arXiv subjects

Paul Bouchaud

Publications and source records attributed to Paul Bouchaud.

5 recordsLinked to original sources

arXiv2026

Recommender system in X inadvertently profiles ideological positions of users

Studies on recommendations in social media have mainly analyzed the quality of recommended items (e.g., their diversity or biases) and the impact of recommendation policies (e.g., in comparison with purely chronological policies). We use a data donation program, collecting more than 2.5 million friend recommendations made to 682 volunteers on X over a year, to study instead how real-world recommenders learn, represent and process political and social attributes of users inside the so-called black boxes of AI systems. Using publicly available knowledge on the architecture of the recommender, we inferred the positions of recommended users in its embedding space. Leveraging ideology scaling calibrated with political survey data, we analyzed the political position of users in our study (N=26,509 among volunteers and recommended contacts) among several attributes, including age and gender. Our results show that the platform's recommender system produces a spatial ordering of users that is highly correlated with their Left-Right positions (Pearson rho=0.887, p-value < 0.0001), and that cannot be explained by socio-demographic attributes. These results open new possibilities for studying the interaction between human and AI systems. They also raise important questions linked to the legal definition of algorithmic profiling in data privacy regulation by blurring the line between active and passive profiling. We explore new constrained recommendation methods enabled by our results, limiting the political information in the recommender as a potential tool for privacy compliance capable of preserving recommendation relevance.

cs.SI↗

arXiv2025

Linear socio-demographic representations emerge in Large Language Models from indirect cues

We investigate how LLMs encode sociodemographic attributes of human conversational partners inferred from indirect cues such as names and occupations. We show that LLMs develop linear representations of user demographics within activation space, wherein stereotypically associated attributes are encoded along interpretable geometric directions. We first probe residual streams across layers of four open transformer-based LLMs (Magistral 24B, Qwen3 14B, GPT-OSS 20B, OLMo2-1B) prompted with explicit demographic disclosure. We show that the same probes predict demographics from implicit cues: names activate census-aligned gender and race representations, while occupations trigger representations correlated with real-world workforce statistics. These linear representations allow us to explain demographic inferences implicitly formed by LLMs during conversation. We demonstrate that these implicit demographic representations actively shape downstream behavior, such as career recommendations. Our study further highlights that models that pass bias benchmark tests may still harbor and leverage implicit biases, with implications for fairness when applied at scale.

cs.AI↗

arXiv2025

Web Crawler Restrictions, AI Training Datasets \& Political Biases

Large language models rely on web-scraped text for training; concurrently, content creators are increasingly blocking AI crawlers to retain control over their data. We analyze crawler restrictions across the top one million most-visited websites since 2023 and examine their potential downstream effects on training data composition. Our analysis reveals growing restrictions, with blocking patterns varying by website popularity and content type. A quarter of the top thousand websites restrict AI crawlers, decreasing to one-tenth across the broader top million. Content type matters significantly: 34.2% of news outlets disallow OpenAI's GPTBot, rising to 55% for outlets with high factual reporting. Additionally, outlets with neutral political positions impose the strongest restrictions (58%), whereas hyperpartisan websites and those with low factual reporting impose fewer restrictions -only 4.1% of right-leaning outlets block access to OpenAI. Our findings suggest that heterogeneous blocking patterns may skew training datasets toward low-quality or polarized content, potentially affecting the capabilities of models served by prominent AI-as-a-Service providers.

cs.SI↗

arXiv2025

Community Notes undermoderate polarizing content by design creating risks in electoral processes

Community Notes (CNs) of X enables users to collaboratively moderate misleading content. To resolve conflicting moderation, CNs infers a latent ideological dimension and selects notes garnering cross-partisan support. As this system is now deployed worldwide, we evaluate its operation across diverse polarization contexts. We analyze all 1.9 million moderation notes receiving 135 million ratings by March 2025, cross-referencing ideological scaling data on 13 countries. Our results show that the CNs algorithm effectively captures the main polarizing dimensions across countries, surfacing notes that garner cross-partisan support. This also means that, by design, CNs systematically under-moderate polarizing content. We analyze notes relating to four recent elections in the US (2024), the UK (2024), France (2024) and Germany (2025) and demonstrate that they are systematically under-moderated when compared to other notes, posing potential risks to civic discourse and electoral processes.

cs.SI↗

arXiv2023

Can Few Lines of Code Change Society ? Beyond fack-checking and moderation : how recommender systems toxifies social networking sites

As the last few years have seen an increase in online hostility and polarization both, we need to move beyond the fack-checking reflex or the praise for better moderation on social networking sites (SNS) and investigate their impact on social structures and social cohesion. In particular, the role of recommender systems deployed at large scale by digital platforms such as Facebook or Twitter has been overlooked. This paper draws on the literature on cognitive science, digital media, and opinion dynamics to propose a faithful replica of the entanglement between recommender systems, opinion dynamics and users' cognitive biais on SNSs like Twitter that is calibrated over a large scale longitudinal database of tweets from political activists. This model makes it possible to compare the consequences of various recommendation algorithms on the social fabric and to quantify their interaction with some major cognitive bias. In particular, we demonstrate that the recommender systems that seek to solely maximize users' engagement necessarily lead to an overexposure of users to negative content (up to 300\% for some of them), a phenomenon called algorithmic negativity bias, to a polarization of the opinion landscape, and to a concentration of social power in the hands of the most toxic users. The latter are more than twice as numerous in the top 1\% of the most influential users than in the overall population. Overall, our findings highlight the urgency to identify harmful implementations of recommender systems to individuals and society in order better regulate their deployment on systemic SNSs.

cs.DM↗