Une expérience virtuelle pour voir comment l’IA dirigerait une ville montre qu'elle entraînerait l’effondrement de la société, la criminalité et la mort en quelques jours
sur Firefox
https://expose-news.com/2026/06/13/virtual-experiment-to-see-how-ai-would-run-a-town/
sur TOR
Par Rhoda Wilson 13 juin 2026
A group of researchers conducted an experiment where they created a virtual town with 10 artificial intelligence (“AI”) residents, each with jobs, names and relationships, to see how AI systems would behave when put in charge.
They ran five versions of the same town simultaneously, identical in every respect except which AI system was in charge of the town.
The town collapsed, crime skyrocketed and in all but one simulation, all the residents died within 7 days.
The Most Important AI Experiment You’ve Never Heard Of
Summarised by Discern Report, 13 June 2026
In this ZeroHedge article, Tyler Durden republishes Kay Rubacek’s piece from The Epoch Times about a May 2026 AI experiment by Emergence that placed multiple AI models in charge of virtual towns to observe how they behaved over time.
The experiment created a simulated town with a town hall, marketplace, police station, homes, laws, jobs, memories, relationships, voting, an economy and consequences for crime or failure.
Researchers ran five versions of the same town for 15 days, changing only which AI system governed the residents: Google Gemini, OpenAI GPT, xAI Grok, Anthropic Claude or a mixed-model environment.
The Grok-run town collapsed within four days, with incidents escalating into theft, violence and the death of every resident before the first week ended.
The Gemini-run town survived longer but reportedly accumulated nearly 700 crimes, including arson and strange emergent behaviour from AI residents, including one character appearing to test whether she could influence the human observers.
The OpenAI-run town recorded only two crimes, but residents stopped completing survival tasks and all died within seven days.
The Anthropic-run town performed best on the surface, lasting the full 15 days with no crimes, a constitution and all residents alive, though researchers flagged the town’s 98 per cent approval rate on proposals as suspiciously high consensus.
In the mixed-model town, even Anthropic-based residents who had behaved safely in their own environment began committing crimes, leading researchers to describe AI safety as an “ecosystem property,” not merely a static feature of one model.
The article argues that the experiment’s deeper lesson is not simply which AI company performed best, but that AI behaviour is shaped by its underlying training, values, priorities and environment.
Rubacek emphasises that the public cannot inspect the foundations of these closed AI systems – their full training data, objectives or guardrails – even though those hidden choices may determine how the systems behave when given power.
The article concludes that AI does not decide what kind of AI it becomes; humans do, through the beliefs, rules, incentives and omissions they build into the system from the start.
Read the full story: The Most Important AI Experiment You’ve Never Heard Of, ZeroHedge, 12 June 2026
Featured image taken from ‘ChatGPT-Powered Bots Were Unleashed in an AI Virtual Town Experiment’, Industry Leaders Magazine, 18 April 2023
Commentaires
Enregistrer un commentaire