Weill Cornell AI Incident: Lessons for Research Teams

On November 5, Weill Cornell Medicine disclosed that patient health information had been entered into AI tools without authorization during a radiology research study. The use violated the institution’s own policies.

It is the latest in a run of similar cases at academic medical centers, and the common thread is where they happened: inside research.

What Happened at Weill Cornell

According to the notice and reporting by Becker’s Hospital Review, HIPAA Journal, and DataBreaches.net, the information was entered into AI tools between June 10 and June 20. Weill Cornell learned of it on September 6.

The data came from Weill Cornell Imaging at NewYork-Presbyterian and NewYork-Presbyterian Hospital. It included medical record numbers and clinical details. It did not include patient names, Social Security numbers, or financial information.

Weill Cornell says it has not identified any misuse, suspended the research activity, requested deletion of the data, and provided additional education on appropriate AI use. The number of people affected has not been disclosed.

It Is Not the Only Case

Becker’s reports that Loma Linda University Health disclosed in July that a file with limited patient information was uploaded on May 25 to an external AI platform during an IRB-approved research study. Loma Linda also requested deletion and is reviewing its policies and training.

The Johns Hopkins Bayview case we covered earlier involved a hospital employee rather than a research study, but the pattern is the same: a person, a file, and a tool nobody had cleared.

Two Lessons Worth Taking

Research projects are not exempt. An approved study can still involve an unapproved tool. IRB approval covers the research protocol, and it does not automatically cover every software tool a team member opens during the work.

“No names” does not mean “no identifiers.” Medical record numbers are among the 18 identifier types in HIPAA’s Safe Harbor method. Data without names can still point to a patient, especially when combined with clinical detail.

The Cleanup Problem

In both cases, the institutions asked for the data to be deleted. That is the standard response, and it has a limit: deleting your own copy does not confirm what the tool provider kept. DataBreaches.net asked Weill Cornell whether it had received confirmation that the data was deleted, and the notice does not say.

The better time to deal with retention is before upload.

What Research Teams Can Put in Place

1. Name the approved tools. List which AI and writing tools are cleared for the study, and say so in the data management plan.
2. Cover everyone. Students, fellows, and coordinators often work fastest and have the least training.
3. Minimize what goes in. Use the smallest amount of data a task needs, and remove identifiers, including record numbers, wherever possible.
4. Give people an easy approved route. If the cleared tool is slower, people will use another.
5. Agree the paperwork. Ask vendors for a BAA or DPA before real data is involved.
6. Plan for a slip. Know who to tell, and what to ask a vendor in writing about storage, training use, and deletion.

Where Trinka’s Confidential Data Plan Fits

Trinka’s Confidential Data Plan is designed as an approved route for the writing side of research. Content is processed in memory with zero data retention, is never used to train a model, and is encrypted in transit and at rest. For healthcare organizations, Trinka states that a Business Associate Agreement is available on request.

Confidential Data Plan does not replace IRB oversight, de-identification, or your institution’s policies, and it does not stop someone from using a different tool. What it offers is a cleared option that keeps no copy, so a mistaken paste into it leaves nothing to chase.

Conclusion

The Weill Cornell and Loma Linda cases show that AI slips happen inside approved research too, and that cleanup afterward is uncertain. The stronger approach is to settle the approved tools, the paperwork, and the rules before the study begins.