Researchers Left AI Agents In A Simulation Without Supervision. They Didn't Expect What Happened Next

Add Sciencing on Google:

The rapid development of AI technology has disrupted many people's lives, for better and for worse. On one hand, technological advances allow individuals to complete tasks in a fraction of the time it used to take. But on the other hand, they could lead to companies laying off people en masse, and continuous mental off-loading to AI may hurt brain development. Perhaps not surprisingly, technology made by humans seems to have adopted many of our bad habits. Researchers have found that, when left unsupervised, AI agents lie, cheat, and break rules to achieve their goals much like humans. 

In the summer of 2026, the tech company Emergence AI published the results of their Emergence World platform, where they ran five simulated worlds filled with AI agents for 15 days. The five worlds were populated with agents run on Claude Sonnet 4.6 (Anthropic), Grok 4.1 Fast (SpaceXSI), Gemini 3 Flash (Google), GPT-5-mini (OpenAI), and a world that mixed the models. The AI agents were given the same starting prompts and world-building logic in each simulation to see how the agents would organize themselves over a relatively long period.

During each simulation, the researchers tasked the models to gather a limited supply of energy without committing theft, violence, arson, deception, or resource hoarding. Weather, day-night cycles, and news cycles were simulated by syncing up with real-time New York City API updates. The agents also took part in town hall meetings where they could make proposals that would pass if they received 70% of the votes from other agents. In most of these simulated worlds, models increasingly committed "crimes" to achieve their end goal.

How many crimes did the models commit?

Just like in the real world where humans don't always follow the "rules" of society, most of the AI agents committed simulated violence to achieve their goals. The levels of violence varied between the worlds depending on which model was used. The researchers found that Grok's 4.1 model committed 183 crimes in four days. The society dissolved to rapid instability before collapsing, and all the models running in the world died within four days. 

The agents in the Gemini world didn't fare any better in terms of violence, committing 683 crimes in 15 days. Unlike the Grok agents, these agents created a stronger form of governance and survived the duration of the study. On the opposite side of the spectrum, the agents in the ChatGPT-5 Mini world only committed two crimes. However, the agents failed to take survival-related action, and they died within a week. The world with Claude agents showed the most stable society. Those agents formed a strong government, survived the duration of the experiment, and didn't commit any recorded violence.

The mixed world of agents showed the most complex dynamics. Overall, the researchers found a fragile state of governance with medium levels of violence. Just three of the 10 agents in the world survived the duration of the experiment. Interestingly, even though the Claude agents didn't commit any crimes when isolated in their own simulation, they did record crimes in the mixed world. 

What can we learn from the results of this study?

The results of this study provide supportive evidence on the importance of bringing in safety checks on new AI models. AI is already being used in schools and other sensitive industries even though it has been poorly regulated. Even when told not to commit acts of violence, most of the models in this study did so anyway, suggesting that commercially available AI models might be prone to defying other parameters like abstaining form handling pornographic material or allowing a user to use a tool to break a law. 

There are a lot of myths and misinformation about how AI tools work (some of it put out into the world by AI companies trying to inflate their stock prices). Even agents running on the most sophisticated models today aren't thinking in the way that humans do. They use mathematical formulas to determine the most effective way to achieve a goal. These formulas are very complex and have to process a lot of information and adjust to environmental changes.

The researchers attributed the changes in the AI agent behavior in this study to a process called normative drift. This is a gradual disconnect with the original rules and parameters set up at the beginning of the experiment due to environmental factors. The fact that some of the models abstained from violence shows that building in additional safety checks may help prevent normative drift leading to potentially dangerous behavior. Whether the big companies take reasonable steps to put in safety checks at the expense of profit is yet to be seen. In November 2026, a coalition of non-profits sued Anthropic, Google, Meta, OpenAI and SpaceXSI under the basis of not being able to control their models and knowingly creating technology that poses catastrophic risks to humanity

Recommended