AI in U.S. healthcare: use ChatGPT without exposing PHI

HIPAA-compliant ChatGPT use starts at the front desk, where staff paste a slice of the chart: a referral letter, a prior authorization appeal, a portal reply, each one headed by the patient’s identifiers.

Nobody pastes a whole chart into a chatbot, just the piece that needs work: the referral letter to cardiology the doctor wants by five, or a discharge summary rewritten in plain English for the family. On older intake forms, that piece still starts with the Social Security number. When a new patient emails a photo ID, a watermarked passport card copy shows what the copy was for.

Name, date of birth, phone, email, Social Security number, medical record number: each of those is on the list of eighteen identifiers in the Privacy Rule’s safe harbor at §164.514(b)(2), where the MRN is letter (H). Our guide to HIPAA and AI tools goes through the whole list. This page is about the few seconds before the paste.

The business associate agreement with your AI vendor settles what the vendor owes the practice. It says nothing about what a medical assistant types into a personal account open in the next tab at 4:45 on a Tuesday. That decision is made at the workstation, by whoever is sitting there, and the workstation is where Nonimo does its work.

Of the eighteen HIPAA identifiers, the ones in this referral (name, date of birth, SSN, phone and email) become placeholders on the workstation when your staff member presses one key, before ChatGPT sees the text.

The swap happens on that Mac or Windows workstation, and only then does the text go to the chat. The AI writes its reply around the placeholders, and Nonimo puts the real details back in. More on our security page.

Nothing is hidden from you: every redacted detail is highlighted on screen, and one click brings it back.

The example says how long the chest pain has lasted, not the day it started, and that is deliberate. The safe harbor counts every element of a date except the year, and a date of onset written into the narrative is still a date.

The Privacy Rule never says yes or no to whether ChatGPT is HIPAA compliant, because HIPAA attaches to the information, not to the tool that receives it.

Health information that meets §164.514(a) and (b) stops being individually identifiable, and the safe harbor route reaches that point by removing the eighteen identifiers and having no actual knowledge that the rest points back to the patient.

The safe harbor asks for removal and nothing more on paper. It is the other route, expert determination, that requires the expert to document “the methods and results of the analysis that justify such determination” under §164.514(b)(1)(ii). For a practice relying on the safe harbor, the duty is to take the identifiers out, and taking them out before the text reaches the AI is the part Nonimo handles.

Nonimo replaces the patient’s identifiers with placeholders and leaves the clinical detail, such as the symptoms and the question for the specialist, because that is what the AI is there to help with. Our guide on what to redact before you paste explains what removal changes.

Nonimo is the desktop app that sits between a HIPAA practice and ChatGPT, on Windows or on a Mac, and it installs from the buttons at the top of this page.

A practice can start at the front desk, where referral letters and portal replies are written, and add workstations from there.

With several workstations, each installs the app, or IT deploys the browser extension by policy, which redacts text automatically in the AI tools, without a keypress, before the text is sent. Either way the work happens on the workstation, before the text reaches any assistant at all.