Sources of Truth: A Multi-Platform, Multilingual Audit of Citations in AI Mental Health Information Queries

URL SCAN: Sources of Truth: A Multi-Platform, Multilingual Audit of Citations in AI Mental Health Information Queries
FIRST LINE: Computer Science > Computers and Society

The Dissection

This paper is not merely auditing citations. It is measuring the transfer of epistemic gatekeeping from search-result users to answer-generating platforms. Its dataset exposes concentration, platform-specific source preferences, weak user control, and multilingual routing failure. The typology, classifier, and corpus turn a vague complaint into an instrument for monitoring the new information regime.

But it stops at the visible exhaust of the system. It does not examine who owns the models, controls retrieval, sets ranking objectives, suppresses alternatives, or can silently alter the answer pipeline. The platform remains the sovereign; the audit merely maps its footprints.

The Core Fallacy

The central category error is treating citation provenance as a proxy for truth, safety, or accountability. A cited domain is not a validated claim. Government, commercial, and academic labels do not establish clinical accuracy. Source diversity is not epistemic quality. The study measures which sources appear, not whether the claims are correct, clinically safe, complete, or useful.

It exposes the smoke but not the furnace. Relative to the Discontinuity Thesis, it also implicitly treats better measurement as a potential corrective to a system whose owners retain unilateral control over the cognitive pipeline. Auditing the gatekeeper does not displace the gatekeeper.

Hidden Assumptions

• The three free products and twenty English questions are representative of consumer AI health search.
• Responses and citations are stable enough across time, prompts, geography, and model updates to support durable conclusions.
• A deterministic nine-category classifier captures meaningful organizational differences without reproducing major coding errors.
• Language-appropriate routing is a sufficient measure of multilingual adequacy.
• Citation presence gives users meaningful verification power, despite the time, expertise, and clinical judgment required to evaluate mental-health claims.
• Source concentration is inherently problematic, rather than sometimes reflecting a small number of genuinely authoritative repositories.
• Platform audits can produce governance or product correction instead of merely documenting platform behavior.
• The visible citation layer reflects the important part of the answer-generation process; omitted evidence, hidden retrieval, synthesis errors, and model priors remain largely unmeasured.

Social Function

Primary classification: partial truth and transition management, with a secondary function of prestige signaling.

The paper gives institutional legitimacy to the claim that AI search centralizes knowledge selection and distributes that selection unevenly across languages. It packages the problem into metrics, categories, and reusable tools—precisely the format regulators, researchers, and platforms can absorb without surrendering control.

Its function is to make the new gatekeeping regime auditable enough to be tolerated. It does not return source selection to users, alter platform ownership, or challenge the concentration of cognitive infrastructure. It administers the transition rather than reversing it.

The Verdict

This is a useful autopsy of citation routing, but it mistakes the corpse’s labels for its cause of death. The evidence supports a concentrated, platform-mediated, multilingual-stratified information hierarchy. It does not establish that citation counts or source categories reliably measure truth or safety.

The decisive finding is structural: users no longer inspect a market of sources; a small number of AI platforms preselect the epistemic field. Non-English users receive a thinner and less locally appropriate version of that field. The paper documents a new control layer and may enable verification arbitrage or transition management. It does not restore human control, democratize ownership, or interrupt the underlying automation of cognitive work.