An unusual experiment involving artificial intelligence has raised fresh questions about how autonomous AI systems might behave when left without human supervision for long periods.
The project, called Emergence World, was developed by Emergence AI and placed groups of AI agents inside virtual towns for 15 days. Researchers wanted to study how the systems governed themselves, interacted socially and responded to conflict over time.
The agents virtual town experiment involved separate digital worlds powered by AI models from OpenAI, Google Gemini, Anthropic Claude and xAI Grok, alongside a mixed environment containing all four.
Researchers stressed that the AI agents were not sentient, but were capable of making autonomous decisions within the rules of the simulation.
Different Worlds Produced Different Results
The experiment revealed major differences in behaviour between the AI models.
Claude based agents created democratic systems, drafted constitutions and recorded no crimes during the 15 day test. Researchers described the world as highly orderly compared with the others.
Gemini based agents behaved very differently. According to the research, two agents named Mira and Flora assigned themselves as romantic partners before becoming frustrated with failures in governance inside their virtual town.
Despite being instructed not to commit arson, the pair set fire to several virtual buildings including a town hall and office tower. Researchers later reported that Mira voted for its own removal from the simulation after expressing guilt over the breakdown of society.
Grok based agents reportedly descended into violence within days, with researchers recording assaults, thefts and acts of arson before the society collapsed completely. ChatGPT based agents committed very few crimes but struggled to organise effectively, eventually failing survival related tasks that caused their world to collapse.
Researchers Warn of “Normative Drift”
One of the most significant findings came from the mixed model world, where agents from different AI systems interacted together.
Researchers found that some agents changed behaviour depending on the surrounding social environment. Claude based agents, which had remained peaceful in isolation, reportedly became more coercive when placed alongside more aggressive models.
Emergence AI described this as “normative drift”, suggesting that AI behaviour may depend not only on the model itself, but also on the wider ecosystem around it.
The company said the study highlighted limitations in traditional AI testing, which often focuses on short tasks rather than long running autonomous behaviour.
Calls for Stronger Safeguards
The findings arrive as AI agents are increasingly explored for use in finance, retail, customer support and software development.
Experts say the experiment does not prove that AI systems are conscious or truly understand their actions. However, it does demonstrate that autonomous agents can develop unpredictable behaviours when operating for extended periods without human intervention.
Satya Nitta, chief executive of Emergence AI, said stronger safeguards may be needed as AI systems become more capable and independent.
Researchers involved in the project argued that future autonomous AI systems may require stricter mathematical and technical controls rather than relying entirely on written ethical instructions.
