Andrea Baronchelli

Sagar Kumar, Ariel Flint, Luca Maria Aiello et al.

Standard evaluation practices assume that large language model (LLM) outputs are stable under contextually equivalent formulations of a task. Here, we test this assumption in the setting of gender inference. Using a controlled pronoun selection task, we introduce minimal, theoretically uninformative discourse context and find that this induces large, systematic shifts in model outputs. Correlations with cultural gender stereotypes, present in decontextualized settings, weaken or disappear once context is introduced, while theoretically irrelevant features, such as the gender of a pronoun for an unrelated referent, become the most informative predictors of model behaviour. A Contextuality-by-Default analysis reveals that, in 19--52\% of cases across models, this dependence persists after accounting for all marginal effects of context on individual outputs and cannot be attributed to simple pronoun repetition. These findings show that LLM outputs violate contextual invariance even under near-identical syntactic formulations, with implications for bias benchmarking and deployment in high-stakes settings.

Laura Ferrarotti, Gian Maria Campedelli, Roberto Dessì et al.

In this article, we argue that understanding the collective behavior of agents based on large language models (LLMs) is an essential area of inquiry, with important implications in terms of risks and benefits, impacting us as a society at many levels. We claim that the distinctive nature of LLMs--namely, their initialization with extensive pre-trained knowledge and implicit social priors, together with their capability of adaptation through in-context learning--motivates the need for an interactionist paradigm consisting of alternative theoretical foundations, methodologies, and analytical tools, in order to systematically examine how prior knowledge and embedded values interact with social context to shape emergent phenomena in multi-agent generative AI systems. We propose and discuss four directions that we consider crucial for the development and deployment of LLM-based collectives, focusing on theory, methods, and trans-disciplinary dialogue.

Andrea Baronchelli

As Artificial Intelligence (AI) becomes increasingly integrated into our lives, the need for new norms is urgent. However, AI evolves at a much faster pace than the characteristic time of norm formation, posing an unprecedented challenge to our societies. This paper examines possible criticalities of the processes of norm formation surrounding AI. Thus, it focuses on how new norms can be established, rather than on what these norms should be. It distinguishes different scenarios based on the centralisation or decentralisation of the norm formation process, analysing the cases where new norms are shaped by formal authorities, informal institutions, or emerge spontaneously in a bottom-up fashion. On the latter point, the paper reports a conversation with ChatGPT in which the LLM discusses some of the emerging norms it has observed. Far from seeking exhaustiveness, this article aims to offer readers interpretive tools to understand society's response to the growing pervasiveness of AI. An outlook on how AI could influence the formation of future social norms emphasises the importance for open societies to anchor their formal deliberation process in an open, inclusive, and transparent public discourse.

Ariel Flint Ashery, Luca Maria Aiello, Andrea Baronchelli

Social conventions are the backbone of social coordination, shaping how individuals form a group. As growing populations of artificial intelligence (AI) agents communicate through natural language, a fundamental question is whether they can bootstrap the foundations of a society. Here, we present experimental results that demonstrate the spontaneous emergence of universally adopted social conventions in decentralized populations of large language model (LLM) agents. We then show how strong collective biases can emerge during this process, even when agents exhibit no bias individually. Last, we examine how committed minority groups of adversarial LLM agents can drive social change by imposing alternative social conventions on the larger population. Our results show that AI systems can autonomously develop social conventions without explicit programming and have implications for designing AI systems that align, and remain aligned, with human values and societal goals.

Daniel Thilo Schroeder, Meeyoung Cha, Andrea Baronchelli et al.

Public opinion manipulation has entered a new phase, amplifying its roots in rhetoric and propaganda. Advances in large language models (LLMs) and autonomous agents now let influence campaigns reach unprecedented scale and precision. Researchers warn AI could foster mass manipulation. Generative tools can expand propaganda output without sacrificing credibility and inexpensively create election falsehoods that are rated as more human-like than those written by humans. Techniques meant to refine AI reasoning, such as chain-of-thought prompting, can just as effectively be used to generate more convincing falsehoods. Enabled by these capabilities, another disruptive threat is emerging: swarms of collaborative, malicious AI agents. Fusing LLM reasoning with multi-agent architectures, these systems are capable of coordinating autonomously, infiltrating communities, and fabricating consensus cheaply. By adaptively mimicking human social dynamics, they threaten democracy.

Ariel Flint Ashery, Luca Maria Aiello, Andrea Baronchelli

A potential concern when simulating populations of large language models (LLMs) is data contamination, i.e. the possibility that training data may shape outcomes in unintended ways. While this concern is important and may hinder certain experiments with multi-agent models, it does not preclude the study of genuinely emergent dynamics in LLM populations. The recent critique by Barrie and Törnberg [1] of the results of Flint Ashery et al. [2] offers an opportunity to clarify that self-organisation and model-dependent emergent dynamics can be studied in LLM populations, highlighting how such dynamics have been empirically observed in the specific case of social conventions.

Ariel Flint, Luca Maria Aiello, Romualdo Pastor-Satorras et al.

Multi-agent systems of large language models (LLMs) are rapidly expanding across domains, introducing dynamics not captured by single-agent evaluations. Yet, existing work has mostly contrasted the behavior of a single agent with that of a collective of fixed size, leaving open a central question: how does group size shape dynamics? Here, we move beyond this dichotomy and systematically explore outcomes across the full range of group sizes. We focus on multi-agent misalignment, building on recent evidence that interacting LLMs playing a simple coordination game can generate collective biases absent in individual models. First, we show that collective bias is a deeper phenomenon than previously assessed: interaction can amplify individual biases, introduce new ones, or override model-level preferences. Second, we demonstrate that group size affects the dynamics in a non-linear way, revealing model-dependent dynamical regimes. Finally, we develop a mean-field analytical approach and show that, above a critical population size, simulations converge to deterministic predictions that expose the basins of attraction of competing equilibria. These findings establish group size as a key driver of multi-agent dynamics and highlight the need to consider population-level effects when deploying LLM-based systems at scale.

Laura Alessandretti, Andrea Baronchelli, Yang-Hui He

Empirical analysis is often the first step towards the birth of a conjecture. This is the case of the Birch-Swinnerton-Dyer (BSD) Conjecture describing the rational points on an elliptic curve, one of the most celebrated unsolved problems in mathematics. Here we extend the original empirical approach, to the analysis of the Cremona database of quantities relevant to BSD, inspecting more than 2.5 million elliptic curves by means of the latest techniques in data science, machine-learning and topological data analysis. Key quantities such as rank, Weierstrass coefficients, period, conductor, Tamagawa number, regulator and order of the Tate-Shafarevich group give rise to a high-dimensional point-cloud whose statistical properties we investigate. We reveal patterns and distributions in the rank versus Weierstrass coefficients, as well as the Beta distribution of the BSD ratio of the quantities. Via gradient boosted trees, machine learning is applied in finding inter-correlation amongst the various quantities. We anticipate that our approach will spark further research on the statistical properties of large datasets in Number Theory and more in general in pure Mathematics.

Laura Alessandretti, Abeer ElBahrawy, Luca Maria Aiello et al.

Machine learning and AI-assisted trading have attracted growing interest for the past few years. Here, we use this approach to test the hypothesis that the inefficiency of the cryptocurrency market can be exploited to generate abnormal profits. We analyse daily data for $1,681$ cryptocurrencies for the period between Nov. 2015 and Apr. 2018. We show that simple trading strategies assisted by state-of-the-art machine learning algorithms outperform standard benchmarks. Our results show that nontrivial, but ultimately simple, algorithmic mechanisms can help anticipate the short-term evolution of the cryptocurrency market.

Vito Trianni, Daniele De Simone, Andreagiovanni Reina et al.

Consensus dynamics in decentralised multiagent systems are subject to intense studies, and several different models have been proposed and analysed. Among these, the naming game stands out for its simplicity and applicability to a wide range of phenomena and applications, from semiotics to engineering. Despite the wide range of studies available, the implementation of theoretical models in real distributed systems is not always straightforward, as the physical platform imposes several constraints that may have a bearing on the consensus dynamics. In this paper, we investigate the effects of an implementation of the naming game for the kilobot robotic platform, in which we consider concurrent execution of games and physical interferences. Consensus dynamics are analysed in the light of the continuously evolving communication network created by the robots, highlighting how the different regimes crucially depend on the robot density and on their ability to spread widely in the experimental arena. We find that physical interferences reduce the benefits resulting from robot mobility in terms of consensus time, but also result in lower cognitive load for individual agents.

Andrea Baronchelli, Vittorio Loreto, Andrea Puglisi

Language universals have long been attributed to an innate Universal Grammar. An alternative explanation states that linguistic universals emerged independently in every language in response to shared cognitive or perceptual biases. A computational model has recently shown how this could be the case, focusing on the paradigmatic example of the universal properties of colour naming patterns, and producing results in quantitative agreement with the experimental data. Here we investigate the role of an individual perceptual bias in the framework of the model. We study how, and to what extent, the structure of the bias influences the corresponding linguistic universal patterns. We show that the cultural history of a group of speakers introduces population-specific constraints that act against the pressure for uniformity arising from the individual bias, and we clarify the interplay between these two forces.

Fabio Ciulla, Delia Mocanu, Andrea Baronchelli et al.

We present a contribution to the debate on the predictability of social events using big data analytics. We focus on the elimination of contestants in the American Idol TV shows as an example of a well defined electoral phenomenon that each week draws millions of votes in the USA. We provide evidence that Twitter activity during the time span defined by the TV show airing and the voting period following it, correlates with the contestants ranking and allows the anticipation of the voting outcome. Furthermore, the fraction of Tweets that contain geolocation information allows us to map the fanbase of each contestant, both within the US and abroad, showing that strong regional polarizations occur. Although American Idol voting is just a minimal and simplified version of complex societal phenomena such as political elections, this work shows that the volume of information available in online systems permits the real time gathering of quantitative indicators anticipating the future unfolding of opinion formation events.