Patient information and AI tools

What HIPAA protects, why the course rule is stricter, and what to do if you slip · about 15 minutes

NoteWhat you’ll learn

By the end of this page you will be able to:

  1. Say what makes health information protected under HIPAA, and why taking out the name is not enough.
  2. Say why the course rule (no real patient, ever) is stricter than the law, and why.
  3. Say what a business associate agreement is, and what to ask before you put anything into a tool.
  4. Say what to do the same day if you put patient information somewhere it should not go.

The case you want to ask about

It is 2 a.m. on night float. A patient came in with a presentation you have never seen. You want to ask a chatbot for a differential. You think: I will leave out the name. So you type “94-year-old retired rancher from Kremmling, admitted Tuesday with fever and a new murmur.”

That sentence has no name in it. It is still protected health information. The rest of this page explains why, and what to do instead.

What the law protects

HIPAA is the federal health privacy law. It applies to covered entities: health plans, and health care providers that send claims and other standard transactions electronically, which includes your hospital. It also applies to their workforce. The regulation defines workforce as employees, volunteers, and trainees whose work the hospital directs, paid or not (45 CFR 160.103). As a medical student on a rotation, you are part of it.

The law protects individually identifiable health information: information about a person’s health, care, or payment for care that identifies them, or where there is a reasonable basis to believe it could be used to identify them (45 CFR 160.103). That second part matters. A sentence can identify a person without a name.

Taking out the name is not enough

HIPAA offers a checklist, called Safe Harbor, for making information no longer identifiable (45 CFR 164.514(b)). It lists eighteen kinds of identifier that must all be removed, for the patient and for their relatives, employer, and household. In plain words:

  • names
  • any place smaller than a state: street, city, county, and most of the ZIP code
  • every part of a date except the year, for dates about the person: birth, admission, discharge, death
  • any age over 89 (these can only be grouped as “90 or older”)
  • phone and fax numbers, email addresses, web addresses, and IP addresses
  • Social Security, medical record, health plan, account, and licence numbers
  • vehicle and device identifiers and serial numbers
  • fingerprints, voice prints, and full-face photographs or similar images
  • any other unique identifying number, characteristic, or code

And one more condition: the hospital must not know that what is left could still identify the person. A rare disease in a small town, or a case that was on the evening news, can identify someone with every item above removed.

Now look at the 2 a.m. sentence again. “94” is an age over 89. “Kremmling” is a town. “Tuesday” is part of an admission date. Three identifiers, no name.

Why the course rule is stricter than the law

The rule in every chapter of this book is: no real patient, ever. Not a de-identified one. Not one you are sure nobody could recognise. There are two reasons.

First, under the regulation, the covered entity decides that information has been de-identified, by Safe Harbor or by a statistician’s formal review (45 CFR 164.514(b)). You are not in a position to make that call alone at 2 a.m.

Second, the habit matters more than the case. The exercises in this course use made-up patients, public cases, or published examples. If you practise with invented patients now, you will not have to decide under pressure later.

The law also has a minimum necessary rule: when information is used or shared, it should be limited to what the purpose needs (45 CFR 164.502(b)). Sharing with another clinician to treat the patient is an exception. A question to a chatbot is not treatment of that patient by another provider.

The contract that decides which tools are allowed

A hospital may share protected health information with an outside company only under a business associate agreement (BAA). That is a contract that makes the company legally responsible for protecting the data. A scribe or model your hospital licenses inside the EHR runs under one. The consumer chatbot on your phone does not.

At CU Anschutz, ChatGPT Edu and Microsoft Copilot are covered when you sign in with your university credentials. The same products through a personal account are not. Check which account you are signed in to before you type. Other tools, including Google’s, are covered only if the institution tells you so in writing.

A covered account is for patient care, under the institution’s policy. It does not change the course rule: course exercises use made-up patients, whatever account you use.

Before you put anything into a tool, ask two questions. Is there a BAA for this tool, in the account I am signed in to? Does the policy say I may put this into it? If nobody can answer in a sentence, the answer is no. The ethics and regulation chapter takes this further.

Your patient may do what you may not

HIPAA applies to covered entities and their workforce. A patient is neither. Your patient may paste their own discharge summary into any chatbot they like. That is their choice about their own information. The same summary, pasted by you, is a disclosure by the hospital’s workforce. The patient communication chapter covers how to talk with a patient who has done this.

If you slip

If you put patient information into a tool by mistake, act the same day.

  1. Stop. Do not add more.
  2. Tell your supervising resident or attending.
  3. Report it to your institution’s patient privacy office. If you do not know how to reach it, your supervisor or the course director will.

Deleting the conversation may not remove it from the company’s systems, so deleting is not the same as reporting. Reporting early is what the process is for. Nobody expects you to fix it alone.

Check yourself

Six short questions. They use made-up patients.

1. You copy these values from a real chart, with nothing else: “Hemoglobin 7.2, MCV 68, ferritin 4.” Under HIPAA, is this protected health information?

  • Yes, because it came from a chart
  • Probably not on its own, but the course rule still says no
  • No, and you may paste it anywhere

Three lab values with no identifier and no unusual detail probably do not identify anyone, so on their own they are likely not protected. But they came from a real patient, and you are not the person who decides they are de-identified. The course rule is no real patient, ever.

2. “Mr. G, 94, retired rancher from Kremmling, admitted Tuesday with a hip fracture.” How many Safe Harbor identifiers does this sentence contain, not counting “Mr. G”?

  • One
  • Two
  • Three or more

The age is over 89, Kremmling is a place smaller than a state, and “Tuesday” is part of an admission date. “Retired rancher” in a small town can also narrow it to one person. “Mr. G” is part of a name, which makes four.

3. A 7-year-old with a very rare genetic disease was on the local news last month. You remove every item on the Safe Harbor list and keep the diagnosis and the hospital. Is the information now safe to share?

  • Yes, every listed identifier is gone
  • No, the rare diagnosis plus the news coverage can still identify the child

Safe Harbor includes “any other unique identifying characteristic,” and it fails if the hospital knows the rest could still identify the person. A publicised rare disease is exactly that.

4. Your patient pastes her own discharge summary into a free chatbot and asks it to explain her medications. Has she broken HIPAA?

  • Yes, the summary is protected health information
  • No, HIPAA applies to the hospital and its workforce, not to her

She is not a covered entity. She may share her own information however she chooses. If you pasted the same summary, it would be a disclosure by the hospital’s workforce.

5. You realise you pasted a real patient’s note into a chatbot an hour ago. What is the first thing to do after you stop typing?

  • Delete the conversation and move on
  • Tell your supervising resident or attending, then report it to the patient privacy office
  • Wait to see whether anyone notices

Deleting may not remove it from the company’s systems, and it does not count as reporting. Tell your supervisor and report the same day. Reporting early is what the process is for.

6. On rounds, a resident asks you to summarise a long consult note with ChatGPT. You open the app on your phone. It is signed in to your personal account. Is that covered by the university’s business associate agreement?

  • Yes, ChatGPT is covered at CU Anschutz
  • No, only the university-login ChatGPT Edu account is covered

The agreement is with the university’s account, not with the product. The same app signed in to a personal account is a consumer tool. Switch to the university login, and follow the institution’s policy for clinical use.

Go deeper