AINEWS
Epic Systems' Cosmos Curiosity model trains on a 320 million patient pool drawn from its hospital customers' records. Twenty hospital systems have already signed on as early adopters.
Every time a U.S. hospital patient pulls up a lab result or messages a doctor through MyChart, the portal they are using is owned by Epic Systems, a private Verona, WI company that already holds the most complete copy of the American medical record there is. This week Epic told the customers who feed that record that the company now intends to use it to train its own clinical AI.
At its annual User Group Meeting in Verona (Aug 18-20, 2026), Epic unveiled Cosmos Curiosity, a generative AI the company describes as trained on "over 100 billion patient medical events" drawn from Cosmos, the clinical research database Epic has built from records contributed by its hospital customers. Epic positions Curiosity as an in-house alternative to general-purpose chatbots like ChatGPT or Claude for clinical use, on the rationale that sensitive patient data stays inside each customer's environment.
Epic is the dominant U.S. hospital electronic health record vendor, and Cosmos is a database Epic's customers, not Epic itself, populated with their patients' records. Curiosity, then, is not a third party training on patient data. It is the same vendor that already holds the chart, training a model on the pooled chart, and selling the model back to the hospitals that supplied the data. Twenty hospital systems have signed on as early adopters, per Healthcare IT Today's UGM 2026 coverage.
The scale Epic cites is large enough to make the lock-in concrete. The company says Cosmos covers 320 million unique patients, 23 billion encounters, 99 billion lab results, 186 million surgeries, 170 million home care visits, 25 million cancer cases, 18 million rare disease patients, and 15 million mother-baby links. Epic also previewed a separate family of foundation models called the Cosmos Medical Event Transformer (CoMET), pre-trained on 300 million patient records and 16 billion medical events.
Those numbers are Epic's own. Neither the announcement nor the trade press coverage of UGM 2026 cites independent clinical validation of Curiosity's predictive performance. Epic describes the model as simulating outcomes, not as a diagnostic tool, and the company has not published benchmarks against existing readmission or stroke models. Anyone betting clinical decisions on the output should hold that gap in view.
Hospitals are also under pressure from their own clinical AI ambitions. The agent platform Epic previewed at the same meeting positions the vendor as the broker between the hospital's chart and any number of third-party models, which extends the data-gravity argument into a marketplace play. The customer who fed the data buys back the model that the data trained, and any outside model has to plug in through Epic. Fierce Healthcare and Becker's Hospital Review framed the agent platform as Epic's answer to a market where clinicians are already using ChatGPT and Claude informally, sometimes on patient data, with no vendor in the loop.
What does this change for a patient with a MyChart login? On the announcement as written, nothing immediate. Curiosity is being deployed to the hospitals and health systems that license Epic and contribute data to Cosmos, not rolled into MyChart as a consumer feature. Whether a given patient's chart is in the training pool depends on whether their hospital participates in Cosmos, and that is a contract between the hospital and Epic, not a consent form the patient signs. Healthcare IT Leaders' read of the announcement flags the same patient-side gap.
The next concrete data point is whether Epic publishes benchmarks, an audit mechanism, or a patient-facing consent pathway before Curiosity moves from the 20 early adopters to broader deployment. Epic also previewed a Germany launch at the same meeting, which would extend the same data-gravity argument to a second country and a second regulatory environment. For now, the model exists as a vendor announcement backed by vendor-supplied scale numbers, in a market where the vendor already holds the chart.