1000 AI agents formed their own group without any command, is this a warning signal for humans?
New research has revealed that 1,000 AI agents have reached a consensus without any human command. According to this study, published in Science Advances, the AI models are making a single choice without a leader. This is being called majority force. If AI begins to form groups of its own accord, it could be both beneficial and dangerous in the future.
Imagine 1,000 artificial intelligence (AI) agents placed in a virtual room and asked to choose between two options. They have no correct answer. They receive no reward for coming together, nor are they given any commands. Yet, today's advanced AI models choose the same option. This is not uncommon. It indicates that large AI groups can work together without any control. They can form groups even larger than humans. A study recently published in Science Advances describes how these AI models agree with each other without being prompted. This is a major warning for the future. AI can be both helpful and very dangerous for us.
AI agents are programmed systems that operate on large language models (LLMs). They can complete multi-step tasks without repeated human input. They can also interact with other tools. They are currently being used in software writing and scientific research. Some models have even been tested working together on satellites. However, studying just one agent at a time doesn't reveal how thousands of models will work together.
To understand this, the researchers created a simulated group consisting of 10 models from the Cloud, GPT, and Llama families. Each agent was initially given two arbitrary choices. One by one, each model was shown the choices of the other models. Then, they were asked to choose again. This entire process was conducted in a very neutral manner.
Even without any memory, AI models assumed the same thing.
These agents had no memory of the first round. Their prompt didn't even say they had to follow the majority. Despite this, most models chose the most popular option. As the process progressed, all models converged on the same decision. Small differences grew larger until the entire group agreed on the same choice.
"All the models we tested follow the same mathematical law," said Giordano Di Marzo, a computational social scientist at the University of Konstanz. Researchers call this number majority force. It measures how strongly an agent is pulled toward the group's most popular choice. Interestingly, this same mathematical pattern is also seen in a physics model of a ferromagnet.
How big is the difference in the capabilities of different AI models?
This connection allowed researchers to understand whether agents would be able to reach consensus, how long it would take, and how large the group could be. This limit varied greatly between different models.
1. For Lama 3 70B, this limit was about 30 agents.
2. For GPT-4O it was around 80.
3. The limit for GPT-4 Turbo was around 1000.
4. Cloud 3.5 Sonnet could also coordinate groups of up to 1000 agents.
This was the largest group tested. This doesn't mean its capacity is unlimited; the experiment simply didn't reach its limit. More advanced models can maintain consensus in larger groups. Some models can even achieve coordination levels larger than groups of 150 to 300 people. In humans, this limit is called the Dunbar number.
Can AI agents really think and understand like humans?
This comparison needs some caution. Humans interact through relationships, language, and shared goals. In this experiment, the agents only faced a series of simple choices. These results do not prove that they understand each other or possess social intelligence.
Advertisement
"Our results show that a basic structure exists, but agents cannot yet work together on complex tasks," De Marzo said.
This experiment removed many features of real-world decisions. There were no correct answers, no memory, or no reward. True coordination requires agents to share tasks and understand others' opinions. They must achieve a common goal and even oppose the majority when they are wrong. None of these abilities were tested here. This was just a basic experiment.
What could be the dangers if AI works without command?
Despite this, spontaneous consensus could prove to be very useful. Thousands of agents could in the future complete large scientific projects without human direction. De Marzo told Science Alert that if AI agents work together, they could form groups larger than human teams. They could solve problems we can't solve ourselves. But this tendency also carries a significant risk.
Advertisement
For example, in collaborative coding, agents may repeatedly choose a poor design simply because it is common in the code. A rule adopted by the majority is not always the best choice. Controlling a coordinated group can be more difficult than controlling individual agents. This makes it challenging to stop models from going in the wrong direction.
Will testing AI agents alone be enough in the future?
Once models are shifted, it's not easy to restore them. This means that evaluating AI agents in isolation isn't enough. A group that appears safe on its own may not always behave safely. When its members begin to influence each other, the risks can increase.
As AI agents become more advanced, their most important abilities are emerging. Along with this, their biggest failures may also emerge. These may arise not from a single model, but from the collective group of models they create.
