---
title: Clinical AI Reaches Patients Faster Than Anyone Tracks Its Mistakes
description: ECRI is expanding its safety-reporting network to track errors from AI tools in patient care, as most cleared AI medical devices reach hospitals untested on outcomes.
author: Darie Nani (Editor-in-Chief)
updated: 2026-08-25T21:23:35.008Z
canonical: https://www.sovereignmagazine.com/article/ecri-ai-errors-patient-care-reporting-network
image: https://cdn.nanimediahouse.com/ecri-ai-patient-safety-206482.webp
categories: Artificial Intelligence
content_type: News
region: United States
publication: Sovereign Magazine
schema_type: Article
---

AI now touches an enormous share of American medicine, and until this week almost no one was systematically counting when it got a patient's care wrong. ECRI, a national patient-safety nonprofit based in Fort Washington, Pennsylvania, said on August 25 that it is expanding its Problem Reporting Network to specifically capture and investigate errors, malfunctions and near misses involving AI tools and AI-enabled medical devices used in patient care.

The organization is asking healthcare providers, health systems and clinicians across the country to report any incident where an AI-enabled tool may have contributed to an error or introduced risk into how care is delivered. ECRI says it will triage and investigate each report and share its findings back with whoever submitted it. The network covers the range of places AI has reached in clinical work, from diagnostic imaging to clinical decision support to patient-facing chatbots.

ECRI says there is no centralized, healthcare-specific mechanism tracking how often clinical AI tools produce incorrect or misleading outputs, or how often those wrong outputs actually reach a patient. Surveys already show that a large majority of US hospitals and health systems use clinical AI in at least one workflow, so the software is making calls in live care while no one keeps a systematic record of when it fails.

> "We don't yet have a clear picture of AI's downstream impact in healthcare."
> — Scott Lucas, ECRI

## Only Three of 1,357 Cleared AI Devices Were Tested on Patients

Little of this software has been tested against what actually happens to patients. The FDA had authorized 1,357 AI/ML-enabled medical devices as of December 2025. Of those, only about 2.5% had prospective clinical trials registered, 0.9% posted results to ClinicalTrials.gov, and just 0.2%, three devices in total, were evaluated against real patient outcomes such as mortality or hospital readmissions. Most reached the market through the FDA's 510(k) pathway, which asks a manufacturer to show a device is substantially equivalent to one already on sale rather than to prove it helps patients.

That clearance route says little about how a tool behaves at the bedside. Epic's sepsis-prediction model, deployed across hundreds of US hospitals, was found in a Michigan Medicine study to miss roughly two-thirds of actual sepsis cases while firing alerts on a large share of patients who did not have sepsis. A separate study of recalled AI medical devices found many had no published clinical validation before authorization, and that a large share of recalls happened within the first year the device was on sale.

## ECRI Already Runs a National Safety-Reporting Network

It runs the ECRI and ISMP Patient Safety Organization, which has gathered and analyzed more than 8 million safety-event reports from healthcare providers nationwide. The expansion turns that same reporting network toward a kind of failure current oversight was never designed to catch.

Scott Lucas, ECRI's vice president of devices, therapeutics and technology, said the organization has long warned about adopting AI with too little scrutiny, and that without a solid dataset the industry cannot improve how these tools are designed and used. The aim, Lucas said, is to feed real-world failure data back to the hospitals choosing these systems, so they can keep the tools that work and drop the ones that quietly fail.

For health-system operators and the investors backing clinical AI, that dataset is the piece that has been missing: a shared, structured account of where these tools break in practice, held by an organization that has done the same for drugs and devices for decades.

## FAQ

**Q: Is reporting an AI-related incident to ECRI mandatory?**
No. ECRI's Problem Reporting Network is voluntary. It asks providers, health systems and clinicians to come forward, with no legal requirement to do so, which is part of why reliable data on how often clinical AI fails has been so hard to assemble.

**Q: How is this different from the FDA's existing device-reporting system?**
The FDA's MAUDE database collects adverse-event reports on medical devices in general, but it was not designed to capture AI-specific failures, such as a model that quietly degrades over time or performs worse for certain groups of patients. ECRI's network is meant to capture and investigate those healthcare-specific AI incidents.

**Q: Can ECRI force an unsafe AI tool off the market?**
No. ECRI is a patient-safety nonprofit, not a regulator. It collects and investigates reports and shares what it learns, but it cannot recall a device or pull a tool out of hospitals. That authority sits with the FDA and with the health systems that decide which tools to run.