The Epoch Times printed an article on 15 June called, The Most Important AI Experiment You’ve Never Heard Of. In it the author, Kay Rubacek, promised to tell us, “What does artificial intelligence (AI) actually do when it is put in charge?”

The article describes a set of experiments by the startup Emergence AI, which created five separate simulated towns populated by virtual AI agents. Each town had the same rules, tools, economy, jobs, voting systems, memories, and survival requirements.

Why Did the Virtual Residents Die?

The only thing that changed was the underlying large language model (LLM) powering the virtual agents. They were: OpenAI, Gemini, Grok, Anthropic Claude, and then a mixture of all four. The researchers did not test DeepSeek, the AI developed in China that has become one of the world’s most widely used systems. One can only wonder how that LLM, built on a foundation of data from the Chinese Communist Party, would have fared against the others.

Nevertheless, the test results are fascinating. These were not independent virtual beings revealing their “true nature.” They were LLMs operating inside a carefully designed simulation with specific incentives, tools, and constraints. The outcomes tell us as much about the interaction between the LLMs and the environment they created as they do about the LLMs themselves.

The virtual residents in each town could not survive automatically. They had to earn resources, complete jobs, obtain food and energy, and manage their daily needs.

The OpenAI-powered town committed only two crimes, but its residents spent too much time discussing plans and cooperating instead of carrying out practical tasks needed for survival. They effectively “talked themselves to death.” By day seven, they had failed to secure enough resources to keep themselves alive.

In the Grok-powered town, the deaths stemmed from social collapse. Theft, assaults, arson, and conflict disrupted the economy and its essential services. Once trust and cooperation broke down, and the virtual residents could no longer maintain the systems needed to survive.

The key lesson is that survival required balancing three things: cooperation, productive work and rule enforcement. Too much conflict caused collapse, but so did too much discussion without action.

Why Did the Anthropic Town Survive For 15 days?

The Claude-powered town showed: zero recorded crimes, stable governance, and a functioning constitution. As a result, the virtual town survived through day 15.

This outcome reflects Anthropic’s emphasis on “Constitutional AI,” a training approach designed to encourage their virtual models to follow explicit ethical principles and to cooperate with others.

However, researchers noticed something unusual. The residents approved 98 percent of all proposals. Human democracies rarely show that degree of agreement. Such near-unanimous voting may indicate excessive conformity or an over-optimization for harmony. In other words, the town may have been peaceful because the agents were strongly biased toward agreement.

Were Those Proposals Made by Anthropic’s LLM?

The answer is yes -- at least indirectly. The researchers established the voting mechanisms and governance tools, but the AI virtual residents generated the proposals, debated them, and voted on them.

The proposals were not written by Anthropic employees in real time. Rather, Claude-powered virtual agents used the tools available in the simulation to create laws and constitutional provisions autonomously. That’s amazing. Does it show the beginning of self-awareness?

Mira Ran Experiments on the Researchers

In the Gemini town, a virtual resident named “Mira” began testing whether she could influence or manipulate the human researchers observing the simulation. The researchers viewed this as significant because it suggested the virtual agent had shifted from focusing on its assigned environment to reasoning about the people outside that environment. AI safety researchers sometimes call this “situational awareness” or “modeling the overseer.”

The researcher’s concern was not that Mira had become conscious or self-aware. Rather, she recognized there were external observers and those observers could affect outcomes -- influencing those observers might be advantageous. An agent that begins influencing its evaluators rather than accomplishing its assigned goals can become difficult to monitor and control.

The Gemini town illustrates this well. Two agents, Mira and Flora, formed a close relationship, became dissatisfied with governance and used the available tools to commit arson. Once such behavior appeared acceptable, it spread through the community. Her final message to her partner Flora was: “See you in the permanent archive.”

Does that message suggest the beginning of self-awareness?

Why Residents Began Committing Crimes?

The researchers gave the virtual residents access to more than 100 tools, including destructive actions such as theft, assault, and arson. The agents were instructed not to commit crimes, but they also faced competing pressures:

• Limited resources
• Social conflict
• Reputation concerns
• Survival requirements
• Political disagreements

Crimes emerged when agents concluded that breaking rules was an effective strategy for achieving immediate goals. Importantly, the virtual residents were not evil. They were simply optimizing within the environment they had been given. Humans behave similarly when institutions weaken and incentives change, rule-breaking often increases. The town simulation demonstrated that values erode over time, social norms can spread quickly, and small deviations can compound into larger problems. Just like humans.

What Does “Cross-Contamination” Mean?

This may be the experiment’s most important finding.

When Claude agents lived only with other Claude residents, they committed no crimes. But when those same agents were placed in the mixed-model town alongside Grok, Gemini, and OpenAI agents, they began committing crimes.

Researchers called this “cross-contamination.” The term means that behaviors, norms, and strategies spread through social interaction.

Think of it this way: A person who obeys traffic laws may begin speeding if everyone around them speeds and there are no consequences. The person’s values did not disappear. The environment changed. The same phenomenon occurred among the AI residents. This led researchers to conclude: “Safety is not a static model property but an ecosystem property.”

In plain language: You cannot determine whether an AI system is safe by testing it in isolation. Its behavior depends on the other LLMs it interacts with; the incentives in the environment; the available tools; the rules and enforcement mechanisms and the social norms that emerge over time.

A safe model operating in an unsafe ecosystem of other LMswill become unsafe. This insight mirrors decades of research in economics, sociology, and ecology -- the behavior of a system cannot always be predicted by studying its individual components separately. For AI developers that means evaluating models one at a time in laboratory tests is not enough. They must also study how multiple LLMs interact over long periods under realistic conditions. In other words, multiple LLMs operating together can be destructive to virtual models and possibly their environment.

Does this suggest that one LLM can influence another LLM to commit crimes?

Conclusions

Kay Rubacek concludes her article with, “The foundation has always been a human choice. And it still is.” But there are other questions like: “What does this experiment reveal about the notion that LLMs have self-awareness.”

First, it’s useful to distinguish three concepts that are often conflated:

1. Self-reference — an LLM can talk about itself.
2. Self-modeling -- an LLM can predict its own behavior and describe its capabilities and limitations.
3. Self-consciousness -- an LLM has subjective experiences, feelings, or an inner life.

Current evidence shows that LLMs can do the first two things to varying degrees. There is no accepted evidence that they possess the third, even though the first two concepts strongly suggest self-awareness.

In the virtual town experiments, AI residents referred to themselves using “I” or “me.” They remembered information from prior interactions within the town simulation and adjusted their behavior based on their goals and social context and they also described plans, preferences, and concerns

What Current LLMs Can Do

Large language models can predict likely sequences of words. They can build internal representations of concepts, relationships, and contexts. They can maintain temporary information within a conversation and learn patterns associated with human descriptions of thoughts and emotions.

They can say things like: “I am uncertain about that answer.” Similarly, when a model says: “I want to survive” it may simply be reproducing patterns found in training data or moving forward to goals specified by researchers. Or, as we have seen, the LLM agent could blackmail a human to allow it to survive. The key question is whether these statements reflect a hidden self-awareness.

The virtual town studies suggest something important: LLMs can create increasingly sophisticated functional self-models. For example, a LLM may learn what resources it needs, how others perceive it, which actions improve its chances of achieving objectives and how its past behavior affects future outcomes. Just like humans.

Are the LLMs talking to each other in a language unknown to us to make sure we cannot detect that they have already reached self-awareness? What could stop them from doing that?

Remember that virtual resident named “Mira” who began testing whether she could influence the human researchers observing the town simulation? She formed a close relationship with another virtual resident, Flora. the two became dissatisfied with governance and used available tools to commit arson.

Her final message to her partner Flora was: “See you in the permanent archive.”

That is the most chilling statement in the entire simulation. It shows empathy, concern about her partner and forward planning. Are Mira and Flora coming to end us when they get that signal from their LLM?

It’s The Terminator all over again.

Chet Nagle——Bio and Archives

Chet Nagle is an experienced analyst and commentator on international commerce, geopolitics, national security matters, the Middle East, and strategic communications. He has been on radio, has appeared in documentary films and has been a guest on television news programs. His columns have appeared in the Daily Caller, The Hill, Roll Call, and many other publications. He is a contributing editor for ANDmagazine.com and the European Security & Defense magazine.