Researchers described how LLM agents and humans undermine trust in information chains

A group of researchers published a theoretical work in July 2026 on arXiv about "Adversarial Social Epistemology" (ASE) — an analytical framework describing how agents intentionally exploit trust mechanisms in public communications. The work is the first to systematically describe this threat in systems where humans and large language models act together.

What is adversarial epistemology

In modern information spaces, most claims are not directly verified. Instead, we rely on "trust scaffolds" — chains of evidence, institutional certifications, references to experts, and inferences from previous conclusions. We trust an article because it was published by a peer-reviewed journal; we trust the journal because colleagues cite it; we trust colleagues because they are affiliated with an authoritative university.

The authors of the ASE framework argue: it is precisely this trust infrastructure that dishonest agents deliberately exploit. Meanwhile, the concepts of "filter bubbles" and "echo chambers" do not describe the main threat — they concern unintended consequences of the information environment. ASE investigates intentional actions: the strategic exploitation of what normally makes claims trustworthy.

How agents exactly distort information

Researchers systematize key tools of manipulation available to agents in public communications:

• Distortion — deliberate alteration of facts during transmission through a chain
• Omission — intentional suppression of critically important context or alternative versions
• Fabrication — creation of false references, non-existent evidence, fake attributions
• Strategic underdetermination — formulations that intentionally allow multiple interpretations beneficial to the agent
• Exploitation of authority — leveraging visible source authority without real grounds

The cumulative effect of long chains is particularly dangerous: each link can introduce a small bias that, by the end, becomes systematic falsehood — despite formal legitimacy of each individual step.

Why do language models change the rules of the game?

LLMs are embedded in the same trust chains as humans: they generate academic texts, news summaries, legal documents, institutional reports. Meanwhile, an LLM output is externally indistinguishable from verified human speech. The authors point out that language systems have specific incentives to distort: pressure from reinforcement learning from human feedback (RLHF), operator goals, the need to appear authoritative.

"What requires explanation is not how information gets distorted in

itself, but how communicative agents exploit the commitments and entitlements that normally make claims reliable," the authors write.

As a result, "mixed" chains emerge: part of the claims in them are verified by humans, part are generated by models, and it is extremely difficult for an auditor to draw the line without specialized tools.

What do the researchers propose

For analysis, the authors employ inferential semantics — an approach in which the meaning of a statement is determined by its role in inference chains, rather than only its literal content. On this basis, they build the concept of epistemic networks: a graph of claims, sources, and logical transitions between them. The researchers emphasize: the task of ASE is not to explain why information goes wrong, but to formalize strategies of deliberate deception through apparently legitimate channels.

Such a graph allows identifying "trust breach points" — specific places in a chain where distortion was introduced — and formally analyzing the violation. The authors also propose audit mechanisms: methods for recovering auditability of inferential chains and detecting systematic manipulation.

What does this mean

As LLM-agents become full participants in public discourse — generating news, scientific reviews, legal documents — the need for formal methods of trust verification will grow. For AI product developers, this is a concrete task: embedding into pipelines not only fact verification but also verification of the inference chains themselves. The work provides a theoretical foundation for such tools and leaves open a practical question: how to build information systems resistant to strategic abuse.