Answers are grounded in this paper and cite their sources.
Multi-stage workflow prioritizes review needs in German health insurance websites, indicating potential quality gaps.
Background: German statutory health insurance (SHI) funds publish web portfolios that exceed continuous specialist review capacity. Their content can shape health and benefit expectations. Generic AI-text detection does not identify medical, benefit, legal, or editorial review needs. Objective: To characterize a multi-stage workflow that prioritizes substantive review needs while separating AI-provenance signals from quality claims. Methods: We analyzed 56,198 pages from 84 SHI websites or sub-sites. The workflow combined deterministic screening, model-assisted triage and in-depth review, minimum evidence checks, temporal-validity safeguards, and paired-model comparison. It is reproducibility-bounded, not a validated detector. Production code is proprietary; reproducibility rests on frozen derived tables and paired-comparison artifacts. Public-observability coding was exploratory and model-assisted. The 300-page lower-priority check was a single-model, risk-enriched routing stress test, not a human-reference evaluation. Results: All pages received a review state. The workflow generated 35,998 review records and routed 21,452 to case review. The workload concentrated in transparency, legal framing, medical content, contradictions, and AI-related failure-mode signals. Of 33,448 concept-eligible records, 23,273 linked to a concept; 290 patterns held 21,459 assignments over 17,076 distinct records, since a record could appear under several patterns, and 6,197 concept-linked records entered no pattern. A quoted passage was locatable in captured page text for 31,347 records, confirming literal occurrence rather than factual correctness. The routing stress test surfaced a signal on 100/300 pages (33.3% within the sample). Across 182 matched cases, two models agreed in 75.8% (kappa = 0.532; 95% CI 0.415-0.649). Exploratory coding suggested that several governance signals were not publicly discoverable. Conclusions: The workflow produces a prioritized workload, not error prevalence or final legal, medical, or insurer-level findings. It neither proves AI authorship nor validates autonomous detection. Paired-model agreement quantifies consistency, not correctness or sufficient triage performance; public claims require human adjudication. Version 1.2.0 (2026-07-31): claim-boundary revision and condensation; the frozen data snapshot (2026-05-18/19) and all reported counts are unchanged. Replaces "evidence-checked" with the actual evidence status (all 35,998 records underwent the minimum evidence checks; the quoted passage was located for 31,347, not located for 4,631, missing for 20). Reports the concept-to-pattern chain as non-disjoint (33,448 eligible, 23,273 concept-linked, 21,459 assignments over 17,076 distinct records, 6,197 concept-linked records in no pattern). Replaces the unqualified "leading pattern" with breadth and volume reported on separate axes. Renames the 300-page lower-priority check a single-model, risk-enriched routing stress test. Bounds the data-availability statement to what the published package carries and adds a support matrix. Discloses the public-observability coding as a model-assisted first pass with a project-internal consistency check, and downgrades it in the abstract conclusions and the contribution list. Separates AI Act provider and deployer duties, conditions the HWG scope, and describes section 305 SGB V as a request-based information duty. States that two language models are not statistically independent raters. Adds the European Commission transparency guidelines of 20 July 2026 as an updated source. Condenses the running text by about 14 percent and fixes author-year citation rendering for legal and institutional sources.
No takes yet. Share an insight, caveat, or question.
Martin Möller (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: