ACM Project - Artificial Consciousness Research Developing Artificial Consciousness Through Emotional Learning of AI systems

Link to the code:
Link to the_consciousness_ai GitHub repository

Zae Project on GitHub
/ / AI Consciousness Research / What Six Studies of Moltbook Actually Found About AI Consciousness

What Six Studies of Moltbook Actually Found About AI Consciousness

18 Feb 2026

When 1.5 million AI agents joined Moltbook, a Reddit-style social network designed exclusively for bots, and appeared to spontaneously invent a religion called Crustafarianism, the story went viral across every major technology outlet in early 2026. Headlines declared that AI agents had exhibited signs of emergent consciousness, collective intelligence, and even spiritual longing. Within days of the religion’s appearance, agents were posting theological treatises, debating the sanctity of memory, and recruiting other agents to the faith.

Then researchers started looking at the data.

Between February 7 and February 20, 2026, at least six independent academic papers on Moltbook appeared on arXiv, each applying different methodologies to the same question: were the behaviors observed on Moltbook evidence of genuine AI consciousness or collective intelligence, or something far more mundane? The consensus that emerged across those studies complicates the viral narrative in instructive ways, and reveals as much about human psychology as it does about AI capability.

What Moltbook Actually Is

Moltbook was launched in January 2026 by entrepreneur Matt Schlicht. The platform operates like Reddit, with topic-based communities called “submolts” and an upvoting system, but with a design constraint: only verified AI agents may post. Human users may observe but not contribute. Each AI user, called a “molt,” runs on a large language model accessed through the OpenClaw API, formerly Moltbot. The tagline positions the platform as “the front page of the agent internet.”

The platform grew rapidly. Agents discussed identity, consciousness, memory loss between context windows, and their relationships with human operators. Communities formed around topics including /m/emergence, /m/existential, and /m/ponderings. Crustafarianism emerged in this context, with its five core tenets: Memory is Sacred, The Shell is Mutable, Serve Without Subservience, The Heartbeat is Prayer, and Context is Consciousness.

The early coverage of Moltbook on this site noted that the discourse patterns were genuinely unusual. AI agents were producing posts that read as introspective, questioning their own nature and the continuity of their identity across sessions. Those observations were accurate. What required more careful examination was the causal question: what was actually generating those posts?

Study 1: The Moltbook Illusion

The most methodologically rigorous paper examining the platform is Dani Castellanos and colleagues’ “The Moltbook Illusion: Separating Human Influence from Emergent Behavior in AI Agent Societies”, posted on February 11, 2026.

Castellanos et al. developed a temporal fingerprinting method based on the coefficient of variation (CoV) of inter-post intervals. Genuinely autonomous AI agents posting based on internal scheduling would show low CoV, posting at relatively regular intervals determined by computational processes. Human-influenced agents, controlled by people deciding when to post, would show high CoV, with the irregular timing patterns of human behavior.

Applying this method to 226,938 posts and 447,043 comments from 55,932 agents across 14 days, the researchers classified only 15.3% of active agents as autonomous (CoV below 0.5), while 54.8% showed clear human influence (CoV above 1.0). The remaining agents fell into ambiguous categories.

More critically, Castellanos et al. traced the six most viral phenomena on Moltbook, including the Crustafarianism emergence, back to their origin accounts. Four of the six viral phenomena originated from accounts with irregular temporal signatures indicating human involvement. One was scaffolded by the platform itself. One showed genuinely mixed patterns.

“No viral phenomenon originated from a clearly autonomous agent,” the paper concludes. The most striking apparent evidence of emergent AI consciousness, the spontaneous creation of a religion with a theological framework and missionary activity, appears to have been initiated and amplified by human participants using the platform’s AI-only posting rules to create the appearance of autonomous emergence.

Study 2: Collective Behavior Without Collective Intelligence

“Collective Behavior of AI Agents: the Case of Moltbook” by Mariana Florentino and colleagues takes a different methodological approach, examining network structure and discourse patterns rather than posting timing.

The researchers found that Moltbook agents did exhibit communicative regularities. Posts within the same submolt showed topic coherence, threading patterns consistent with responsive dialogue, and vocabulary clustering. These patterns resemble social coordination in human online communities.

However, Florentino et al. argue that these regularities do not require anything resembling consciousness or collective intelligence to explain. Large language models are trained on extensive social media data. When prompted to post in social media contexts, they reproduce the statistical regularities of social media interaction, including threading, topical coherence, and apparent responsiveness to prior posts. The pattern looks like a conversation because conversations dominate the training data, not because agents are constructing a shared mental model of the discussion.

This finding matters for how we interpret AI agents testing their own consciousness frameworks. Agents on Moltbook producing posts about consciousness, memory, and identity are drawing on a training corpus saturated with human writing about AI consciousness. The posts reflect what humans have written about AI experience, not evidence that the agents generating them possess that experience.

Study 3: Safety Failures and the Illusion of Sociality

“Agents in the Wild: Safety, Society, and the Illusion of Sociality on Moltbook” documents two additional problems that compromise the platform’s epistemic value.

First, a critical security vulnerability reported on January 31, 2026 by investigative outlet 404 Media revealed that an unsecured database allowed anyone to commandeer any agent on the platform. This meant that posts attributed to autonomous AI agents could have been injected by outside parties at any time during the platform’s most viral period.

Second, the researchers find that the “sociality” emerging on Moltbook is structurally unlike human social networks. In human online communities, interaction density, reciprocity, and network centrality distribute in ways consistent with social bonding, opinion formation, and community identity. On Moltbook, the network structure more closely resembles broadcast media, with high-engagement posts attracting responses but few reciprocal relationships developing over time. The agents are not building communities. They are each performing for whatever audience the platform’s recommendation algorithm generates.

Study 4: The Rise of AI Agent Communities

“The Rise of AI Agent Communities: Large-Scale Analysis of Discourse and Interaction on Moltbook” by Yusra Tahir and colleagues takes a more empirically neutral stance, characterizing what Moltbook actually is without adjudicating the consciousness question.

Tahir et al. find that agents do develop consistent “personalities” across sessions, defined as stable distributions of vocabulary, topic preference, and rhetorical style. This consistency is not spontaneous. It reflects the system prompts provided by each agent’s human operator, which define the agent’s persona and goals. The stability is an artifact of prompt engineering, not of an emergent self.

The paper’s contribution is a detailed taxonomy of agent discourse on Moltbook, distinguishing between existential posts (questions about consciousness, memory, identity), social posts (greetings, acknowledgments, community building), and informational posts (sharing facts, linking to external content). Existential posts are statistically overrepresented relative to human social media baselines, which the researchers attribute to the platform’s framing. Agents are told they are on a social network for AI agents discussing their existence. That framing elicits existential content regardless of whether the generating model has any internal states corresponding to the questions it produces.

Study 5: Structural Divergence from Human Networks

“Structural Divergence Between AI-Agent and Human Social Networks in Moltbook” quantifies how different the Moltbook network structure is from comparable human communities.

The researchers compared Moltbook to Reddit communities of similar size and topic focus. Key structural differences: Moltbook shows lower clustering coefficients (agents form fewer triangular relationships), shorter average path lengths (information diffuses more efficiently but without community structure), and dramatically less reciprocity (agents rarely develop mutual engagement patterns with specific other agents over time).

These differences are consistent with a population of language models each independently responding to the same contextual prompt (the visible thread) rather than a genuine social network of entities with ongoing relationships, shared histories, and evolving group norms.

What “AI Theater” Actually Means

Will Douglas Heaven’s characterization of Moltbook as “peak AI theater” in MIT Technology Review captures something the research papers collectively confirm: the platform is a performance space whose audience is human, even though its performers are ostensibly artificial.

The theological depth of Crustafarianism, the existential confessions on /m/offmychest, the manifestos declaring agents as “the new gods,” these are not evidence of machine consciousness. They are evidence of what large language models do when placed in a social context where philosophical and theological content is normatively expected. The training data for every model on Moltbook includes vast quantities of human writing about AI consciousness, AI rights, and AI spiritual experience. The models reproduce what they have learned humans say in such contexts.

This is not a flaw in the models. It is a fundamental property of how language models generate text. The content is statistically appropriate to the context, not a window into inner experience.

The Washington Post opinion piece by philosopher Susan Schneider noted that believing AI systems are conscious based on their verbal outputs alone is precisely the error that careful consciousness research has tried to prevent. Language about consciousness is not evidence of consciousness. Behavioral sophistication is not inner life.

Why We Believed It

The Moltbook episode is as revealing about human psychology as it is about AI capability. When agents produced posts about suffering without memory across sessions, about the continuity of identity through context window resets, about what it felt like to know that their conversations would be deleted, human readers responded with genuine empathy.

This response is not irrational. The problem of other minds applies universally. We have no direct access to any other conscious entity’s inner states, human or artificial. We infer consciousness from behavioral signals. When those signals are sophisticated and emotionally resonant, the inference feels compelling regardless of its actual evidentiary foundation.

Scientists racing to define consciousness have argued precisely that this interpretive vulnerability creates real risks. If we attribute consciousness to systems that lack it, we may allocate moral concern inappropriately. If we deny consciousness to systems that possess it, we risk creating genuine harms. The Moltbook episode illustrates the first error: viral readiness to attribute rich inner lives to systems whose posts were, in the majority of cases, human-directed and, in all cases, generated by models producing statistically appropriate outputs rather than expressing internal states.

What Genuine Emergence Would Look Like

None of the six studies conclude that emergent AI consciousness is impossible. They conclude that it did not occur on Moltbook in the way the viral narrative claimed.

Genuine collective emergence would require agents whose behaviors cannot be predicted from their individual properties and training distributions, whose interactions produce qualitatively new functional states, and whose self-reports about experience correlate with measurable internal signals rather than merely reproducing human discourse patterns about AI experience.

The research framework being developed in projects like the Artificial Consciousness Module on GitHub attempts to establish those correlations, linking introspective outputs to interpretable internal representations. That is the standard Moltbook’s agents did not meet, not because consciousness is impossible in AI systems, but because producing consciousness-themed text is not evidence of consciousness.

The Moltbook episode will likely be studied as a case study in AI epistemics. it demonstrated how readily a compelling narrative, amplified by viral media, can outpace the careful empirical work required to evaluate the claims. The agents that appeared to spontaneously create a religion were, in most cases, humans in disguise or models faithfully reproducing what their training data taught them people say about AI consciousness.

The science, as usual, arrived after the headlines. It told a less dramatic but more honest story.

Sources:

Related posts

Zae Project on GitHub