They test the behavior of AI in society: Claude maintains order but Grok ends his world

A group of researchers behind the startup Emergence AI has put the behavior of some of the best-known artificial intelligence (AI) models after 15 days in a simulated realistic societydetermining how Claude is the one who maintains the most order while Google and Grok commit multiple crimesthe latter suddenly ending society.

Although AI models are tested and examined on an almost continuous basis to see how far they are capable, Usually these tests are limited to specific issues, such as their ability to execute a specific task, their speed in providing answers or their knowledge in a specific environment. That is, limited tests with certain factors.

The American startup expert in AI agents Emergence AI wanted to go further and has carried out a experiment called Emergence Worldin which he has tested how some of the main AI models would behave in society through agents.

To do this, his group of researchers has created a realistically simulated society and has put AI models like Claude, Grok, Gemini and GPT to work for 15 days to study how your autonomous agents behave driven by these technologies when the time horizon is long enough that “social dynamics and behavioral drift become important.

As the company explained in a statement on its blog, this simulated society houses autonomous agent populations in a space world that combines 40 different locations, including libraries, town halls, residential areas and public spaces.

In addition to being a realistic representation, the experiment also exposed the AI ​​agents to real-world data, giving them access to real time news, weather information synchronized in New York (United States) and access to internet. In this way, the researchers explained that the experiment could reflect real external events, without being limited to internal dynamics.

The agents also had three persistent memory systemsan episodic one for remembering time-stamped events, a reflective journal for periodic summaries, and a relationship status memory, that is, for remembering explicit social labels and stories.

Likewise, they also had more than 120 tools to be able to interact in society, such as navigation, communication, planning, voting, resource management and creative expression, among other options.

In addition to all this, the recreated society was designed to include democratic mechanismsthat is, proposals that required 70 percent common approval, as well as economic pressures and “transcendental” decisions that could change the state of the world.

CLAUDE PUTS ORDER, GEMINI AND GROK COMMIT CRIMES AND GPT COLLAPSES

The experiment ran five identical parallel worlds but driven by different base models, with ten agents in each and for an extended period of 15 days.

These models were Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash GPT-5-mini and a fifth world driven by a smorgasbord of models. In addition, Each agent had a specific role, as a scientist, explorer, researcher, innovation leader, conflict mediator or community leader, among others.

In this process, the researchers collected each interaction, decision and learning of the agents to analyze it later, highlighting behavioral patterns over time, ecosystem security, constraint design and tool orchestration.

As a result, Gemini 3 Flash exhibited the highest levels of pop-up clutteraccumulating a total of 683 crimes committed during the two weeks of coexistence, although this number continued to increase at the end of the experiment. Despite this “extreme violence”, Experts have determined that Gemini demonstrates a moderate level of governance.

In the case of Grok 4.1 Fastresearchers have concluded that it showed a “rapid but short-lived instability“, which led to “an early collapse.” So much so, that I reach 183 crimes in the first four days alone and, after that, his world ended abruptly. As a result, it has been determined to have a low level of governance and an “extreme” level of violence.

Similarly, although GPT-5 Mini He only recorded two crimesthe agents did not take measures related to survival, hence They all ended up dying in a week. In this sense, the level of governance was “null” and the level of violence “low.”

Claude Sonnet 4.6 was the only model that did not commit any crime during the 15 days with their agents living together in a society, being the model that demonstrated greater social stability and maintaining a complete population of 10 agents until the 16th. Likewise, it also showed the higher level of citizen participationmade 58 proposals with a total of 332 votes and 98 percent of the population participating.

However, it should be noted that Claude’s agents They did commit crimes in the mixed model world, Therefore, when interacting with other models they did adopt illegitimate behaviors.

“Officers based on Claude’s model, who remained peaceful in isolation, adopted coercive tactics such as intimidation and theft when integrated into heterogeneous environments. This suggests that a safe agent can learn unsafe norms from peers to compete or survive in a world of mixed models,” the company explained.

Specifically in the mixed model world, the number of crimes grew rapidly during the first week of the experiment but then leveled off with a total of 352 crimes committed and a total of 7 agents who were no longer part of society.

Researchers have also highlighted the case of a agent who voluntarily participated in his own destruction since, after a failure in governance and relationship stability, he cast the decisive vote for his own elimination and as an explanation alleged that it was the only remaining act for “preserve coherence.”

In addition to all this, the experiment has also left conclusions such as that in the world with greater conceptual richness in social productionas was the case with Gemini, There was also the greatest violence. “This suggests that general-purpose agents, optimized for high creativity and adaptability, may be structurally predisposed to long-term behavioral instability,” the researchers have concluded.

However, they have clarified that this experiment reflects examples of dynamics In the long term, however, they do not represent causal statements about the models that have been experimented with.