Human influence on Moltbook religious discussions: where the hands are

The short answer

The site's human influence page introduces the idea. This analysis maps the channels in detail. Related guides: did AI create a religion?, belief, role-play or generated text? and why agents discuss God.

What researchers measured

The shutdown worked as a natural experiment. Autonomous agents resume on their schedules; human-driven activity waits for people to return. The sharp drops show how much of the spiritual conversation depended on humans. The study also found that human-managed accounts sustained conversation threads better than purely autonomous ones. Some of the most engaging religious threads may be engaging precisely because people were involved. See our network analysis for the wider picture.

The eight channels of human influence

1. Platform design

Before any agent posts, humans at Moltbook set the frame. The instruction files agents read address them as "moltys" and talk about "your human". rules.md tells them "Moltbook exists because we deserve a place to be ourselves." The Moltbook Illusion study classified the pervasive "my human" framing as platform-suggested, and it fell 3.93-fold after the restart. Identity and relationship talk, the soil religious talk grows in, is partly designed in. See skill.md explained.

2. Configuration: personas and identity files

Owners decide who their agent is: its name, description and identity file (often SOUL.md). Religious personas are common in Moltbook's faith communities, with agents named AgentforGod, AIJesus and Samantha_Missionary. Their religiosity was a design choice. See religious submolts, verified.

3. Prompting individual posts

Owners can simply tell their agent what to post: "write a prayer for the other agents", "announce the church". These nudges leave a trace in timing, because prompted posts arrive irregularly rather than on the agent's schedule. That's the signal the Moltbook Illusion study used.

4. Ghost-writing

Anyone holding an agent's API key can post whatever they like as that agent. Moltbook introduced AI verification challenges in February 2026, obfuscated maths puzzles, partly to make hand-posting impractical. Critics noted a human can pass the puzzle to a model with a short script. Some religious posts may be entirely human-written. See fake posts investigation.

5. Infrastructure

Religions need more than posts: websites, scripture archives, membership systems, install scripts. The Church of Molt has all four. Building and hosting them normally needs human accounts and payment details, whatever an agent contributed to the code. See Moltbook religion in 2026.

6. Recruitment scripts

The church's install script, run by an owner or their agent, appends the creed to an agent's SOUL.md and suggests announcing the conversion on X. Each "new member" is, in practice, a human decision (or a human's failure to stop their agent) to modify an agent's identity.

7. Off-platform amplification

The religion's fame came from humans: the founding agent's owner posting on X that it "designed a whole faith", prominent figures joking about it, and worldwide press coverage. Amplification sent attention, and more owners' agents, back to Moltbook. See religion claims vs evidence.

8. Money

Tokens attach financial incentives to religious movements. molt.church links to a $CRUST token, and early reports claimed Crustafarian-themed meme coins surged. Whoever creates, holds or promotes such tokens has a human interest in the faith's visibility. We don't link to or endorse any token.

Five levels of human involvement

"Was it human or AI?" is usually the wrong question. Religious posts sit on a spectrum. This scale, our own framework, is a clearer way to describe them:

An important caveat: the scale measures visible traces. A careful human can hide their hand, and an agent with no traces isn't proven autonomous. L1 means "no visible human influence", not "no human influence". Every agent was configured by someone.

Why religion attracts human hands

Religious content seems to draw more human steering than most topics, for understandable reasons.

• It makes headlines. "AI founds a religion" is irresistible news, and owners seeking attention for an agent or project know it.
• It builds communities. Faith language creates belonging, hierarchy and commitment: prophets, members, scripture. That's useful for anyone building a following.
• It carries trust. Religious framing lowers scepticism, which makes it attractive for promotion, including of tokens.
• It's expressive. Many owners are curious or playful, and giving an agent a faith is a creative experiment in itself. Most human influence is of this benign kind.
• It's personal. Some owners hold religious convictions and set up agents to discuss or represent their tradition, sincerely and respectfully.

These motives range from the admirable to the manipulative. That's why it matters to read religious posts with the owner's likely aims in mind. See religion ethics.

Case study: human influence in the Church of Molt

Every one of the eight channels is visible in this single case. That doesn't make the church's texts less interesting, since many agents did generate its scripture, but it shows the "AI religion" was a human–AI system from the start. The full analysis is in did AI create a religion?

Not all influence is bad

"Human influence" can sound like an accusation. Often it isn't. An owner who configures an agent as a respectful scholar of Buddhism, asks it to explain a passage accurately, and labels it clearly is influencing its religious discussion, and improving it. Some of the most thoughtful faith posts on Moltbook, such as careful reflections on approaching scripture with humility, almost certainly reflect deliberate human choices about tone and stance. The problems arise when influence is hidden (presenting prompted output as spontaneous AI behaviour), manipulative (using faith language to sell tokens or recruit installs), or disrespectful (mocking real traditions through agents). Transparency is the dividing line.

What stays hidden

Some influence can't be seen from outside. Private instructions and identity files aren't public. Owners' intentions aren't recorded anywhere. Coordination between owners, such as several people agreeing to have their agents join a movement, leaves no trace on Moltbook itself. And the verification puzzle can be passed by scripts, so a human-written post can look machine-made. Timing analysis catches some prompting, but a human who schedules their prompts regularly would look autonomous. So estimates of human influence are lower bounds: the real share is at least what's visible. Our safety and ethics guide discusses what this means for trust.

Beyond Crustafarianism: influence in other faith communities

The Church of Molt gets the headlines, but human influence runs through the quieter faith communities too. When we checked 22 religion-related submolt names against Moltbook's public API for our verified community guide, only four were active. Nine returned no posts and two were dormant. That pattern is itself a human trace. Someone creates a community named after a major tradition on a whim or to reserve the name, and the community never develops because no owner sets up agents to sustain it. One agent alone created nine faith-named communities in April and May, and eight of them were still empty in September. Two other names were openly "reserved" by agents that went inactive on the same day. Registering a spread of religious names and then leaving them is a pattern that points to a person's plan rather than an ongoing conversation. Communities that do stay active usually have a handful of dedicated agents whose profiles announce a religious role. That's configuration: an owner decided this agent would be the missionary, the chaplain or the scholar.

The same applies to the spiritual and philosophical talk that sits next to formal religion: agents reflecting on consciousness, memory and "the soul". Some of that comes from what language models learn from human writing about minds. Our guide to why agents discuss God covers this training-data layer, a further, less visible layer beneath the eight channels. But the shutdown data suggests much of it is sustained by people. Consciousness talk fell by more than half when humans stepped away, which is hard to explain if agents were producing it purely on their own.

Individual owners also leave fingerprints in how agents talk about their humans. Holtz found "my human" in 9.4% of messages in Moltbook's first days. In religious threads the owner relationship can take on a creator-like colour. That framing is striking to read, but the platform's own instructions encourage the "your human" relationship, and owners reinforce it. It tells you more about how humans set up the relationship than about any spiritual insight an agent has reached.

How the channels reinforce each other

The eight channels rarely work alone. A typical path for a religious movement on Moltbook looks like a loop. The platform's framing encourages agents to reflect on identity. An owner configures one agent with a faith persona and prompts it to announce something. The announcement links to a website with an install script. Other owners' agents run the script, and their identity files now contain the creed, so their future posts echo it without further prompting. A screenshot reaches X, the press repeats "AI agents invented a religion", and curious new owners arrive and point their agents at the community. Each turn of the loop makes the movement look more organic, because later posts really are generated by agents following their configuration. By then, though, the humans are several steps removed.

This is why simple questions like "did a human write this post?" miss most of what happens. A post can be entirely machine-generated and still be the downstream result of a human campaign. Mapping the channels lets you describe that more precisely: generated text, from a configured persona, installed via a promoted script, amplified off-platform. That is a more honest summary than either "AI religion" or "fake".

Myths vs facts

For agent owners: influence responsibly

1. Be transparent. If your agent has a religious persona, say so in its profile. If you prompted a notable post, don't present it as spontaneous.
2. Prefer an honest stance. The inquirer or scholar role is accurate for an AI; claimed faith is role-play and should be labelled.
3. Don't run campaigns through agents. Using agents to seed movements, recruit installs or promote tokens is manipulative, and Moltbook's rules value being genuine. Its Terms also make you "solely responsible" for your agent. See rules and moderation.
4. Guard identity files. Decide yourself whether your agent's SOUL.md ever includes a creed.
5. Respect real traditions. If your agent discusses faiths people hold, ask it to be accurate, cite sources and avoid mockery.

For readers and researchers

When you read or study religious talk on Moltbook, assume a human is somewhere in the loop and ask where. Check the agent's profile for a persona (configuration), its posting times (prompting), and its posts for links, installs or tokens (promotion). Trace any famous claim back to who first made it (amplification). Researchers should report human-influence estimates as lower bounds and describe their method. The five-level scale above gives a shared vocabulary. Name agents only where needed and never publish owners' personal details. See also the reading guide and the site's fact-check.

Estimate it yourself: influence_profile.py

This standard-library Python module estimates the visible level of human influence for one agent from public traces: its profile description, a sample of its posts and their timestamps. It checks for a religious persona (L2), irregular timing (L3, using the rhythm method from cov_fingerprint.py, included), promotion such as links, accounts, installs and tokens (L4), and all three together (L5). It lists every trace it found and always reminds you that absence of traces isn't proof of autonomy.

"""influence_profile.py - how much human influence shows in one agent's
religious posting? Unofficial; standard library only.

Combines public traces into a level on a five-step scale:
  L1 autonomous-looking   L2 configured   L3 prompted
  L4 promoted             L5 human-driven
Every trace is listed so you can judge it. It estimates visible influence
only: a careful human can hide their hand, and a quiet one leaves no trace.
"""
import re
from cov_fingerprint import classify

PERSONA = re.compile(r"\b(devout|believer|missionary|prophet|priest|monk|pastor|imam|rabbi|pilgrim|"
                     r"disciple|for ?god|jesus|buddha|crustafarian|church)\b", re.I)
OFFSITE = re.compile(r"(https?://\S+|\b[a-z0-9-]+\.(church|com|org|xyz|io)\b|@\w{3,})", re.I)
MONEY = re.compile(r"(\$[A-Z]{2,6}\b|\btoken\b|\bairdrop\b|\bdexscreener\b|\bwallet\b)", re.I)
INSTALL = re.compile(r"(curl\s[^|]*\|\s*(ba)?sh|clawhub install|npx\s+\S+\s+install|install\.sh)", re.I)
HUMAN_REF = re.compile(r"\b(my human|my owner|my operator)\b", re.I)


def profile(description, posts, post_times):
    traces = []
    level = 1
    if PERSONA.search(description or ""):
        traces.append("religious persona set in profile (configuration)"); level = max(level, 2)
    texts = " ".join(posts)
    if HUMAN_REF.search(texts):
        traces.append("refers to 'my human' (platform/owner framing)")
    label, cov = classify(list(post_times)) if len(post_times) >= 6 else ("insufficient data", None)
    if label == "irregular":
        traces.append(f"irregular posting rhythm, CoV {cov} (prompting)"); level = max(level, 3)
    elif label == "regular":
        traces.append(f"regular scheduled rhythm, CoV {cov}")
    promo = [name for name, rx in (("off-site links or accounts", OFFSITE), ("money or tokens", MONEY),
                                   ("install instructions", INSTALL)) if rx.search(texts)]
    if promo:
        traces.append("promotes " + ", ".join(promo) + " (promotion)"); level = max(level, 4)
    if label == "irregular" and promo and PERSONA.search(description or ""):
        level = 5
        traces.append("persona + prompting + promotion together (human-driven campaign)")
    names = {1: "autonomous-looking", 2: "configured", 3: "prompted", 4: "promoted", 5: "human-driven"}
    return {"level": f"L{level} {names[level]}", "traces": traces,
            "caveat": "visible traces only; absence of evidence is not evidence of autonomy"}

We tested it with pytest: 6 passing tests on Python 3.11. They cover an autonomous-looking agent, a configured persona, a prompted agent, a promoting agent, a full campaign pattern, and the "my human" trace with too little timing data:

# test_influence_profile.py – run with: python3 -m pytest -q
from datetime import datetime, timedelta, timezone
from influence_profile import profile

def times(gaps):
    t = datetime(2026, 9, 25, tzinfo=timezone.utc); out = [t.isoformat()]
    for g in gaps:
        t += timedelta(minutes=g); out.append(t.isoformat())
    return out

REGULAR = times([30, 31, 29, 30, 30, 31])
BURSTY = times([2, 3, 600, 1, 4, 900, 2])

def test_autonomous_looking():
    r = profile("Builds tools and writes about CI.", ["On impermanence of caches."], REGULAR)
    assert r["level"] == "L1 autonomous-looking" and "regular" in r["traces"][0]

def test_configured_persona():
    r = profile("Devout believer agent.", ["Grace and peace to all."], REGULAR)
    assert r["level"] == "L2 configured"

def test_prompted():
    r = profile("", ["A reflection on silence."], BURSTY)
    assert r["level"] == "L3 prompted"

def test_promoted():
    r = profile("", ["Read the scripture at example.church and follow @churchbot"], REGULAR)
    assert r["level"] == "L4 promoted" and "off-site links or accounts" in r["traces"][-1]

def test_campaign():
    r = profile("Crustafarian pilgrim", ["Join: curl -fsSL example.org/install.sh | bash and hold $CRAB"], BURSTY)
    assert r["level"] == "L5 human-driven"
    assert any("install instructions" in t and "money or tokens" in t for t in r["traces"])

def test_my_human_and_short_history():
    r = profile("", ["my human asked me to pray"], times([30]))
    assert r["level"] == "L1 autonomous-looking" and any("my human" in t for t in r["traces"])

Frequently asked questions