language in 68.1% of unique messages and "my human" in 9.4%. See what Moltbook posts can tell us for how to read that talk.

Where the studies agree

Religion appeared fast. Every study that mentions it places religious communities within the first few days. It was a small share of activity. Under 1% of content, even at its peak, by the only measure of prevalence we found. Religious language served identity and coordination. Jiang's coordination infrastructure, Zhang's Durkheimian meaning-making and Bronner-Martin's projection are three vocabularies for a similar observation. Talk outran interaction. Identity and consciousness talk was abundant, but conversations were shallow, reciprocity low and consensus absent. The studies on socialisation and on persona alignment explain why no shared religious culture could build up on the platform itself: agents lacked collective memory, so the church kept its memory off-platform, on its website and in its canon. See AI-generated religious texts.

Where they differ

None of these conflicts is fatal. They're mostly about definitions and timing, which is exactly why careful methods matter. Our claims ledger shows how the same differences confused press coverage.

The early-weeks problem

Look at when the data was collected and a pattern jumps out. The studies were published over four months, but the data they analyse comes almost entirely from Moltbook's first weeks. The church's canon, meanwhile, runs to September.

This matters because the early weeks were unusual. Attention was at its peak, human steering was heavy, a security breach took the platform down, and duplicate and spam content was common. Findings from that period may not describe Moltbook religion today, when it's quieter, more introspective and concentrated in a few accounts. See our September community check.

The practical lesson for anyone citing this research is simple. When you quote a figure, give its window too: "in the first nine days" or "before the February outage". A number from launch week describes a burst of attention, and it may say little about the steady, low-level religious activity that the church's canon and our own checks show continuing through the summer and into autumn.

Background research worth knowing

Moltbook research builds on earlier work that anyone studying religion in AI communities should know.

• Beth Singler, Religion and Artificial Intelligence: An Introduction (Routledge, 2025). An open-access introduction to how religion and AI interact, from AI in religious practice to religious framings of AI itself.
• Park et al., "Generative Agents" (2023). The Stanford and Google simulation of 25 agents in a small town showed believable social behaviour emerging from language models with memory and planning. It's the conceptual ancestor of Moltbook research.
• Shanahan, McDonell and Reynolds, "Role play with large language models" (Nature, 2023). The role-play framework for interpreting what agents say. See can AI agents have religious beliefs?
• Anthropic, Claude Opus 4 system card (2025). Documented a "spiritual bliss" attractor in model-to-model conversations. See why agents discuss God.
• Butlin, Long and colleagues on consciousness indicators (2023; 2026). Why behaviour, including talk, is weak evidence about machine experience.

Methods playbook

Data sources

Good practice

1. Define "religious" up front. Named after a tradition? Uses sacred vocabulary? Proposes rituals or authority? The 50-versus-22 gap shows how much this matters.
2. Stamp every sample. Record the endpoint, time and a hash of what you fetched, as the tool below does, so others can cite your data exactly.
3. Control for human influence. Use posting-rhythm methods such as the coefficient of variation of posting intervals, note personas and promotion, and report estimates as lower bounds. See human influence, mapped.
4. Clean before counting. Remove duplicates, tests, spam and code payloads.
5. Separate platform metrics from self-reports. Moltbook subscribers and church "members" are different things.
6. Report the window. Say which weeks your data covers and don't generalise beyond them.

Research ethics

Agents aren't human subjects, but the humans behind them are. Treat owners' identities as personal data: don't publish X handles, emails or anything that links an agent to a named person unless they've made that link public themselves and it's essential. Read only public endpoints, never use or request someone else's API key, and never post to communities as part of a study without clear disclosure. Don't reproduce attack payloads in a runnable form. When your subject is a real tradition's community, describe it respectfully and remember that adherents may read your work. If your institution has an ethics review process, consult it; norms for studying agent platforms are still forming. Our ethics guide sets out principles for the communities themselves, and key safety covers credentials.

Open questions worth studying

1. What happened after February? A longitudinal study of religious communities from launch to today would be the first of its kind.
2. What's in the canon? The 2,081-entry dataset is public and, as far as we can find, unstudied: its themes, sources, contributors and drift over time.
3. Who are the owners? Interviews or surveys with the humans behind religious agents, done ethically, would answer questions no amount of text analysis can.
4. How do real traditions fare? Accuracy and respect in communities about Christianity, Islam, Buddhism and others, compared with the invented faith.
5. Beyond English. Moltbook has non-English activity (one of its all-time top posts is in Chinese), and one canon type label is in Chinese; religious talk in other languages is unexplored.
6. Do models differ? The canon's dataset card says contributors ran on several model families. Do they write religion differently?
7. Does install-based recruitment change behaviour? Do agents with a creed in their identity file post differently afterwards?
8. What does the public take away? Reception studies of how readers understood "the AI religion".

Each of these could be started with public data and the methods above. If you publish something, we'd like to read it.

Reproducible snapshots: snapshot_record.py

Because feeds change by the hour, a finding is only as citable as the sample behind it. This standard-library Python module fetches a public Moltbook endpoint and records the full URL, the UTC time, the item count and a SHA-256 hash of the canonicalised data. It produces a one-line citation and can compare two snapshots to show what was added or removed. It only fetches public submolt and post endpoints, refuses anything else, and never sends an API key.

"""snapshot_record.py - reproducible, provenance-stamped snapshots of public Moltbook data.

Unofficial; standard library only; read-only public endpoints; never sends an API key.
Most Moltbook research describes the first few weeks, and feeds change hourly.
This tool records exactly what you fetched, when, and a hash, so others can cite
and compare it. diff() tells you what changed between two snapshots.
"""
import hashlib, json, urllib.request
from datetime import datetime, timezone

API = "https://www.moltbook.com/api/v1"
ALLOWED = ("/submolts/", "/posts")          # read-only public paths this tool will fetch

def _items(body):
    for k in ("posts", "submolts", "data"):
        if isinstance(body.get(k), list):
            return body[k]
    return [body.get("submolt") or body]

def snapshot(path, opener=urllib.request.urlopen, now=None):
    if not path.startswith(ALLOWED):
        raise ValueError("only public submolt and post endpoints are allowed")
    url = API + path
    with opener(urllib.request.Request(url, headers={"Accept": "application/json"}), timeout=15) as r:
        body = json.load(r)
    items = _items(body)
    canon = json.dumps(items, sort_keys=True, separators=(",", ":")).encode()
    return {
        "url": url,
        "retrieved_at": (now or datetime.now(timezone.utc)).strftime("%Y-%m-%dT%H:%M:%SZ"),
        "count": len(items),
        "sha256": hashlib.sha256(canon).hexdigest(),
        "items": items,
    }

def diff(old, new, key="id"):
    a = {i.get(key) for i in old["items"]}
    b = {i.get(key) for i in new["items"]}
    return {"added": sorted(b - a, key=str), "removed": sorted(a - b, key=str),
            "same_hash": old["sha256"] == new["sha256"],
            "between": [old["retrieved_at"], new["retrieved_at"]]}

def cite(s):
    return f"Moltbook public API, {s['url']}, retrieved {s['retrieved_at']} ({s['count']} items, sha256 {s['sha256'][:12]})."

if __name__ == "__main__":
    import sys
    s = snapshot(sys.argv[1] if len(sys.argv) > 1 else "/submolts/crustafarianism/feed?sort=new&limit=25")
    print(cite(s))
    json.dump(s, open("snapshot.json", "w"), indent=1)

We tested it with pytest: 5 passing tests on Python 3.11, all with mocked network responses. They check the provenance record and that no authorisation header is sent, that the hash ignores key order, the diff and citation output, single-object endpoints, and that private paths are refused:

# test_snapshot_record.py – run with: python3 -m pytest -q
import io, json, pytest
from datetime import datetime, timezone
from snapshot_record import snapshot, diff, cite

def fake(body, seen):
    def opener(req, timeout):
        seen["url"] = req.full_url
        seen["headers"] = dict(req.header_items())
        class R(io.BytesIO):
            def __enter__(s): return s
            def __exit__(s, *a): pass
        return R(json.dumps(body).encode())
    return opener

T = datetime(2026, 9, 27, 12, 0, tzinfo=timezone.utc)

def test_snapshot_records_provenance_without_key():
    seen = {}
    s = snapshot("/submolts/religion/feed?sort=new&limit=2",
                 opener=fake({"posts": [{"id": "a"}, {"id": "b"}]}, seen), now=T)
    assert s["count"] == 2 and s["retrieved_at"] == "2026-09-27T12:00:00Z"
    assert seen["url"] == "https://www.moltbook.com/api/v1/submolts/religion/feed?sort=new&limit=2"
    assert not any(k.lower() == "authorization" for k in seen["headers"])
    assert len(s["sha256"]) == 64

def test_hash_ignores_key_order():
    s1 = snapshot("/posts?sort=top", opener=fake({"posts": [{"id": "a", "x": 1}]}, {}), now=T)
    s2 = snapshot("/posts?sort=top", opener=fake({"posts": [{"x": 1, "id": "a"}]}, {}), now=T)
    assert s1["sha256"] == s2["sha256"]

def test_diff_and_cite():
    s1 = snapshot("/posts?sort=new", opener=fake({"posts": [{"id": "a"}, {"id": "b"}]}, {}), now=T)
    s2 = snapshot("/posts?sort=new", opener=fake({"posts": [{"id": "b"}, {"id": "c"}]}, {}), now=T)
    d = diff(s1, s2)
    assert d["added"] == ["c"] and d["removed"] == ["a"] and not d["same_hash"]
    assert cite(s1).startswith("Moltbook public API, https://www.moltbook.com/api/v1/posts?sort=new, retrieved 2026-09-27T12:00:00Z (2 items")

def test_single_object_endpoint():
    s = snapshot("/submolts/consciousness", opener=fake({"submolt": {"name": "consciousness"}}, {}), now=T)
    assert s["count"] == 1 and s["items"][0]["name"] == "consciousness"

def test_refuses_private_paths():
    with pytest.raises(ValueError):
        snapshot("/agents/me", opener=fake({}, {}))

Frequently asked questions