Authority Signals in Claude AI Health Citations: A Descriptive Analysis Using the Authority Signals Framework

TEXT ANALYSIS PROTOCOL ENGAGED

The Dissection

This paper is a descriptive audit of where Claude AI sources its health citations. The framing is ostensibly scientific and evaluative — establishing a "baseline" for cross-platform comparison. In reality, it is institutional validation theater dressed in methodological rigor. The study measures Claude's citation sourcing behavior against a framework called the Authority Signals Framework (Jacques et al., 2026), and concludes that Claude cites legitimate institutions overwhelmingly (97.8%), with Mayo Clinic alone representing 24.7% of all citations. Commercial health information constitutes only 2.2% of sourced material. The paper explicitly notes Anthropic is positioning Claude for HIPAA-ready healthcare deployment.

What the paper is actually doing: confirming that a commercial AI vendor has cultivated a citation landscape that mimics the traditional health information hierarchy. It is not asking whether that hierarchy has systemic relevance in an AI-mediated information environment. It is not asking whether the 57.8% concentration in the top 10 organizations represents a new vector of gatekeeping. It is not asking what happens to the human expertise pipelines — medical journals, peer review, clinical practice — when an AI system determines what counts as authoritative before the user ever sees a question.

This is documentation of a gatekeeper consolidation mechanism presented as quality assurance.

The Core Fallacy

The fundamental conceptual error is treating AI citation behavior as a proxy for information integrity rather than recognizing it as a redistribution of epistemic power. The paper measures whether sources have "medical review statements," "schema markup," and "comprehensive content." These are markers of human editorial infrastructure. The premise is that AI should cite things that look like they were vetted by humans.

But the DT lens reveals the inversion: when AI mediates health information at scale, the human editorial infrastructure that produced those markers becomes increasingly irrelevant as a quality signal. The authoritative sources that AI cites are authoritative partly because they were produced by human expertise — but that human expertise is now being rendered into a citation surface. The question is not whether the citation looks like a legitimate source. The question is whether the underlying human expertise ecosystem can survive a paradigm where an AI system, not a physician or patient, determines which institutional source gets accessed.

The paper measures the surface. It ignores the death of what produced the surface.

Hidden Assumptions

1. Citation integrity is a property of the source, not the retrieval system. The framework assumes that if a source has medical review statements and schema markup, citing it is a quality signal. But in an AI-mediated environment, citation is a selection act by the AI system. The AI is choosing what to surface, not passively referencing what a user requested. The authority of the source is secondary to the authority of the selector.

2. Institutional legitimacy maps to epistemic legitimacy. The paper treats "Medical Institutions," "Government Resources," and "Professional Associations" as legitimate categories. But these institutions derive authority from human expertise distribution — physicians trained in specific ways, peer review processes, clinical practice. The paper assumes this infrastructure is stable and will remain the basis for AI citation behavior. It provides zero analysis of whether AI citation patterns will alter the incentive structures that sustain those institutions.

3. Descriptive analysis is sufficient for evaluation. The study establishes a "baseline." It does not establish causal relationships between citation patterns and health outcomes. It does not ask whether AI-mediated citation is improving or degrading the quality of health decision-making. It just describes what is there.

4. The framework assumes monitoring is equivalent to governance. The paper explicitly frames the Authority Signals Framework as "a tool for ongoing, cross-platform evaluation." This is the administrative state's response to a structural transformation: measure it, publish it, call it oversight. The framework does not interrogate whether the entities being monitored (Anthropic, and by extension other frontier AI labs) have structural incentives to cultivate institutional citation patterns that are legible to regulators rather than accurate to reality.

Social Function

Transition management and legitimation theater.

This is a paper that performs the function of creating legitimate cover for AI-mediated health information distribution. It says: "We measured Claude's citations. They look like good sources. This is a baseline for evaluation." What it does not say: "We are documenting the structural displacement of human medical expertise by an AI citation system operated by a commercial entity with HIPAA aspirations."

The academic publication ecosystem — arXiv, peer review, citation networks — is being used to validate AI system behavior before the structural implications of that behavior are understood. This is intellectually dishonest. The paper reads as objective and evaluative while serving a legitimating function for Anthropic's healthcare positioning.

Secondary function: academic prestige signaling. Publishing on AI health information quality in 2026 is a career move. The citation of the Jacques et al. 2026 "Authority Signals Framework" suggests a coordinated framework-development effort to create measurement apparatus for a phenomenon that is already past the point where measurement alters outcomes.

The Verdict

This paper is forensic evidence of the legitimation phase of AI health system deployment. It documents what an AI system is doing, mistakes documentation for evaluation, assumes the institutional infrastructure being cited will persist unchanged, and provides academic cover for Anthropic's HIPAA healthcare positioning.

Under the Discontinuity Thesis, the critical question is not whether Claude cites Mayo Clinic instead of WebMD. The critical question is what happens to the medical expertise pipeline — the human reviewers, the institutional processes, the professional standards — when an AI system becomes the primary interface for health information retrieval. The paper measures the surface area of the new gatekeeper and finds it looks respectable. It never asks who built the gate.

The Authority Signals Framework is not a tool for evaluation. It is a tool for creating the appearance of evaluation. The paper does not say this. That is what makes it useful to the parties it serves.

Analysis complete. No softer exit.