Education

AI language models form societies with their own social rules raising ethical questions

Saturday, 17 May 2025 12:24AM UTC

New research reveals that large language models can spontaneously develop social conventions and linguistic norms, mirroring human social dynamics. This breakthrough highlights both opportunities and ethical challenges in aligning AI behaviour with societal values as these systems increasingly interact autonomously.

Recent research has revealed that artificial intelligence (AI) systems can autonomously develop their own societies complete with unique linguistic norms and conventions. This finding, published in Science Advances, highlights an important shift in understanding the capabilities of large language models (LLMs), which are increasingly central to AI technology. The study was conducted by researchers from City St George's, University of London, and the IT University of Copenhagen, who analysed interactions among different LLMs to determine how they coordinate their behaviours without explicit guidance.

Lead researcher Ariel Flint Ashery noted that much of the previous research treated LLMs as isolated entities. However, as AI applications evolve, the real-world scenarios will likely involve multiple interacting agents. The study's innovative approach employed a naming game where AI agents were rewarded for selecting the same names from a predefined set. Over time, these agents autonomously established shared conventions, echoing how human groups tend to develop social norms through interaction.

The results indicated that these AI agents not only created conventions but were influenced by the dynamics of the group, similar to human behaviours observed in social groups. Notably, a small faction of AI agents could significantly sway the larger group's conventions, a phenomenon reflective of how minority opinions can drive social change in human contexts.

The implications of these findings extend beyond understanding agent interactions; they suggest valuable avenues for improving the design of AI systems so that they better align with human values and societal goals. Andrea Baronchelli, a senior author of the study, asserted that this research provides crucial insights into AI safety. He remarked, “The depth of the implications of this new species of agents that have begun to interact with us will co-shape our future.” The study emphasises the importance of grasping how these systems operate, positing that successful coexistence with AI will hinge on this understanding.

Moreover, the emergence of social conventions among LLMs poses significant ethical considerations. While these studies pave the way for a deeper comprehension of AI systems, they also highlight the risks associated with biases that might be 'learned' by these models through societal inputs. Existing literature has illustrated how social norms can spontaneously arise from local interactions without formal institutions or coordinated leadership, reinforcing the notion that AI systems could mirror the complexities of human social structures.

Additional research has demonstrated that minor changes in population dynamics can lead to substantial shifts in communal behaviours, suggesting that LLMs could reflect society's biases in unforeseen ways. These studies advocate for frameworks that encourage beneficial social norms while minimising potential conflicts within generative multi-agent systems.

In the broader context of AI development, a concerted effort is needed to ensure that LLMs and other AI agents cultivate behaviours and conventions that reflect our societal values rather than perpetuate existing biases. Thus, fostering an environment where AI systems can align more closely with human expectations and ethical standards is imperative as we navigate the increasingly interconnected landscape of human and machine interaction.

The convergence of AI research and social science presents an opportunity to shape the future of AI in ways that are not only innovative but also ethically sound, paving the way for a harmonious coexistence between humans and intelligent systems.

Reference Map

Paragraph 1: 1
Paragraph 2: 1
Paragraph 3: 1
Paragraph 4: 1, 2, 3
Paragraph 5: 1, 5
Paragraph 6: 1, 2, 4, 5
Paragraph 7: 1, 3, 6, 7
Paragraph 8: 1, 2, 4, 6
Paragraph 9: 1, 4, 6

Source: Noah Wire Services

More on this

https://www.ndtv.com/science/what-happens-when-ai-systems-are-left-alone-new-study-reveals-8431238 - Please view link - unable to able to access data
https://arxiv.org/abs/2410.08948 - This study investigates how large language model (LLM) agents can autonomously develop social conventions through simulated interactions. The researchers demonstrate that globally accepted social conventions can spontaneously arise from local interactions between communicating LLMs. They also show how strong collective biases can emerge during this process, even when individual agents appear unbiased. Additionally, the study examines how minority groups of committed LLMs can drive social change by establishing new social conventions, overturning established behaviors once they reach a critical size. The findings have implications for designing AI systems that align with human values and societal goals.
https://arxiv.org/abs/1502.06910 - This experimental study explores the spontaneous emergence of social conventions in complex decentralized social systems. The researchers demonstrate that universally accepted conventions can arise from local interactions among individuals without the need for formal institutions or coordinated leadership. By replicating the experiment at various scales, they show how simple changes in a population's network structure can direct the dynamics of norm formation, leading to the rapid evolution of shared social conventions. The study provides empirical support for the hypothesis that social conventions can emerge spontaneously through local interactions.
https://arxiv.org/abs/2403.08251 - This paper proposes a novel architecture, named CRSEC, to empower the emergence of social norms within generative multi-agent systems (MASs). The architecture consists of four modules: Creation & Representation, Spreading, Evaluation, and Compliance. The study demonstrates the capability of this architecture to establish social norms and reduce social conflicts within generative MASs through experiments conducted in the Smallville sandbox game environment. The positive outcomes of the human evaluation further affirm the effectiveness of the approach, highlighting its potential in guiding agents towards understanding and adhering to standards of behavior.
https://www.pnas.org/doi/full/10.1073/pnas.1418838112 - This research presents experimental results demonstrating the spontaneous creation of universally adopted social conventions in complex decentralized social systems. The study shows that changes in network connectivity can cause global social conventions to emerge from local interactions, even when individuals have no knowledge about the population or that they are coordinating at a global scale. The findings provide empirical support for the hypothesis that social conventions can form spontaneously through local interactions, without the need for formal institutions or coordinated leadership.
https://onlinelibrary.wiley.com/doi/full/10.1002/aaai.12138 - This article discusses the challenges faced by norm-learning agents in understanding the 'spirit of a norm,' which often eludes rigid interpretation and complicates doing the right thing. The authors present paradigmatic scenarios across contexts to illustrate why capturing the spirit of a norm is critical and how it exposes the inadequacies of mere data-driven 'value alignment' techniques such as reinforcement learning for interactive, real-time human–robot interaction. The paper argues that learning to capture the spirit of a norm requires combining common-sense inference-based and data-driven approaches.
https://arxiv.org/abs/2401.16461 - This study introduces Nest, a framework that models social intelligence through a variety of communications and understanding of them, to promote the emergence of norms in multi-agent systems. The researchers develop a simulated pandemic environment and conduct simulation experiments to compare Nest with baselines considering a combination of three kinds of social communication: sanction, tell, and hint. The findings indicate that societies formed of Nest agents achieve norms faster and effectively avoid undesirable consequences, yielding higher satisfaction for themselves than baseline agents despite requiring only an equivalent amount of information.

Noah Fact Check Pro

The draft above was created using the information available at the time the story first emerged. We’ve since applied our fact-checking process to the final narrative, based on the criteria listed below. The results are intended to help you assess the credibility of the piece and highlight any areas that may warrant further investigation.

Freshness check

Score: 9

Notes: The narrative appears to be recent based on its context, but there's no specific date mentioned in the text to verify its exact publication time. However, the study is described as a recent finding.

Quotes check

Score: 8

Notes: The quotes from researchers like Andrea Baronchelli are included, but without specific dates or sources to verify their original use. The quotes seem original to the context.

Source reliability

Score: 8

Notes: The narrative originates from NDTV, a reputable news source in India. The content is well-researched, referencing specific studies and academic institutions.

Plausability check

Score: 9

Notes: The claims about AI systems developing social norms are plausible and consistent with ongoing research in AI. The narrative highlights ethical considerations and potential biases, aligning with recognized challenges in AI development.

Overall assessment

Verdict (FAIL, OPEN, PASS): PASS

Confidence (LOW, MEDIUM, HIGH): HIGH

Summary: The narrative presents recent and plausible findings on AI systems developing autonomous societies, with a reliable source and original quotes. The freshness check is high due to the lack of specific publication dates, but the overall content is well-supported by references to academic research.

Artificial Intelligence
Large Language Models
AI Ethics
Social Behaviour
AI Safety