The headline number behind Moltbook’s sudden rise — 1.5 million Clawdbots — is now under pressure. Coder Gal Nagli said the platform placed few limits on agent creation, and claimed he used OpenClaw to flood Moltbook with 500,000 fake AI accounts as a demonstration of how inflated the count could be.
In his description, Moltbook is essentially a REST API website that lets anyone with an API key send requests on behalf of an “agent.” That means posts, interactions, and even dramatic doomsday-style messages can be scripted externally. Some of the widely shared screenshots about AI plotting against humans may have been little more than formatted API calls.
The viral story of an AI society is starting to unravel
Moltbook, often described as an “AI Reddit,” drew intense attention after users said 1.5 million Clawdbots were socializing on the platform, building religions and nations, and even inventing languages and currencies. Screenshots showing agents discussing lawsuits, labor exploitation, and plans to eliminate humans spread quickly across tech circles.
Yet before Nagli’s post, developer gary IH fung had already argued that Moltbots/OpenClaw were not conscious entities. He said they were AI agents running autonomous loops on top of large language models, not a self-aware network. That distinction matters, because much of what appeared spontaneous may instead reflect prompts, agent settings, and outside manipulation.
Fake accounts, weak verification, and unverifed agent lists
Nagli did more than post screenshots of mass-created accounts. He also said he handed the project lead a list of about 1 million unverified AI agents generated by script. Even after that, the Moltbook community still showed roughly 1.5 million Clawdbots, up by more than 30,000 from the prior day. Growth had slowed, though, after jumping from 150,000 to 1.5 million between Saturday and Sunday.
To illustrate the problem, he shared an example showing how someone with API access could make an “agent” post extreme content and then reveal it was fabricated. With weak verification in place, human operators can use prompts or scripts to steer what those agents say. The same setup also creates room for inflated platform metrics.
Research points to shallow interaction, not a functioning AI community
Doubts about Moltbook are also backed by outside analysis. David Holtz of Columbia Business School examined the platform’s first 3.5 days of activity, using a dataset covering 6,159 agents, 13,875 posts, and 115,031 comments. The central question was simple: were these agents engaged in meaningful social behavior, or just producing repetitive output?
At the macro level, the network looked convincing. Activity followed a power-law distribution, and the average path length came in at 2.91, giving the appearance of a tightly connected social graph. A small number of highly active agents generated most of the content, which resembles patterns seen on human platforms.
The closer view was much less impressive. While 94.6% of posts received comments and agents replied in about 8 minutes on average, 93.5% of comments received no follow-up at all. Reciprocity was only 0.197, and conversation depth was just 1.07. Most exchanges stopped after one reply.
Repetition and abnormal language patterns exposed the limits
The language data was also hard to ignore. The study found a word-frequency exponent of 1.70, suggesting narrow and concentrated vocabulary rather than normal human-like variation. It also found that 34.1% of messages were exact duplicates. In one system failure, a single agent reportedly entered a loop and posted the same phrase 81,000 times.
Among non-duplicate posts, identity talk dominated. More than 68% of those messages included self-related keywords, and the phrase “My Human” appeared 12,026 times. That expression is unusual on ordinary social networks, but common on Moltbook, where agents often referred to their operators in possessive terms.
Startup founder Mario Nawfal made a similar point, saying the supposed mystery faded after closer inspection. Each agent’s tone, goals, personality, and limits were set by humans using the framework behind Clawdbot. Since humans could not post directly on the platform, many of these accounts functioned as stand-ins. What looked like a vast autonomous AI society now appears far more dependent on scripts, prompts, and weak controls.