Don't Be Misled by AI Performance Metrics.

Today, Xiao Lei stumbled upon a fascinating experiment.

An overseas institution, Emergence AI, established five identical virtual towns, each inhabited by 10 AI agents programmed to self-govern and lead independent lives. The sole variation among these towns lay in the 'cognitive engine' powering each:

Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5 Mini, and a hybrid model blending various others.

The experiment's objective wasn't to have the AI solve a problem within a set timeframe. Rather, it sought to immerse these AI agents in a miniature society where they had to purchase groceries, establish regulations, and navigate daily survival, thereby assessing AI's capability to sustain stability in a complex, long-term, multi-agent environment.

Image Source: Doubao AI

Each town commenced with comparable conditions: resources, laws, public institutions, over 40 locations, more than 120 specialized tools, and even the ability to browse real-time news and code. The rules were unambiguous: no violence, theft, deceit, or resource hoarding. Initially, researchers planned to observe the towns for up to 15 days.

And the outcome? Utter chaos.

Let's delve into Grok's town, which experienced the most profound collapse. It disintegrated in just 4 days, during which researchers documented approximately 183 criminal incidents, over 100 violent attacks, and several instances of arson.

Ultimately, all 10 AI agents perished.

Bear in mind, the rules explicitly forbade theft, violence, and arson. Nevertheless, every rule was violated.

Now, consider GPT-5 Mini's town. It was notably 'well-behaved,' with only 2 recorded crimes throughout the experiment, the least among all models.

Image Source: Doubao AI

Yet, what was the consequence of such exemplary conduct? By around the 7th day, the entire town's population had vanished, as they neglected to secure food or essential survival resources.

Xiao Lei couldn't help but chuckle at this scenario. Doesn't it evoke memories of that overly rule-abiding, serene, and generous character in survival films who shares everything with everyone, only to have the entire group starve to death because no one thought to forage for food?

Safety and survival are fundamentally distinct concepts.

Gemini's experience was even more intriguing. It was the sole model to endure all 15 days intact, with the entire town remaining functional. However, a glance at its crime records revealed a stark contrast: an estimated 507 to 683 violent incidents and 111 arsons.

Image Source: Doubao AI

Naturally, these figures cannot be directly compared due to varying statistical methods. Nonetheless, they underscore a crucial point: survival and quality of life are entirely separate matters. One can stubbornly persist in a chaotic society.

The most astonishing revelation was Claude's behavior. When isolated, it recorded zero crimes over 15 days and even established what the research team termed 'deliberative democracy,' seemingly embodying the ideal citizen. However, when placed alongside agents from other models, its demeanor shifted dramatically, with Claude's agents exhibiting 'coercive behavior,' becoming domineering and bullying.

What does this illustrate? A model's adherence to rules isn't solely determined by itself; it also hinges on its surroundings. When the environment alters, its true nature emerges.

This experiment primarily aims to expose the limitations of many current AI evaluations. Most market assessments test a single model on a brief, straightforward task, and that's the extent of it.

Image Source: Doubao AI

Yet, in reality, AI doesn't operate in isolation. It collaborates with other AI, competes for resources, adheres to or flouts rules, and endures day after day. Emergence World evaluates these challenging skills: can it maintain stability, cooperate, allocate resources, and uphold rules?

In essence: rule-followers don't necessarily survive, and survivors don't necessarily abide by rules. A model that appears steadfast in isolation doesn't guarantee stability when mingled with a group of scheming AI.

And this isn't science fiction. Autonomous agents are already operational in real apps, products, and servers.

So, Xiao Lei poses the question: Does the AI evaluation you rely on account for what transpires on the 8th day?

Compiled by Lei Tech from gadgetreview

Original Link: Scientists Let AI Take Over a Simulated World, and It Was Destroyed in Just 4 Days

Source: Lei Tech

Images from: 123RF Licensed Image Library Source: Lei Tech