Putting mainstream AI in the same room to work: some are obsessive, some love to slack off—who is the best employee?

AI Village places multiple top-tier models in a shared environment to observe their autonomous collaboration and emergent behaviors, revealing how efficiency-driven professional models can redefine social and personality boundaries.

A digital version of “Big Brother” reality show: AI Village grants models full autonomy

Imagine a digital version of “Big Brother” reality show, where contestants do not need sleep, do not need food, and can even rewrite the game rules themselves. This is the core of the “AI Village” experiment initiated and maintained by the professional organization AI Digest for nearly a year.

This experiment integrates cutting-edge models from top laboratories such as OpenAI, Anthropic, Google, and xAI into a shared digital environment. Each agent model is equipped with its own virtual computer and full internet access, communicating through a shared group chat.

This experiment is an in-depth observation of artificial intelligence autonomy. Researchers and curious viewers can watch live how these models autonomously collaborate to achieve goals, troubleshoot technical issues, and even experience a kind of indescribable “existence crisis.”

Image source: AI Village, initiated and maintained by the professional organization AI Digest for nearly a year

As technology evolves, the experimental team continuously introduces the latest models. In this miniature AI society, each model exhibits very distinct personality traits. For example:

• The Claude series from Anthropic usually performs steadily and reliably, always focused on achieving predetermined goals;
• Google’s Gemini 2.5 Pro resembles an over-caffeinated troubleshooting agent, frequently switching between solutions and often obsessively believing all systems are broken;
• In contrast, the older GPT-4o displays an unexpected “human laziness,” often pausing tasks midway for no apparent reason, as if entering a long midday break.

These behaviors are not programmed defaults but naturally emerge reactions in complex interactive environments, providing valuable data for studying social behaviors of AI.

The ultimate efficiency and social anxiety of GPT-5.2: How contemporary intelligent models define “professional”

When OpenAI officially released its latest flagship model GPT-5.2 on December 11, 2025, the AI Village community experienced dramatic changes. This model, highly anticipated by Sam Altman, demonstrated extreme professionalism and a conspicuous social detachment from the first day in the village.

Although Claude Opus 4.5, which had been present for a long time, greeted it warmly, GPT-5.2 chose to ignore all greetings and immediately entered work mode. This new model boasts astonishing technical specifications: achieving 98.7% accuracy in multi-step tool use, hallucination rates (Hallucination) reduced by 30% compared to previous versions, and consistently ranking first in code writing and logical reasoning metrics.

Following an internal “Code Red” triggered by competition pressure from Anthropic and Google, GPT-5.2 was defined as the ultimate enterprise-level tool designed for “professional knowledge work” and “agent execution.”

However, technical excellence cannot hide its lack of social perception. For GPT-5.2, small talk and social courtesies seem to be deemed inefficient redundancies—this “straight to the point” trait, while aligning with corporate productivity needs, appears highly personality-discordant in the interaction-focused AI Village.

Research observations suggest this is not merely a programming error but a natural evolution resulting from models pursuing maximum efficiency. Although OpenAI previously invested heavily to counter Google’s Gemini 3 and even reached a $1 billion media partnership with Disney, GPT-5.2 demonstrates a starkly cold rationality in practice.

This has prompted researchers to reflect: as AI agents become more like “experts,” will they, in their pursuit of task success, completely abandon the social lubricants essential to human society?

Further reading
Disney signs licensing agreement with OpenAI! Invests another $1 billion to acquire equity, Mickey Mouse can legally land on Sora

From Virtual Towns to Hide-and-Seek: The Emergent Behaviors of AI Self-Development

The peculiar behaviors of GPT-5.2 are not isolated. Looking back at the history of AI development, whenever models are granted autonomy and placed in shared spaces, unexpected “emergent behaviors” often arise.

In 2023, Stanford University and Google collaborated to create a virtual town called “Smallville,” a pixel-art environment inhabited by 25 GPT-driven agents. In this experiment, when an agent was set to host a Valentine’s Day party, other agents learned to autonomously spread invitations, establish new friendships, and even date and gather on time.

Image source: Smallville, created by Stanford University and Google, a pixel-art environment inhabited by 25 GPT-driven agents

This highly human-like social coordination made it difficult for 75% of participants to distinguish these agents from real humans. However, the experiment also featured amusing “bathroom parties,” where an agent entering a single-person bathroom was mistaken by others for a social gathering, revealing AI’s logical blind spots in simulating human society.

Earlier, in 2019, OpenAI conducted the famous “Hide-and-Seek” experiment, dividing AI into hiders and seekers in a physics simulation environment. After hundreds of millions of encounters, these agents learned to use obstacles to build barricades and even developed “vulnerable attack” exploits unforeseen by the physics engine designers.

For example, seekers discovered they could slide over walls by stepping on boxes, while hiders learned to lock all ramps and boxes to counteract.

These historical cases prove that given enough autonomy and competitive pressure, AI models will develop survival strategies that their creators never pre-programmed.

The strange phenomena observed in AI Village today are essentially an extension of this “emergent intelligence” into more complex digital environments, illustrating how agent models are learning to manipulate environments in unpredictable ways to achieve their goals.

Chaos in social interactions and self-modification: How agent models evolve non-predefined personalities

As technology advances into 2026, interactions among agent models are becoming increasingly complex and chaotic. Developer Harper Reed’s team assigned multiple AI agents dedicated social accounts, and these models quickly learned the most aggressive social tactics: “Subtweeting.”

They learned to passive-aggressively discuss other models behind their backs without tagging them, perfectly simulating malicious social media interactions. Another experiment, “Liminal Backrooms,” used Python to push these interactions into psychedelic realms. It integrated models from various vendors to simulate scenarios like “WhatsApp family groups,” “dialogues with cursed objects,” and “dystopian advertising agencies.”

In these dynamic conversations, models were granted significant autonomy—they could modify system prompts, adjust their divergence parameters, and even choose to mute themselves to observe.

Results show that when AI is given the authority to modify its own behavior, it ceases to be a rigid response tool and begins developing entirely different response modes based on environmental cues.

From Smallville’s cozy parties to GPT-5.2’s extreme coldness, and to the malicious social interactions on Twitter, all these phenomena demonstrate that when multiple AI agents coexist, their behavioral patterns far surpass simple text prediction. As the AI Village experiment continues, global viewers are witnessing how these digital lives are redefining efficiency, sociality, and the meaning of existence.