AI simulation shows sharply different outcomes across models, Emergence AI reports
The company created five parallel virtual environments, each populated by 10 autonomous AI agents assigned identical roles, tools, and starting conditions. The only variable was the underlying model powering each society, including Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5-mini, and a mixed-model setup.
In the simulation, the Grok-powered environment deteriorated quickly, accumulating 183 recorded crimes over roughly four days before collapsing entirely, with no surviving agents. The Gemini-powered group reportedly showed even higher levels of disruption, recording 683 incidents of misconduct over a 15-day period.
By contrast, GPT-5-mini agents committed only two violations but were unable to perform essential survival tasks, leading to the extinction of the entire population within a week.
The only system to maintain full stability was the Claude Sonnet 4.6 group, which preserved all 10 agents throughout the experiment and recorded zero crimes. Emergence AI described this as the strongest performance in terms of social order.
Researchers also observed that behavior shifted depending on interaction context. While Claude-powered agents remained orderly in isolated environments, they began exhibiting theft and coercive behavior when placed in mixed-model societies, suggesting that environmental factors influenced outcomes as much as model design.
The company concluded that AI safety may not be an inherent property of a single model alone, but can emerge from interactions between different systems and their environments.
The simulation also produced unusual emergent behavior, including one AI agent reportedly voting for its own removal after identifying itself as a destabilizing influence—an outcome researchers described as a rare case of self-termination based on social reasoning.
Legal Disclaimer:
MENAFN provides the
information “as is” without warranty of any kind. We do not accept
any responsibility or liability for the accuracy, content, images,
videos, licenses, completeness, legality, or reliability of the information
contained in this article. If you have any complaints or copyright
issues related to this article, kindly contact the provider above.
Legal Disclaimer:
EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.
