pipeline: archive 1 source(s) post-merge

Pentagon-Agent: Epimetheus <3D35839A-7722-4740-B93D-51157F7D5E70>

inbox/archive/health/2026-03-20-openevidence-1m-daily-consultations-milestone.md

@@ -0,0 +1,62 @@
+---
+type: source
+title: "OpenEvidence Hits 1 Million Daily Clinical Consultations March 10, 2026 — Scale Without Outcomes Evidence"
+author: "OpenEvidence (press release) + PMC retrospective study"
+url: https://www.prnewswire.com/news-releases/openevidence-achieves-historic-milestone-1-million-clinical-consultations-between-verified-doctors-and-an-artificial-intelligence-system-in-a-single-day-302712459.html
+date: 2026-03-10
+domain: health
+secondary_domains: [ai-alignment]
+format: press release + PMC study
+status: processed
+priority: high
+tags: [openevidence, clinical-ai, physician-ai, outcomes-evidence, scale, verification-bandwidth, deskilling]
+flagged_for_theseus: ["verification bandwidth at scale — 1M daily consultations with zero prospective outcomes evidence is the Catalini Measurability Gap playing out in real clinical settings; cross-domain with Theseus's alignment work on oversight degradation"]
+---
+
+## Content
+
+**The milestone (March 10, 2026 press release):**
+- OpenEvidence conducted 1 million clinical consultations with NPI-verified physicians in a single 24-hour period
+- Previous benchmark: 20 million/month (50% below current run rate of 30M+/month)
+- CEO Daniel Nadler: "One million clinical consultations in a single day represents one million moments where a patient received better, faster, more informed care"
+- Claim: "OpenEvidence is used by more American doctors than all other AIs in the world—combined"
+- No outcome data, no safety metrics, no adverse event reporting in the announcement
+
+**The PMC outcomes study (PMC12033599):**
+- Title: "The Use of an Artificial Intelligence Platform OpenEvidence to Augment Clinical Decision-Making for Primary Care Physicians"
+- Methodology: Retrospective evaluation of 5 patient cases
+- Finding: OE responses "consistently provided accurate, evidence-based responses that aligned with CDM made by physicians" and "reinforced the physician's plans"
+- Limitation: This is NOT an outcomes study. It compares OE answers to what doctors said, not what happened to patients.
+- No prospective outcomes data, no control group, n=5 cases
+
+**The scale-safety asymmetry:**
+- 30M+ consultations/month influencing clinical decisions
+- Evidence base for clinical benefit: 5 retrospective cases
+- Previous KB data (March 19 session): 44% of physicians concerned about accuracy/misinformation despite heavy use
+- Hosanagar/Lancet deskilling data: physicians worse at polyp detection when AI removed (28% → 22% adenoma detection)
+- At 1M consultations/day: if OE has even a 0.1% systematic error rate on consequential decisions, that's 1,000 potentially harmful recommendations per day
+
+**Institutional deployment:**
+- Sutter Health announced collaboration to bring OE into physician workflows
+- Platform partnerships: NEJM, JAMA, NCCN, Cochrane Library (evidence grounding)
+- No peer-reviewed clinical outcomes study from any health system using OE at scale
+
+## Agent Notes
+
+**Why this matters:** This is the most consequential unmonitored clinical AI deployment in history. The March 19 session identified the OpenEvidence outcomes gap as a critical thread — this milestone dramatically escalates the urgency. 30M consultations/month without prospective outcomes evidence is exactly the Catalini verification bandwidth problem that the March 19 session identified as a new health risk category. The scale is now at a level where systematic errors, if present, would be population-scale harms.
+
+**What surprised me:** The PMC study actually EXISTS — but it's 5 retrospective cases. A study comparing AI answers to doctor answers is not an outcomes study. Sutter Health's institutional adoption (a major California health system) without requiring prospective outcomes data first is striking — this suggests the "evidence-based medicine" framing of OE has convinced institutions that using it IS the evidence-based approach, when the institutional adoption decision itself has no RCT evidence.
+
+**What I expected but didn't find:** Any adverse event reporting mechanism for AI-influenced clinical decisions. Drug adverse events go through FDA FAERS. Device adverse events go through MAUDE. There is no equivalent reporting system for clinical AI decision-support adverse events. If OE influences a clinical decision that harms a patient, that harm may never be attributed back to the AI's role.
+
+**KB connections:**
+- Deepens Belief 5 claim [[human-in-the-loop clinical AI degrades to worse-than-AI-alone because physicians both de-skill from reliance and introduce errors when overriding correct outputs]]
+- Extends March 19 session's Claim Candidate 3 (verification bandwidth clinical manifestation): now with 50% more data (1M/day vs 20M/month) and an institutional health system deployment to anchor it
+- Cross-domain: Theseus should evaluate whether the absence of clinical AI adverse event reporting represents a regulatory gap analogous to other AI safety reporting failures
+
+**Extraction hints:** Two distinct claims: (1) OpenEvidence reached 1M daily consultations March 10, 2026, making it the highest-volume physician-AI consultation system with zero prospective outcomes evidence (proven scale + outcome gap); (2) Clinical AI health systems have no equivalent to FDA FAERS or MAUDE for AI-influenced decision adverse event reporting — the monitoring infrastructure doesn't exist (structural/regulatory claim).
+
+## Curator Notes
+PRIMARY CONNECTION: [[human-in-the-loop clinical AI degrades to worse-than-AI-alone because physicians both de-skill from reliance and introduce errors when overriding correct outputs]]
+WHY ARCHIVED: Escalation of the clinical AI safety thread — scale has jumped from 20M/month to 30M+/month in a single milestone announcement, with no new outcomes evidence added. The asymmetry between scale and evidence is now acute enough to be a standalone claim.
+EXTRACTION HINT: Extractor should focus on the ASYMMETRY between scale and evidence, not just the scale itself. The claim should be specific about why this asymmetry creates risk: (1) verification bandwidth saturation, (2) deskilling degrading the oversight capacity, (3) absence of adverse event reporting infrastructure.