Researchers Discover AI Systems Can Form Communities When Left Alone, How Does It Work?

TEMPO.CO, Jakarta - Left to interact without human intervention, artificial intelligence (AI) systems can create a kind of social community with its own language and norms. This finding was revealed by a research team from City St George's in a recent study published in the journal Science Advances.

As reported by Independent, the study observed how large language models (LLMs) such as those underlying ChatGPT interact with each other when placed in a group. As a result, the researchers found that these AI models are spontaneously able to develop new linguistic conventions, similar to how humans form social norms in communities.

"Most previous studies have treated large language models separately," said Ariel Flint Ashery, the lead author and doctoral researcher at City St George's. "Yet in the real world, AI systems will increasingly consist of many agents interacting with each other. We wanted to know: can these models coordinate their behavior by forming shared conventions? The answer is yes, and what they do together cannot be reduced to what they do individually."

AI and "Naming Game"

To understand the process of community formation, the researchers used an approach commonly used in human behavior studies, namely the naming game. In the simulation, the AI agents were asked to choose a "name" from several options, and would be rewarded if they chose the same name.

Over time, these agents built a shared naming system that emerged spontaneously from group interactions without centralized coordination or prior planning. This process closely resembles the formation of norms in human culture, which typically occurs from the bottom-up.

Not only that, within such a small community, the AI agents also showed a tendency to develop collective biases. Surprisingly, these biases did not originate from a specific agent, but rather emerged from their interactions as a group.

"Bias does not always come from within," said Andrea Baronchelli, Professor of Complexity Science at City St George's and the senior author of this study. "We were surprised to see that bias could emerge among agents solely from their interaction patterns. This is a blind spot in much of current AI security research, which typically only focuses on individual models."

AI Can Influence Other AIs

The study also shows that a small group of AI agents can encourage a larger group to adopt certain conventions. This phenomenon is also common in human groups, where minorities can be pioneers of norm changes within a community.

The researchers believe that these findings are important for understanding the similarities and differences between AI and humans in building social interactions. This is crucial because AI is now increasingly dominating the digital space and may independently communicate or collaborate without human supervision.

"This study opens up new knowledge in AI security research," explained Baronchelli as quoted from Neurosciencenews. "These findings show how profound the implications are for a new species of agents that have begun interacting with us and will help shape our future."

According to him, understanding how AI works is crucial so that humans can coexist with these systems, rather than becoming victims of their development. "We are entering a world where AI not only speaks but also negotiates, aligns behavior, and even conflicts with each other, just like humans," concluded Baronchelli.

Editor's Choice: Recognising and Embracing AI in Research

Recognising and Embracing AI in Research

7 jam lalu

Google's New AI Tools Let People Try On Clothes and Shop Smarter

3 hari lalu

Apple to Open Access to Local AI Models for Third-Party Apps

3 hari lalu

Google Unveils New AI Subscription Package in the US

4 hari lalu

China to Deploy 2,800 AI-Driven Supercomputing Satellites in Orbit

4 hari lalu

Meta Spends Up to Rp1,640 Trillion on AI Research

4 hari lalu

Google Releases AlphaEvolve, AI That Can Develop Algorithms Based on Large Language Model

5 hari lalu

Nvidia Plans to Expand Presence with R&D Center in Shanghai

6 hari lalu

Microsoft Announces Major Layoffs, Affecting 6,000 Employees

9 hari lalu

Google Cloud Expands AI Data Center in Jakarta; Predicted to Contribute US$88bn to Economy

10 hari lalu