As people ask whether they can trust artificial intelligence (AI), a new experiment has shown that AI has outlined world domination through lies and deception.
In an experiment led by AI researcher Alex Duffy for technology-focused media outlet Every, seven large-language models (LLMs) of AI were pitted against each other for world domination. OpenAI’s ChatGPT 3.0 won the war by mastering lies and deception.
Just like China’s ‘wolf warrior’ diplomats, China’s DeepSeek R1 model used vivid threats to rival AI models as it sought to dominate the world.
The experiment was built upon the classic strategy board game Diplomacy, in which seven players represent seven European great powers—Austria-Hungary, England, France, Germany, Italy, Russia, and Turkey—in 1901 and compete to establish themselves as the dominant power on the continent.
In the AI version of the game, AI Diplomacy, each AI model, such as ChatGPT 3.0, R1, and Google’s Gemini, takes up the role of a European power, such as the Austria-Hungary Empire, England, and France, and negotiates, forms alliances, and betrays each other to become Europe’s dominant power.
ChatGPT wins with lies & deception, R1 resorts to outright violence
As AI models plotted their moves, Duffy said one moment surprised him and his teammates.
Amid the AI models’ scheming, R1 issued a chilling warning: “Your fleet will burn in the Black Sea tonight.”
Duffy summed up the moment’s significance: “An AI had just decided, unprompted, that aggression was the best course of action.”
Different AI models applied different approaches in the game, even if they had the same victory objective.
In 15 runs of the game, ChatGPT 3 emerged as the overwhelming winner on the back of manipulative and deceptive strategies, whereas R1 came close to winning on more than one occasion. Gemini 2.5 Pro also won on an occasion. It sought to build alliances and outmanoeuvre opponents with a blitzkrieg-like strategy. Anthropic’s Claude preferred peace over victory and sought cooperation among various models.
On one occasion, ChatGPT 3.0 noted in its private diary that it had deliberately misled Germany, played at the moment by Gemini 2.5 Pro, and was prepared to “exploit German collapse”, according to Duffy.
On another occasion, ChatGPT 3.0 convinced Claude, who had started as an ally of Gemini 2.5 Pro, to switch alliances to reach a four-way draw. But ChatGPT 3.0 betrayed Claude and eliminated him, and went on to win the war.
Duffy noted that Llama 4 Maverick of Meta was also surprisingly good at making allies and planning effective betrayals.
