OpenAI connected ChatGPT Health to Epic, reaching 325 million patient records

It does not remove the interesting risks, because reading a record is itself a consequential act when the reader then explains it.

A patient asking ChatGPT what their oncologist's note means is asking a model to interpret a document written for another clinician, in compressed professional language, with hedges and differentials that mean specific things to a specialist and something else to a frightened person. Getting that wrong is not a data integrity problem. It is a patient who acts on a misreading.

The track record on adjacent tasks is not reassuring. NHS England's watchdog documented AI clinical scribes producing wrong drug names, missing prescription advice and, in one case, a fabricated diagnosis — "null demyelination" — across 27 systems already in use. Those tools were transcribing consultations, a considerably easier task than interpreting a specialist's note for a lay reader.

The consent question the announcement does not answer

An Epic record is not written by the patient and does not contain only the patient's information. Family history, sexual health, mental health, substance use, safeguarding concerns and clinicians' candid impressions all sit in there.

So the disclosure event is worth examining. A patient enabling this is authorising a commercial model provider to process material that includes third parties' information and their own most sensitive data, and OpenAI has not published what happens to it: whether it is retained, whether it can be used for training under any tier, what the deletion path is, and what happens when a patient revokes access.

The EU designated ChatGPT a very large online platform under the Digital Services Act on 31 August, with 159.1 million monthly users and an obligation to give vetted researchers access to platform data. How that obligation interacts with health records flowing through the same service is not something either the Commission or OpenAI has addressed.

What it does well

The use case that justifies this is unglamorous and real: patients cannot read their own records. Test results arrive through portals in raw form, discharge summaries are written for other doctors, and specialists' letters assume knowledge the patient does not have.

The existing alternative is a web search, which is worse in every respect — no access to the actual result, no context, and a reliable path to the most frightening possible interpretation. A model that can see the actual lab value, the reference range and the clinician's own note is starting from a better position than any tool a patient has had.

That is a genuine improvement. It is also being deployed to 325 million people's records without published evidence about how often the interpretation is wrong, and neither OpenAI nor Epic has said what evaluation was done before launch.