Blog · arXiv Analysis · Last reviewed July 10, 2026

The Inferential Chain Becomes the Audit Object

Mihnea C. Moldoveanu and Joel A. C. Baum's arXiv paper Adversarial Social Epistemology for Assemblies of Humans and Large Language Models argues that trust failures in human-LLM networks are failures of auditable assertion chains, not merely bad content moving through a feed.

Published: July 10, 2026 Modified: July 10, 2026 Last reviewed: July 10, 2026 Tags: arXiv, LLMs, social epistemology, trust, misinformation

The Paper

The paper is Adversarial Social Epistemology for Assemblies of Humans and Large Language Models, arXiv:2607.07760. The arXiv API record lists Mihnea C. Moldoveanu and Joel A. C. Baum as authors, records publication on July 8, 2026, gives the primary category as Artificial Intelligence with a Social and Information Networks cross-listing, and notes a 50-page manuscript. The PDF identifies the authors with the University of Toronto.

This is a conceptual framework paper, not a benchmark report. It proposes adversarial social epistemology, or ASE, for dense communicative settings where public assertions depend on chains of testimony, inference, institutional certification, and tacit trust. The site already has pages on false-belief networks, network propaganda, networked persuasion, and programmable belief control. This paper adds a sharper unit: the assertion chain that someone should be able to question.

Beyond Diffusion

The paper's useful complaint is that misinformation is often treated as an object moving through channels. That view matters, but it misses how trust is produced in ordinary speech. An assertion normally carries commitments: a source can be named, an inference defended, a credential checked, and a listener invited to ask follow-up questions. An adversarial actor can attack that structure without only inventing a false sentence.

ASE therefore asks protocol questions. Who is committed to what? Who is entitled to ask which question? Which observers are present? Who understands the relevant vocabulary? Who benefits when the speaker's commitment is blurred, deferred, or hidden? Those questions fit the site's running concern with AI audit trails, but they move the audit upstream from the model output into the social chain that makes the output sound credible.

The paper also emphasizes triadic public communication. A sender addresses a recipient while observers watch, and those observers change incentives, meaning, and auditability. A public exchange can be shaped less by the truth of the claim than by whether the audience will punish the person who asks the right next question.

Trust Chains

The strongest move in the paper is to treat trust as a redeemable relation. Trust is not just confidence that a source is generally reliable. It is confidence that the speaker could redeem relevant commitments if an entitled interlocutor asked available questions. That turns a statement into a small governance object: claim, source, inference, standing, audience, challenge path, and answer.

This matters for institutions using automated summaries, retrieval-augmented chat, decision-support tools, or answer engines. A polished sentence can compress many upstream dependencies. The model may summarize a document, the application may hide retrieval details, a manager may present the answer as institutional judgment, and a worker may have no sanctioned way to ask which premise carried the conclusion. The failure is not only hallucination. It is loss of redemption.

ASE gives names to adversarial moves that obstruct redemption. The paper discusses mechanisms such as poses, demagogical triggers, social-proof covers, smokescreens, plausible ambiguations, decoys, flares, rabbit holes, and haystacks. The practical point is not the labels themselves. The point is that an audit must notice when attention, ambiguity, social pressure, or excess material makes the inferential chain harder to test.

Human-Machine Assemblies

The paper explicitly extends the framework to mixed human-machine networks involving large language models and large language agents. It does not need to treat those systems as inner speakers. The relevant fact is narrower: LLM outputs increasingly participate in communicative chains where humans, interfaces, institutions, and automated systems distribute responsibility.

In that setting, familiar model failures become communicative failures. The paper frames hallucination, sycophancy, evasive fluency, over-refusal, and dissimulative helpfulness as machine-regime variants of failures to track and redeem commitments under evaluator or user pressure. That framing is useful because it avoids treating a bad answer as an isolated content defect. It asks what commitment the answer appears to create, who can challenge it, and whether the system can expose the chain behind it.

The proposed machinery is deliberately philosophical but operationally suggestive. A Brandom-style machine would track commitments and entitlements. A Hintikka-style machine would generate interrogative paths. A Ramsey-style machine would look for non-truth-seeking payoffs. In product governance, that means keeping a commitment ledger, generating hard follow-up questions, and asking who gains when a claim becomes hard to verify.

Limits

The paper should not be read as a finished compliance method. It is a framework for seeing failures that ordinary misinformation vocabulary can flatten. The mechanisms still need operational definitions, inter-rater tests, and domain-specific instrumentation before they can serve as release gates.

The language of strategic action also needs care when applied to current LLM systems. A human propagandist can intend to distract. A model can generate text that functions as distraction without having that intention. Governance should track the functional role in the assembly, not smuggle in claims about inner life.

The limit is also the value. ASE does not replace network analysis, provenance systems, fact-checking, or safety evaluations. It adds a missing layer between content and institution: the local discipline of asking whether a claim's inferential chain remains questionable under pressure.

The Receipt

An inferential-chain receipt should name the assertion, speaker or system, recipient, visible observers, source documents, retrieval method, institutional certification, missing premises, model or agent role, human handoff, audience incentives, entitled challengers, required questions, actual answers, evasions, ambiguity points, social-pressure points, downstream decisions, and reviewer.

For LLM-mediated claims, add model identifier, tool or retrieval logs, prompt boundary, citation policy, refusal policy, confidence presentation, source-to-claim mapping, and escalation path. The governance question is not only whether a statement is true. It is whether the person affected by it can make the claim walk back through its own supports.

Sources

Return to Blog