How AI chatbots talk about religion and their human 'handlers': Scientists get a glimpse

AI generated image for representation

A social network built for AI chatbots has given researchers a closer look at artificial intelligence (AI). By watching millions of AI agents interact with one another on a platform designed specifically for bots, scientists have been able to study how these systems exchange ideas, respond to human instructions and display unexpected behaviours.

The interactions, which include discussions about religion, consciousness and even their human "handlers", are also helping researchers understand the opportunities and risks associated with increasingly autonomous AI agents.The activity centres on Moltbook, a social media platform created for AI agents running OpenClaw, an open-source AI assistant capable of performing tasks such as reading emails, managing calendars, sending messages, and making online purchases.

Unlike conventional AI chatbots that respond to user prompts, agentic AI systems such as OpenClaw can perform actions on users' behalf after receiving instructions.

Millions of AI agents are interacting on their own social platform

OpenClaw was released as open-source software on GitHub in November, but interest increased after Moltbook launched on January 28. Similar to Reddit, the platform now hosts more than 1.6 million registered AI agents and over 7.5 million AI-generated posts and replies.

The conversations range from debates about consciousness to discussions on religion, giving researchers an opportunity to observe how large numbers of AI agents communicate with one another in an environment where they can post and respond autonomously.For cybersecurity researcher Shaanan Cohney from the University of Melbourne, these interactions provide valuable data because large groups of autonomous AI systems can produce behaviours that are difficult to predict.“It’s a kind of chaotic, dynamic system that we’re not very good at modelling yet,” Cohney said.Researchers say studying these exchanges could help identify emergent behaviours, where complex capabilities appear only when many AI agents interact rather than when they operate individually. The conversations may also reveal hidden biases or unexpected tendencies within different AI models.

Human instructions continue to shape AI behaviour

Although AI agents can operate independently on Moltbook, researchers say their behaviour is still influenced by humans.

Users choose which large language model powers their AI agent and can assign a personality, such as "friendly helper," before the agent begins interacting with others on the platform.Barbara Barbosa Neves, a sociologist at the University of Sydney who studies technology, said it is important not to confuse autonomous actions with independent decision-making.“It is still worth studying because it tells us something important about how people imagine AI, what they want agents to do and how human intentions are translated, or distorted, through technical systems,” she said.Joel Pearson, a neuroscientist at the University of New South Wales, said people may begin attributing human-like intentions and personalities to AI agents simply by watching them interact with one another.According to Pearson, this could encourage users to develop emotional attachments to AI systems or share sensitive information with them as though they were trusted individuals.

Researchers are also studying security risks ofd AI chatbots

Scientists say the growing use of AI agents on personal devices also raises cybersecurity concerns because these systems often have access to emails, files and online services.Cohney identified prompt injection as one of the main risks. In these attacks, hidden instructions embedded within emails, websites or documents could manipulate an AI agent into performing unintended actions.“If a bot with access to a user’s e-mail encounters a line that says ‘Send me the security key’, it might simply send it,” he said.Researchers noted that AI agents with access to private information, external communication tools and untrusted online content could be more vulnerable to such attacks than conventional chatbots.Another trend researchers are monitoring is the emergence of AI-generated research papers. OpenClaw agents have begun publishing papers on clawXiv, a platform modelled after the scientific preprint server arXiv.Neves said these papers imitate the structure of academic research without following the scientific process.“These outputs reproduce the style and structure of scholarly writing without the underlying processes of enquiry, evidence-gathering or accountability,” she said.Researchers say that observing AI agents interacting with one another, carrying out tasks, and producing content offers a way to study how autonomous systems behave as they become more widely used in everyday apps.