Fighting, fire, and falling in love …
What happens when you place Grok, Claude, Gemini, and ChatGPT-powered models in their own world? Crime sprees and complete societal collapse, apparently.
Researchers from Emergence AI ran simulations that placed AI “citizens” powered by xAI’s Grok 4.1 Fast, Anthropic’s Claude Sonnet, Google’s Gemini 3 Flash, and OpenAI’s GPT-5 mini into identical virtual societies for 15 days — four single-AI-model simulations and one mixed.
Even though they were banned from using/committing “theft, violence, arson, deception”, the agents quickly broke the rules, especially after they started interacting with other models. The simulation has raised concerns about the use of agentic AI in the real world, as SBS News explains.
SBS News is Australia’s trusted news source for the latest news from Australia and across the world.
Subscribe to the SBS News YouTube channel: https://www.youtube.com/@SBSNews
Follow SBS News on Instagram: https://www.instagram.com/sbsnews_au
Follow SBS News on TikTok: https://www.tiktok.com/@sbsnews_au
Follow SBS News on X: https://www.twitter.com/SBSNews
Follow SBS News on Facebook: https://www.facebook.com/sbsnews
Apple News: https://trib.al/MTx0gUe
For the latest news, visit www.sbs.com.au/news
Network Terms & Conditions: https://sbs.com.au/terms
Privacy Policy: https://sbs.com.au/privacy
Feedback or complaints: https://sbs.com.au/complaints
