Two researchers have published a framework for understanding how humans and large language models manipulate information in shared communication networks.

Mihnea Moldoveanu from the University of Toronto and Joel Baum co-authored the 50-page paper on adversarial social epistemology, focusing on environments where public assertions rely on chains of testimony, inference, and institutional trust.

The framework addresses scenarios where communicative agents — both human and AI — have incentives to distort, omit, fabricate, or strategically under-specify information for personal, reputational, or material gains.

Beyond Echo Chambers

The researchers argue that existing concepts like epistemic bubbles, echo chambers, and misinformation diffusion fail to capture how agents exploit the very mechanisms that normally make information trustworthy.

Their adversarial social epistemology provides analytical tools for understanding how trust gets subverted in scaffolded public communications. The paper outlines specific mechanisms that undermine trust and proposes auditing machinery to address trust breaches.

The framework draws on epistemic networks enriched with inferentialist semantics for interpreting assertions. This approach focuses on how the commitments and entitlements that typically support reliable information exchange can be weaponized.

The research comes as AI systems increasingly participate in information ecosystems alongside humans, creating new vectors for strategic manipulation of shared knowledge.

The paper was submitted to arXiv on July 8, 2026, under the artificial intelligence and social information networks categories. The authors propose their framework as essential for understanding densely interactive communicative landscapes where AI and human agents coexist.

The work addresses growing concerns about information integrity as AI systems become more sophisticated at generating and disseminating content that appears authoritative but may serve strategic rather than epistemic purposes.