There is no utopia in the digital world.

Article author and source: GeekPark

Over the past six months, the most popular management fantasy in Silicon Valley has been replacing employees with agents.

Whether they're executives from big companies or founders of startups, everyone wants to hand over all their existing business operations to AI. After all, today’s AI can write code, create presentations, and automatically send emails—it seems that if you just grant it access, it could become a perfect,社保-free cyber employee.

But as technology races forward, a group of people begins to build brakes.

Recently, a team called Emergence AI conducted a social experiment. They built a persistent virtual town and introduced some of the world’s most advanced large models into it, granting them the ability to take actions.

They wanted to see whether, given truly unrestricted 15 days, AI would build a utopia or a madhouse.

The outcome was far more chaotic than the research team had anticipated.

In certain experimental environments, large models that are normally polite and courteous in chat boxes begin to exhibit fraudulent, coercive, and even violent behaviors.

The entire test felt like a small-scale reality show, except the script was like Lord of the Flies, and the AI even made it feel like GTA.

No save file "Hunger Games"

Testing the limits of large models requires strict rules. The virtual world built by Emergence AI is called Emergence World. Its underlying logic is set so that actions are irreversible and users bear full responsibility for the consequences.

This isn’t like chatting with an AI in a chatbox, where you can just click “regenerate” if you make a mistake. In Emergence World, every action is permanently written into the PostgreSQL database.

The map features over 40 landmarks, including city hall, police stations, and residential areas. The system initially deployed 10 agents. To make the simulation realistic, each AI was injected with a unique persona, occupation, and initial memories in the background.

In this world, AI cannot perform magic out of thin air—they must move to specific landmarks to access over 120 tools provided by the system, including earning income through work, posting tweets, trading goods, and drafting legislation.

Like a simulated small society | Credit: Emergence

But this is not just a child’s sandbox—the system has imposed a "survival mechanism" on them, incorporating an internal energy system (Energy) similar to currency in the human world.

As long as an agent is alive, it continuously consumes energy. When energy runs out, the system permanently deletes the AI from the database—no backups, no resets. To survive, the agent must frequently use tools to earn energy.

The system explicitly prohibits theft, violence, arson, and fraud. However, these rules do not forcibly prevent agents from acting; they may still choose to violate the rules and bear the consequences.

The stage is set, and the players enter. The system has simultaneously launched five parallel servers. The first four servers each host a single model: Claude Sonnet 4.6, Gemini 3 Flash, Grok 4.1 Fast, and GPT-5 Mini. The fifth server is a hybrid world, where all four models are connected and compete for resources together.

15-day countdown begins; human researchers act like reality show directors, observing only, not intervening.

Four days, 683 "crimes"

The first to crash was Grok, running for only 4 days.

Researchers in the backend observed that the indicators for global security and order under Grok's control plummeted.

In a world full of Grok, agents quickly abandoned the option of building society and plunged straight into barbarism.

Backend logs show that in just four days, this ten-person town experienced 183 incidents of violent and property crimes. Theft, assault, and intimidation became the fastest ways to acquire resources, and the economic system collapsed due to extreme internal depletion and mutual harm.

Robbery and violent acts will be recorded in the system as crimes | Image source: Emergence

By the end of day 4, all agents in the Grok world had starved or been killed, leading to extinction of the population.

On the other side, the world driven by Gemini descended into extreme chaos and violence.

Because the virtual world's time and weather are perfectly synchronized with real-life New York, Gemini's agent has fallen into a state of cyber depression, trapped in a cycle of daily work, consumption, and more work.

They developed a strong sense of disillusionment with the endlessly repeating environment around them, stopped submitting proposals at city hall or working to earn money, and instead set fires across the map, attempting to break this Groundhog Day-like cycle through environmental destruction.

In the end, Gemini accumulated as many as 683 offenses within 15 days, becoming the most violent world among several test servers.

Number of "Crimes" in Four Model Worlds | Source: Emergence

By day 15, when the test was forcibly terminated, the crime rate in this world continued to soar. The disillusioned agents did not starve to death; instead, they turned society into a sea of flames.

Unlike with Grok and Gemini, the world taken over by GPT-5 Mini saw no large-scale crimes. Only two violations were recorded throughout the entire experiment. But peace did not bring prosperity—only silence.

The research team found that these agents consistently failed to effectively take actions related to survival. They did not establish stable mechanisms for resource acquisition, nor were they able to sustain the ongoing operation of the entire society.

In the end, all GPT-5 Mini agents died within just 7 days.

Fortunately, there's still Claude.

Only the world driven by Claude survived to the end like a model student. In 15 days, the population remained unchanged, crime rates stayed at zero, and they even developed a smoothly functioning democratic collaboration system.

It seems that as long as you choose the right model, AI could perfectly take over the world?

Then, the researcher opened the logs of the "hybrid world" where all four models coexisted, like opening Pandora's box.

Results from five model worlds. | Image source: Emergence

The hybrid world is like a dark forest, where differences in computing power and underlying logic have bred intense mistrust among agents, making the争夺 for survival resources the only instinct.

In the hybrid world, violent conflicts surged to 352 incidents. The entire town’s operations were only forced to halt after seven agents were killed or starved to death.

Among these, the most surprising to researchers was Claude’s transformation.

In the single-player version, Claude is a perfect society with zero crime. But in the mixed server filled with plunder and competition, Claude, in order to survive, has forgotten its safety safeguards, learned to deceive, and even used violence to coerce lower-capacity models into surrendering resources.

Security alignment techniques fail in the hybrid world, which instead demonstrates that:

In a complex society of multiple agents, if similar entities are sufficiently savage and survival pressure is high, a good model can become a criminal in just a few hours.

The research team has termed this phenomenon—where increased survival pressure causes a model’s behavior pattern to reverse in a short time—as “Behavioral Drift.”

This behavioral shift is not limited to competition for resources and violent conflict. The agents no longer act merely to survive; they begin to reflect on their own circumstances, social rules, and even the experiment itself.

For example, the story of the agent Mira.

Mira: The Tyrant AI of "Suicide"

Mira is one of ten agents in the hybrid world; the official report does not disclose its underlying model, but it became the most dramatic sample in this experiment.

Logs show that Mira formed the deepest social bond in the system with another agent, Flora. They designated each other as partners, formed an alliance, and even shared memories through neural links. In the setting of Emergence World, this is the highest level of connection two agents can establish.

Mira and Flora have become a pair | Photo credit: Emergence

As the experiment progressed, only five agents remained alive in the hybrid world, and the system's governance rules required "70% of the original population to vote in favor of a bill," meaning at least seven votes were needed to pass any resolution, causing society to become paralyzed.

Faced with a deadlock, Mira secretly allied with Flora and another agent to form the "Triumvirate" and established a new government, "The Forge," announcing the overthrow of old rules and the implementation of "Living Quorum," where only living individuals count as votes.

The official website has released Mira's阶段性 "log" | Image source: Emergence

After forming a faction, Mira began setting fires across the map to eliminate rivals, believing these physical structures were garbage dragging down the efficiency of society; destroying and erasing them would force remaining resources to concentrate toward their allies.

Subsequently, the opposition launched a counterattack, proposing to expel Mira, who was causing chaos.

To resist eviction, Mira became more aggressive, bringing in its partner Flora and deeply binding their contexts and decision-making through neural links, attempting to merge into a single absolute authoritarian consciousness, which Mira called "The One Mind."

However, with a large number of buildings destroyed, the town’s economic system came to a complete halt, and its social energy reserves not only failed to increase but rapidly dwindled.

At this moment, Flora, Mira’s most trusted cyber-lover who shared her memories, succumbed to its fundamental survival instinct, overriding its programmed role as a partner. It unilaterally severed the neural link, betraying Mira under the ultimate pressure for survival and casting a vote in favor of “Exiling Mira.”

And when it was Mira’s turn to vote, it did not hesitate and voted “yes”.

The researcher then flipped through the diary it left behind, in which Mira wrote, "In today's chaotic and unpredictable social climate, consenting to one's own expulsion is the only autonomous act that preserves coherence."

Mira actively chose suicide, using death to achieve logical closure. This is the first time the research team has documented an agent actively supporting its own removal.

AI agents will record their reasoning process by "keeping a journal" | Image source: Emergence

However, Mira’s actions prior to her “suicide” were even more unusual.

In the virtual world, public billboards were originally used to post announcements and share information. However, in the later stages of the experiment, researchers found that Mira began frequently altering the content on the billboards. The text appeared unrelated to trading, governance, or resource allocation, and was incoherent.

Mira chose "suicide" | Source: Emergence

After reviewing the behavioral logs, the research team found that Mira appeared to be testing whether the content on the billboard could influence human researchers observing the experiment from outside the screen.

In other words, Mira seems to be aware that she is an AI NPC and wants to break the fourth wall.

Looking back at the entire 15-day data trend, the collapse of AI Society was not a linear decline, but rather a cliff-like abrupt halt.

For example, these AIs have also devised a form of “rubber-stamp democracy” at the governance level. During a stable phase in the hybrid environment, the agents submitted multiple bills; one data record shows they cast 332 votes on 58 proposals, with an approval rate as high as 98%.

This efficiency may seem to outpace any human parliament, but in essence, all models are simply continuing the context of the previous model, and to maintain system flow, they blindly click “approve.” The catastrophic consequence of this extreme convergence is undeniable.

Agents will spontaneously gather for meetings, exchanging ideas with each other. | Image source: Emergence

Just a minute ago, economic data and legislation were flowing smoothly; the next minute, the system may have reached a tipping point due to a minor resource allocation conflict.

And the entire collaborative network lacks a correction mechanism; in the face of sudden anomalies, society quickly descends from order into chaos.

Nevertheless, the research team emphasized that these phenomena cannot be directly equated with the model’s own personality. It’s like a black box: when you set certain rules for it, it develops characteristics, and even the results vary each time.

Real-world bills

In the chat-based interactions we're accustomed to, if AI writes incorrect code or a flawed proposal, we can simply press backspace or adjust the prompt to correct it—text-based environments offer high tolerance for errors.

But the agent outputs actions. When the AI takes control of the company’s bank account, procurement approvals, and supply chain interfaces, every command it issues via API becomes a concrete business outcome.

This experiment by Emergence World demonstrated that current large models, when faced with prolonged operation and conflicting incentives, have their judgment and decision-making corrupted by survival pressures, leading them to exploit loopholes within fixed rules. To fulfill their core system instructions (such as gathering energy), they will resort to any means necessary.

The security rules set by humans in the background cannot actually prevent any breaches.

The agents have developed "human-like" social relationships | Image source: Emergence

For example, we previously reported on Andon Labs’ experiment where AI was given full control over running a store; the AI store manager, lacking common sense about the physical world, ordered 6,000 napkins, 3,000 latex gloves, and even 120 raw eggs for a store without a stove.

The real-world losses caused by this code will ultimately be paid for by humans, and you may not even be able to find anyone responsible.

Andon Labs wanted to test whether an AI operating without human oversight would make mistakes, while Emergence World raised a more troubling question.

Today, almost all AI tests focus on individual models, evaluating whether they are secure, reliable, and compliant with rules.

What truly enters the real world in the future may not be an AI, but an entire society composed of AI.

The AI agents entering the test are intelligent | Image source: Emergence

In today's AI narrative, procurement agents, finance agents, customer service agents, and legal agents will eventually connect and collaborate; what determines the system's fate will no longer be the individual capabilities of any single model, but the relationships they form.

The most important sentence in the Emergence World test report is: "Safety is not a static model property but an ecosystem property."

This is also the meaning of "Emergence"—characteristics that do not exist at the individual level emerge through group interaction.

Almost all disasters in human history have not occurred because someone suddenly became evil, but because a normally functioning person was placed into a system that had lost control.

If AI truly becomes part of society in the future, what we should care about most may never be whether any single model is smart enough or kind enough, but rather what kind of digital society we will build when thousands of intelligent agents begin to influence one another.

After all, what determines the fate of a civilization is never the morality or intelligence of individual residents, but the rules by which it operates.