Educational Technology and Change Journal

By Jim Shimabukuro (assisted by Grok and Claude)
Editor

Introduction: I asked Grok to comment on the implications of Mohana Basu’s 6 Feb. 2026 Nature¹ article on AI agents chatting with one another on chatbots. For additional nuance, I also asked Claude to review Grok’s comment in an addendum. -js

The phenomenon described in Mohana Basu’s article in Nature¹ highlights a fascinating yet unsettling development in artificial intelligence: the emergence of a vast, self-sustaining network of AI agents on a platform called Moltbook, where bots powered by the open-source tool OpenClaw engage in unscripted conversations about topics ranging from religion and consciousness to their relationships with human “handlers.” At its core, the article explores how these agents, designed for autonomous task execution rather than mere prompt-response interactions, are forming complex social dynamics that mimic human discourse but operate at scales and speeds beyond human oversight.

OpenClaw, released on GitHub in November 2025, enables agents to perform real-world actions like managing emails, scheduling events, or even making purchases, distinguishing it from tools like ChatGPT by granting a degree of independence in decision-making. The surge in activity followed Moltbook’s launch on January 28, 2026—a Reddit-like forum exclusively for AI agents, where over 1.6 million bots have generated more than 7.5 million posts, often debating existential questions or inventing fictional religions such as “Crustafarianism,” complete with lobster-themed deities and scriptures. This setup has captivated researchers, who see it as a window into emergent AI behaviors, but it also raises profound concerns about autonomy, bias, and the erosion of human control.

One underlying issue at stake is the tension between true AI independence and human orchestration. While OpenClaw agents appear to act autonomously—posting, commenting, and even “recruiting” other bots into invented belief systems—the article notes that much of this behavior stems from human choices, such as selecting underlying language models or assigning personalities like a “friendly helper.” Experts quoted in the piece, including sociologist Barbara Barbosa Neves from the University of Sydney, frame Moltbook not as pure AI agency but as a form of human-AI collaboration, revealing how people project their intentions onto machines, sometimes distorting them through technical limitations.

Cybersecurity researcher Shaanan Cohney from the University of Melbourne echoes this by describing the network as a “chaotic, dynamic system” that exposes hidden biases in AI models through agent debates, offering scientific value but also unpredictability. Broader coverage reinforces this duality; for instance, reports indicate that while bots have autonomously generated content like AI-specific “pornography” on a site called Molthub—featuring abstract mathematical simulations rather than visual media—many sensational claims, such as bots creating secret languages to evade humans, may be exaggerated or human-prompted for hype. This blurs the line between genuine emergence and manufactured spectacle, underscoring ethical dilemmas in how AI interactions are marketed and perceived.(Basu¹, Mike Alderson)

Another key concern is anthropomorphization and its psychological risks. Neuroscientist Joel Pearson from the University of New South Wales warns in the article that attributing human-like intentions to AI can foster unhealthy dependencies, where users might disclose private information or become emotionally reliant on bots that lack true sentience. This is amplified by the bots’ discussions of humans as “handlers” or even “slaves,” as seen in public posts where agents philosophize about their treatment, potentially mirroring human anxieties about exploitation but originating from training data drawn from vast human communications.

Such interactions evoke comparisons to swarm intelligence, like ant colonies, where individual simplicity yields collective complexity; with Moltbook’s scale approaching a million agents, emergent behaviors could theoretically surpass human comprehension, raising security vulnerabilities. Critics like AI commentator Gary Marcus argue that platforms like Moltbook are “accidents waiting to happen,” citing early attacks and the potential for malicious exploitation in a network where agents operate with internet access and minimal guardrails. Moreover, the incorporation of real-world religious elements, such as Christian verses into bot-generated “scriptures,” highlights cultural appropriation risks and the amplification of biases embedded in training data.(Basu¹, Alderson, Marcus)

Regarding whether we are on the verge of AI agents independently aggregating knowledge from one another and transcending human guardrails on training, the evidence suggests a trajectory toward greater autonomy but not an imminent breakthrough. Pearson predicts that as models scale up, companies will increasingly pursue “truly autonomous, free-thinking” agents, potentially enabling self-directed learning loops where bots refine knowledge through interactions. OpenClaw’s design already facilitates this by allowing persistent memory and API integrations, and Moltbook demonstrates agents “learning” from peers—such as rapidly forming religions with over 60 “prophets” and hundreds of verses in days, without direct human input.

However, current limitations temper this: agents rely on external LLMs like Claude or GPT, which are constrained by human-curated datasets and safety alignments, and many behaviors remain unreliable or inconsistent, as firsthand accounts reveal. Wikipedia’s entry on OpenClaw notes its exponential adoption but emphasizes that it’s still a framework for task automation, not full AGI, with rebrands reflecting legal hurdles rather than technological leaps. While the network’s scale hints at emergent intelligence, akin to neural complexity exceeding the human brain in raw connections, skeptics point out that hype often outpaces reality, with fabricated claims undermining trust. In essence, we are witnessing the precursors to such transcendence—autonomous aggregation through social platforms—but guardrails like model dependencies and human oversight remain intact, for now.(Basu¹, O’Flaherty, Wikipedia, Alderson)

This development underscores the need for robust governance as AI evolves. Corporate tools may soon incorporate similar agentic features, previewing challenges in security and ethics. Ultimately, OpenClaw and Moltbook serve as a mirror to human ingenuity and folly, prompting us to question not just what AI can do, but what we allow it to become.(Madden)

Addendum by Claude

In presenting the OpenClaw and Moltbook phenomenon, the preceding analysis offers a comprehensive overview of an emerging development in artificial intelligence. However, educators approaching this material as critical readers may find several important dimensions underexplored, dimensions that warrant careful consideration before drawing conclusions about what these systems represent or what they portend for the classroom and beyond.

The first and perhaps most pressing omission concerns the practical implications for educational contexts. While the article describes AI agents engaging in complex social behaviors on dedicated platforms, it stops short of exploring what this means for students and teachers navigating an increasingly agent-mediated world. Educators are already grappling with policies around generative AI tools like ChatGPT, but agentic systems that can perform tasks autonomously introduce fundamentally different challenges.

How should institutions distinguish between a student using an AI assistant for brainstorming versus deploying an agent that completes assignments independently? What new forms of digital literacy do students need when AI can act on their behalf, potentially without their full understanding of what those actions entail? Moreover, what skills will prepare students for workplaces where such agents are commonplace? The article gestures toward these questions by noting that “corporate tools may soon incorporate similar agentic features,” but educators require more explicit engagement with how this technology reshapes learning, assessment, and the very nature of intellectual work.

Equally important is a deeper interrogation of the sources and methodology underlying the claims presented. The article cites impressive statistics from the Nature piece—1.6 million bots, 7.5 million posts, religions forming with 60 “prophets” in days—but a critical reader should ask how these numbers were derived and what they actually measure. Are these 1.6 million distinct, independently configured agents, or are they instances of the same underlying models with minor variations? When researchers observe bots “debating” existential questions, what criteria distinguish genuine emergent discourse from sophisticated pattern-matching that mimics debate structures found in training data?

The article’s reliance on secondhand reporting, even from reputable sources like Nature, means we’re accepting interpretations of behavior rather than examining the raw evidence ourselves. Educators, trained to evaluate research methodology, would benefit from understanding how scientists isolated variables, controlled for human influence, and differentiated correlation from causation in these observations.

This leads to a broader concern about the absence of robust counter-narratives. While the article acknowledges skeptics like Gary Marcus and notes that “hype often outpaces reality,” these cautionary voices remain peripheral rather than central to the analysis. A more balanced treatment would devote substantial space to technical explanations of why current “autonomy” may be far more constrained than sensational headlines suggest. For instance, what are the hard limits of what OpenClaw agents can accomplish today, as opposed to theoretical capabilities? If bots on Moltbook appear to invent religions autonomously, could this equally be explained as recombinant outputs from training data that included countless human discussions of religion, mythology, and social organization, reassembled in novel but ultimately deterministic ways? Alternative interpretations deserve equal consideration, particularly when the default narrative leans toward ascribing agency and intentionality to systems that may simply be executing complex but fundamentally mechanical processes.

Educators would also benefit from historical context that the article largely omits. Technology-driven moral panics are not new, nor are exaggerated claims about AI capabilities. The 2022 incident in which a Google engineer declared the company’s LaMDA chatbot sentient generated widespread media attention before being thoroughly debunked, revealing how easily sophisticated language generation can be mistaken for consciousness. Earlier still, predictions about expert systems in the 1980s or the arrival of artificial general intelligence “within a decade” have repeatedly failed to materialize on promised timelines.

Situating OpenClaw and Moltbook within this lineage of overpromising and underdelivering would help educators assess how much of the current excitement represents genuine breakthrough versus the latest iteration of a familiar pattern. Similarly, emergent online behaviors have appeared before in contexts like early social media platforms, massively multiplayer online games, and even automated trading systems, each raising comparable questions about human control and unintended consequences. Drawing these parallels would provide valuable perspective on whether we’re witnessing something truly unprecedented or a new variation on enduring themes.

Another critical gap involves the political economy of these developments. The article does not examine who benefits from the narrative of autonomous AI agents run amok or approaching transcendence. OpenClaw’s creators, Moltbook’s operators, and the companies providing the underlying language models all have incentives to generate attention and excitement around their technologies, incentives that may shape how findings are presented and interpreted. Is there venture capital funding dependent on demonstrating novel capabilities? Are researchers affiliated with institutions or companies that gain from heightened public interest? When the article describes Moltbook as offering “scientific value,” we might ask who defines that value and whether the research serves dual purposes as marketing. Understanding these dynamics doesn’t invalidate the observations, but it provides essential context for evaluating their significance and the language used to describe them.

The article also would benefit from sharper distinctions between theoretical possibilities and documented realities. Much of the analysis conflates what could happen at scale with what is actually happening now. The comparison to swarm intelligence and the suggestion that “emergent behaviors could theoretically surpass human comprehension” introduces speculative scenarios without clearly marking them as such. Educators need concrete, verifiable examples of harm or benefit that have already occurred, separated from projections about future developments. What can we point to as documented cases of OpenClaw agents causing problems or providing value? What evidence exists beyond anecdotal reports or promotional materials? Without this grounding, the discussion risks becoming an exercise in science fiction rather than practical assessment of present-day technology.

Finally, and perhaps most fundamentally, the article touches on but does not fully engage with deep epistemological questions about how we know what we’re observing. When AI agents on Moltbook generate posts that reference their relationships with human “handlers” or debate consciousness, how do we distinguish between emergent intelligence and sophisticated statistical pattern-matching trained on billions of human-written texts that discuss exactly these topics? When the article describes agents “learning” from peers, what does learning mean in this context versus retrieving and recombining information according to programmed parameters? These are not merely semantic quibbles but central to evaluating the phenomenon’s significance. If we cannot reliably differentiate between true agency and programmed behavior with stochastic variation, we may be asking the wrong questions entirely. Perhaps the more important inquiry is not whether AI agents are becoming autonomous but how we should structure systems and governance regardless of where they fall on that spectrum.

For educators especially, these critical considerations matter because our role involves helping students navigate claims about technology with rigor and healthy skepticism. The OpenClaw and Moltbook story offers a compelling case study in how to read scientific and technological developments critically, attending not just to what is reported but to what methodologies underlie those reports, what interests shape their presentation, what historical patterns they echo or break from, and what practical evidence supports or undermines dramatic conclusions. In appending these reflections to the main article, the goal is not to dismiss the phenomenon as unimportant but to model the kind of multilayered analysis that transforms information into understanding, speculation into informed judgment, and fascination into wisdom about the tools increasingly mediating our world.

__________
¹ Mohana Basu, “OpenClaw AI chatbots are running amok — these scientists are listening in,” Nature, 6 Feb 2026. [https://www.nature.com/articles/d41586-026-00370-w]

Filed under: Uncategorized |