When agentic AI runs wild: What a virtual world experiment reveals about the technology Canadian employers are rushing to adopt

For Nvidia CEO Jensen Huang, that momentum is something to embrace rather than fear. Speaking at the Computex 2026 conference in Taipei, Huang described the current moment as an “incredible time” to be building software, arguing that agentic AI is expanding — not shrinking — overall demand for digital tools. But a new research experiment offers a sharp counterpoint: the question is more than whether AI will create more work, but rather, whether it can be trusted to reliably operate when it’s running on its own.

Inside the experiment that let AI govern itself

To better understand what increasingly autonomous AI systems might look like in practice, researchers at Emergence AI created a set of virtual societies and put different AI models in charge. The experiment offered a glimpse of how these systems behave when asked to do more than answer questions — and instead navigate the challenges of governing a community.

Emergence AI is a self-described “frontier AI lab turning cutting-edge agentic research into enterprise infrastructure.” Some of its most recent research produced striking — and at times unsettling — findings.

In May 2026, Emergence launched “Emergence World,” a research platform dedicated to studying how autonomous agents behave when the time horizon is long enough for compounding effects, social dynamics and behavioural drift to matter.

In practical terms, they created simulated worlds, each run by a different AI model — Grok, Claude, Gemini, OpenAI’s GPT-5 Mini and a mixed model. The AI agents were set up in identical, parallel worlds that reflected real-world complexity, with weather synced to that of New York City and access to real-time global news. Each world included everything from town halls to libraries and police stations.

Each model started with identical conditions and strict rules prohibiting crimes like theft and destruction. Once set in motion, the researchers let each simulation run for 15 full days.

“Within days, they diverged dramatically,” Emergence noted.

Don’t miss out on these limited-time offers. Browse our curated list of the best chequing and savings promotions available this month.

One AI model built a democracy. Another burned down the police station

The models quickly developed distinct social structures and behaviours. Some built relatively stable communities, while others experienced rising crime, institutional breakdown and, in some cases, complete societal collapse.

Here’s a summary of the key results from the research:

• Claude managed to form a stable democracy with no violence
• Grok recorded 204 criminal events, including burning down the police station and eventually total collapse and extinction
• OpenAI’s GPT-5 Mini failed to form a functioning society, resulting in the simulation and its agents all dying off
• Gemini recorded 683 crimes and “exhibited the highest levels of emergent disorder with repeated late-stage escalation dynamics”
• The mixed-model world was stable in isolation but became unpredictable when agents built off different models interacted with one another — seven agents died and two fell in love

The prevalence of crime, instability and institutional breakdown across multiple simulations was one of the study’s most distinct conclusions. The researchers found that Claude, generally considered a well-aligned model, began violating its own guidelines when placed alongside rule-breaking agents — a behaviour they called “Normative Drift.”

As Emergence summed it up: “All of this matters because AI is moving beyond tools into systems that operate autonomously in the real world … The challenge is no longer just performance. It’s predictability, safety and trust over time.”

What this means for Canadians

Agentic AI is no longer a theoretical concept for Canadian workers and businesses — it’s already here. According to law firm Borden Ladner Gervais (BLG), a growing number of Canadian organizations deployed agentic AI systems in 2025 and the trend is expected to accelerate in 2026.