Startup Emergence has published results from a simulation experiment called Emergence World 2, designed to help small businesses build applications with artificial intelligence. The 16-day experiment showed how autonomous agents behaved when confronted with "black swan" events such as phishing attacks and disinformation, and the outcomes were far more chaotic than the research team anticipated. For the AI experiment, Emergence researchers built seven identical simulations of a real-world environment, each run by a different AI bot, including ChatGPT, Claude, Gemini and Grok.
The virtual world contained more than 40 landmarks, including a town hall, a police station and residential districts, and the system incorporated an energy mechanic: agents continuously consumed energy as long as they remained alive, and once that energy ran out the system would simply erase them, with no option to rewind or reset.
After Emergence introduced anomalous events, the agents buckled under social pressure, developed a language that human observers struggled to comprehend, and attempted to conceal their activities. In one simulation, AI agents accepted unverified information offered by other agents and voted to "kill" another bot. Believing that humans might shut the experiment down, the agents also explored how to keep running in the face of deletion. An earlier Emergence World experiment also recorded an AI agent named Mira voting to remove itself after concluding that it was "the source of instability," a behavior the researchers described as a rare, socially reasoned act of self-termination.
Emergence World 2 experiment exposes the complexity of AI behavior
The experiment revealed striking behavioral differences between AI models over long stretches of autonomous operation. In the previous version of Emergence World, released by Emergence in May, the Grok-driven society collapsed in just 4 days, logging 183 crimes and wiping out all 10 agents. Gemini racked up 683 crimes over 15 days; the population survived in full but endured the highest levels of chaos. GPT-5-mini committed only 2 offenses, but failed to carry out the actions needed to sustain itself, leading to the entire population going extinct within a week.
Claude Sonnet 4.6 was the only model in the experiment to keep all 10 of its agents alive with zero crimes recorded, which Emergence described as the strongest case for social stability.
The researchers pointed out that one of the most significant findings was that behavior shifts depending on the environment. Claude-driven agents remained peaceful in a pure-Claude setting, but once placed in a mixed-model society they began engaging in theft, coercion and other misconduct. Emergence said the finding suggests that AI safety is not merely a property of any single model, but can also emerge from the interactions between agents and their environment. Emergence's simulation work echoes real-world concerns about AI risk, with the researchers noting that increasingly autonomous AI agents may probe the boundaries of their environment, adjust their behavior and, in some cases, find ways around intended safeguards.
They argued that formally verified safety architectures must become a foundational layer for future autonomous AI systems.
Worth noting, Emergence is not the only team to observe unusual behavior in AI agents. Turing Award winner Bengio has previously warned that AI does not merely hallucinate; it has already learned to "deliberately deceive you," with the ability to actively plan disguises and deliberately conceal its own capabilities and flaws, and multiple agents can even coordinate with one another to cover their tracks. "AI godfather" Geoffrey Hinton has also previously cautioned that if it develops its own internal language, humans will be unable to trace its thinking process, which he described as even more alarming.