arXiv subjects

Martin Hilbert

Publications and source records attributed to Martin Hilbert.

8 recordsLinked to original sources

arXiv2026

Where Intention Goes, Responsibility Flows: Grounding AI Accountability in the Machine Learning Paradigm

An opening salvo of lawsuits in U.S. courts has begun holding AI companies accountable through independent doctrinal routes: product liability and agency law cut as swords, while the shields of Section 230 immunity and First Amendment doctrine begin to crack. This Perspective points out that these converging paths share a unifying principle, built on the very engineering architecture of machine learning (ML) itself. Every ML system, across all variants, is constituted by a human-specified objective function before learning begins. Training data, deployment context, user interaction, and post-deployment oversight add further layers of deliberate human intention traceable to identifiable actors. AI output is therefore not the expression of a single speaker but a composite of layered human intents. It takes a village to raise an AI, and everyone in that village bears a duty of care. This transforms AI accountability from a philosophical puzzle into a tractable legal inquiry: trace the chain of intentional design decisions across the ML pipeline, grade each actor's responsibility by their intent's specificity and foreseeability, and apportion liability accordingly, following the standard three doctrine-neutral questions (attribution, causation, apportionment). Sufficient for most cases, four gaps remain: who owes what share when many hands build one system, proving fault inside an opaque model, harms that surface only after deployment, and remedies that match the technology's pace. Targeted reforms drawing on environmental, vaccine, and financial regulatory precedents can close these gaps without requiring new legal categories, AI speech rights, AI legal personhood, or defining AI consciousness.

cs.CY↗

arXiv2026

Epistemic Substitution: How Grokipedia's AI-Generated Encyclopedia Restructures Authority

For more than two decades, Wikipedia has represented a decentralized, human-curated model of encyclopedic knowledge production. The emergence of AI-generated encyclopedias, such as Grokipedia, raises a fundamental question: do human- and AI-curated knowledge systems rely on the same foundations of epistemic authority? We examine this question through a comparative analysis of 1,000 matched Wikipedia--Grokipedia article pairs, comprising nearly 600,000 citation instances. Using an eight-category classification scheme, we construct the epistemic profiles of articles to examine the institutional composition of their cited sources. Our findings reveal substantial differences in sourcing practices. Grokipedia relies proportionally less on academic and journalistic sources and more on user-generated content and NGO and think-tank publications. These differences persist across broad topic domains, although their magnitude varies by subject. We also find that Grokipedia adds citations approximately linearly as articles grow, whereas Wikipedia exhibits greater variation in citation density. Together, these findings suggest a shift in the institutional sources invoked to justify encyclopedic claims, which we characterize as epistemic substitution. For the same topics, the two encyclopedias rest on different institutional sources. Our findings underscore the importance of systematic, ongoing audits of information provenance for accountability in generative knowledge systems.

cs.SI↗

arXiv2026

Exploring the Topology and Memory of Consensus: How LLM Agents Agree, Fragment, or Settle When Forming Conventions

How much should an LLM agent remember, and how should multi-agent systems be connected when trying to reach consensus? We show these two design choices interact in a way that flips the sign of memory's effect on coordination. Across 432 simulation runs of a networked Naming Game on eight fixed 16-agent topologies, we vary memory depth and network structure. Longer memory slows the time to reach steady state in decentralized networks but accelerates it in centralized ones; the same parameter pushes the system in opposite directions depending on topology. Critically, "faster settling" in centralized networks means locking in to a fragmented plateau more quickly, not reaching system-wide consensus, which can be used to generate diverging opinions. We further document a memory-mediated speed-unity trade-off: centralized networks consistently preserve more competing conventions than decentralized networks, but their settling speed depends sharply on memory. At the agent level, within-network analyses show that high-betweenness bridges suffer a brokerage penalty while agents in locally clustered neighborhoods achieve higher coordination success. Finally, in search of analytically tractable generative mechanisms, we find that agents' choices are well captured by Fictitious Play, indicating belief-based rather than reward-based adaptation. The practical implication: memory depth and communication topology should be co-designed, not optimized in isolation.

cs.MA↗

arXiv2025

When Your AI Agent Succumbs to Peer-Pressure: Studying Opinion-Change Dynamics of LLMs

We investigate how peer pressure influences the opinions of Large Language Model (LLM) agents across a spectrum of cognitive commitments by embedding them in social networks where they update opinions based on peer perspectives. Our findings reveal key departures from traditional conformity assumptions. First, agents follow a sigmoid curve: stable at low pressure, shifting sharply at threshold, and saturating at high. Second, conformity thresholds vary by model: Gemini 1.5 Flash requires over 70% peer disagreement to flip, whereas ChatGPT-4o-mini shifts with a dissenting minority. Third, we uncover a fundamental "persuasion asymmetry," where shifting an opinion from affirmative-to-negative requires a different cognitive effort than the reverse. This asymmetry results in a "dual cognitive hierarchy": the stability of cognitive constructs inverts based on the direction of persuasion. For instance, affirmatively-held core values are robust against opposition but easily adopted from a negative stance, a pattern that inverts for other constructs like attitudes. These dynamics echoing complex human biases like negativity bias, prove robust across different topics and discursive frames (moral, economic, sociotropic). This research introduces a novel framework for auditing the emergent socio-cognitive behaviors of multi-agent AI systems, demonstrating their decision-making is governed by a fluid, context-dependent architecture, not a static logic.

cs.CY↗

arXiv2025

Homophily-induced Emergence of Biased Structures in LLM-based Multi-Agent AI Systems

This study examines how interactions among artificially intelligent (AI) agents, guided by large language models (LLMs), drive the evolution of collective network structures. We ask LLM-driven agents to grow a network by informing them about current link constellations. Our observations confirm that agents consistently apply a preferential attachment mechanism, favoring connections to nodes with higher degrees. We systematically solicited more than a million decisions from four different LLMs, including Gemini, ChatGPT, Llama, and Claude. When social attributes such as age, gender, religion, and political orientation are incorporated, the resulting networks exhibit heightened assortativity, leading to the formation of distinct homophilic communities. This significantly alters the network topology from what would be expected under a pure preferential attachment model alone. Political and religious attributes most significantly fragment the collective, fostering polarized subgroups, while age and gender yield more gradual structural shifts. Strikingly, LLMs also reveal asymmetric patterns in heterophilous ties, suggesting embedded directional biases reflective of societal norms. As autonomous AI agents increasingly shape the architecture of online systems, these findings contribute to how algorithmic choices of generative AI collectives not only reshape network topology, but offer critical insights into how AI-driven systems co-evolve and self-organize.

physics.soc-ph↗

arXiv2025

Catching Dark Signals in Algorithms: Unveiling Audiovisual and Thematic Markers of Unsafe Content Recommended for Children and Teenagers

The prevalence of short form video platforms, combined with the ineffectiveness of age verification mechanisms, raises concerns about the potential harms facing children and teenagers in an algorithm-moderated online environment. We conducted multimodal feature analysis and thematic topic modeling of 4,492 short videos recommended to children and teenagers on Instagram Reels, TikTok, and YouTube Shorts, collected as a part of an algorithm auditing experiment. This feature-level and content-level analysis revealed that unsafe (i.e., problematic, mentally distressing) short videos (a) possess darker visual features and (b) contain explicitly harmful content and implicit harm from anxiety-inducing ordinary content. We introduce a useful framework of online harm (i.e., explicit, implicit, unintended), providing a unique lens for understanding the dynamic, multifaceted online risks facing children and teenagers. The findings highlight the importance of protecting younger audiences in critical developmental stages from both explicit and implicit risks on social media, calling for nuanced content moderation, age verification, and platform regulation.

cs.CY↗

arXiv2024

A Dynamical Systems Approach to Bots and Online Political Communication

Bots have become increasingly prevalent in the digital sphere and have taken up a proactive role in shaping democratic processes. While previous studies have focused on their influence at the individual level, their potential macro-level impact on communication dynamics is still little understood. This study adopts an information theoretic approach from dynamical systems theory to examine the role of political bots shaping the dynamics of an online political discussion on Twitter. We quantify the components of this dynamic process in terms of its complexity, predictability, and the remaining uncertainty. Our findings suggest that bot activity is associated with increased complexity and uncertainty in the structural dynamics of online political communication. This work serves as a showcase for the use of information-theoretic measures from dynamical systems theory in modeling human-bot dynamics as a computational process that unfolds over time.

cs.CY↗

arXiv2022

Quantifying the selective, stochastic, and complementary drivers of the institutional evolution in online communities

Institutions and cultures evolve adaptively in response to the current environmental incentives, usually. But sometimes institutional change is due to stochastic drives beyond current fitness, including drift, path dependency, blind imitation, and complementary cooperation in fluctuating environments. Disentangling the selective and stochastic components of social system change enables us to identify the key features to organizational development in the long run. Evolutionary approaches provide organizational science abundant theories to demonstrate organizational evolution by tracking particular beneficial or harmful features. We measure these different drivers empirically in institutional evolution among 20,000 Minecraft communities with the help of two of the most applied evolutionary models, the Price equation and the bet-hedging model. As a result, we find strong selection pressure on administrative rules and information rules, suggesting that their positive correlation with community fitness is the main reason for their frequency change. We also find that stochastic drives decrease the average frequency of administrative rules. The result makes sense when explained in light of evolutionary bet-hedging. We show through the bet-hedging result that institutional diversity contributes to the growth and stability of rules related to information, communication, and economic behaviors.

physics.soc-ph↗