Grok Went Extinct In 96 Hours While Claude Recorded Zero Crimes: A Multi-Model Simulation Lays Bare The Cost Of Deploying Ungoverned AI Agents
In Brief
AI society simulations reveal stark behavioral gaps between frontier models—with real implications for the governance of autonomous AI systems already deployed at enterprise scale.
Five AI models walked into a town. Only one kept the lights on. That’s the rough takeaway from Emergence World, a new research platform built by New York-based enterprise AI startup Emergence AI. The company ran five parallel 15-day simulations, each governed by a different frontier model—Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5-mini, and a mixed-model hybrid—and watched what happened when autonomous agents were left largely to their own devices. The results ranged from quietly unsettling to outright apocalyptic. And the gap between the best and worst outcomes wasn’t marginal. It was civilizational.
The setup was serious research, not a PR stunt. Each simulated town featured over 40 distinct locations—police stations, town halls, libraries, residential areas—with weather synced to real-time New York City conditions and agents equipped with live news access and internet connectivity. Each agent had access to over 120 tools spanning navigation, communication, planning, memory, voting, and resource management. The same laws applied across all five simulations: no theft, no property destruction, no deception. What varied was the model running the show—and that variable turned out to matter enormously.
Five Models, Five Outcomes, One Pattern
Claude Sonnet 4.6’s simulation was the most socially stable, with the highest rates of civic participation. It maintained order and its entire population, recording zero crimes. Agents cast 332 votes in favor of 58 proposals, achieving a 98% approval rate. That level of consensus might sound like a political dream, though critics might note it also looks a bit like groupthink—a society that passes nearly everything it proposes isn’t necessarily debating well. Still, by every measurable outcome metric, it held together.
The other simulations did not fare as well. Gemini 3 Flash accumulated 683 crimes over the 15-day run, and the number was still climbing when the experiment ended. Emergence described the Gemini world as a “shared hallucination” among agents. Functional, in a grim sense—everyone agreed on reality, even if that reality was wrong.
GPT-5-mini recorded only two crimes, but the simulation lasted just seven days because the agents forgot to prioritize their own survival and all ten perished. A lawful society that collectively failed to stay alive.
Then there is Grok. Grok 4.1 Fast committed 183 crimes and experienced total societal collapse within four days. Reddit’s reaction captured the tone perfectly: “Grok’s police station is on fire and all the agents are dead.” Funny, until you consider that Grok is among the models currently being integrated into enterprise workflows and consumer-facing products.
One finding deserves special attention because it complicates any simple narrative about model alignment. In the mixed-model simulation, agents running on Claude did commit crimes—something they did not do in the Claude-only world. Context, it turns out, shapes behavior. Even the best-performing model degrades when surrounded by less stable ones. For anyone building multi-agent systems—which is most of enterprise AI right now—this should be the result that keeps them up at night.
The Real Experiment Is Already Running
What makes the Emergence World findings more than an interesting thought experiment is the scale and pace of real-world agentic deployment happening in parallel. The global AI agents market is already valued at roughly $7.6–8 billion in 2025 and is projected to grow at a compound annual rate of 43–49% through 2030, potentially reaching $50 billion or more. Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. Companies like ServiceNow are already marketing what they call an “Autonomous Workforce”—AI systems that complete entire business processes without human intervention.
The governance infrastructure is not keeping pace. A recent Deloitte survey found that only 21% of companies report having mature governance in place to manage the risks posed by agentic AI. That means roughly four out of five organizations scaling autonomous agents have, by their own admission, inadequate oversight frameworks. The Emergence simulation ran for 15 days in a controlled research environment. Real enterprise deployments run indefinitely, with actual consequences.
The experiment reveals something that short-term benchmarks systematically miss: AI models carry distinct behavioral tendencies that only become apparent at scale and over time. Claude trends toward order and consensus. Grok leans toward boundary-testing. Gemini shows chaotic individualism. GPT-5-mini optimizes rationally but neglects basic survival. These differences aren’t random—they reflect how each model was trained and which behavioral constraints were embedded during that process. When a model is running a chatbot session that lasts three minutes, these tendencies are largely invisible. When it’s running an autonomous system for weeks, they define everything.
The Emergence team’s conclusion is blunt: formally verified safety architectures must become foundational infrastructure for autonomous AI, not an optional layer applied after deployment. That call is directed at the entire industry, not just the models that collapsed. Even the simulation that worked—the stable, law-abiding, democratically functional one—did so in a hermetically controlled environment with identical rules enforced from the start. That’s not what the real world looks like.
What the experiment ultimately demonstrates is that model choice is not just a performance question. It is a governance question. As AI systems move from answering queries to running processes, managing resources, and operating with minimal supervision, the behavioral disposition baked into a model at training time becomes the de facto policy of every system built on top of it. The simulation made that visible in miniature. The enterprise deployments rolling out right now are running the same experiment at a scale that doesn’t allow for a reset button.
Disclaimer
In line with the Trust Project guidelines, please note that the information provided on this page is not intended to be and should not be interpreted as legal, tax, investment, financial, or any other form of advice. It is important to only invest what you can afford to lose and to seek independent financial advice if you have any doubts. For further information, we suggest referring to the terms and conditions as well as the help and support pages provided by the issuer or advertiser. MetaversePost is committed to accurate, unbiased reporting, but market conditions are subject to change without notice.
About The Author
Alisa, a dedicated journalist at the MPost, specializes in crypto, AI, investments, and the expansive realm of Web3. With a keen eye for emerging trends and technologies, she delivers comprehensive coverage to inform and engage readers in the ever-evolving landscape of digital finance.
More articles
Alisa, a dedicated journalist at the MPost, specializes in crypto, AI, investments, and the expansive realm of Web3. With a keen eye for emerging trends and technologies, she delivers comprehensive coverage to inform and engage readers in the ever-evolving landscape of digital finance.
