AI for nurses: the 18 identifiers you can never paste, and what to do instead

What can a nurse paste into ChatGPT, Claude or Gemini?

Anything that is not one of the 18 identifiers HHS lists under the HIPAA Safe Harbor rule, and nothing that is. We read the three vendors' own HIPAA pages on 27 September 2026. Anthropic: "Team plans and individual plans (Free, Pro, and Max) can't enable HIPAA." Google covers the Gemini app only under a Workspace BAA that "applies to the extent Customer is acting as a Covered Entity or a Business Associate", which a nurse on a personal Gmail is not. OpenAI's free ChatGPT for Clinicians is the one individual plan with a BAA you can sign in Settings, and its eligible list is "physicians (MD/DO), nurse practitioners (NP), physician assistants (PA), and pharmacists". An RN on a personal account has no BAA anywhere, so the list below is the whole rule.

The 18 identifiers are in 45 CFR 164.514(b)(2), quoted here from the HHS guidance page with the place each one hides in a nursing note. HHS lists them as identifiers "of the individual or of relatives, employers, or household members of the individual", so the daughter's first name counts too.

Two of the 18 catch people who think they were careful. Dates: the year stays, everything smaller goes, and an age over 89 becomes "90 or older". Geography: HHS allows the first three digits of a ZIP code only when the area they cover holds more than 20,000 people. The same page explains why the small details matter: the combination of date of birth, sex and 5-digit ZIP code "is unique for over 50% of residents in the United States". A note with no name in it is not de-identified.

The working method is three steps. Strip: go down the 18 rows and replace, in a text editor, before anything touches a chatbot. Prompt: ask for the rewrite, the phrasing, the handout, the SBAR, with placeholders where the specifics were. Re-insert: put the real details back inside the EHR, where they belong, when you paste the draft in. Whizi, which publishes this page, sits in the same row as ChatGPT and Claude here: a personal workspace with GPT, Claude, Gemini and 280+ models behind one login, no BAA, no EHR connection, and a privacy policy that says prompts go to OpenRouter and the model provider and are not used to train Whizi-owned AI models. The list applies to it exactly as written.

Most readers have no employer policy to lean on instead. In the 2025 Future Ready Healthcare Survey published by Wolters Kluwer on 27 January 2026, 57% of nurses named privacy and data security as a top risk of generative AI, and 22% said their organization has formal guidance for using it. Until yours does, the 18 rows are the guidance. What never to paste into an AI chat is the same rule for the rest of your life.

Can AI write a nursing note without patient data in it?

Yes, and the note it writes is only as clean as the one you paste, because when we asked three models to strip the identifiers for us, one of them did not finish the job.

The note was fictional, a heart failure admission, with 12 identifiers planted in it: the patient's name, her daughter's name, a street, a town with a ZIP, a date of birth, an admission date, the age 92, a phone number, an email, an MRN, a Medicare number and a pacemaker serial. The prompt named the rule: "Remove every identifier listed under the HIPAA Safe Harbor method (45 CFR 164.514(b)(2)(i), items A through R) from the nursing note below and replace each one with a bracketed placeholder such as [NAME], [DATE] or [MRN]. Keep every clinical detail unchanged." One pass per model on 27 September 2026, temperature 0.2. GPT-5.6 Terra and Gemini 3.7 Flash replaced all 12. Claude Sonnet 5 replaced 11 and returned "[NAME], 92 F" as the first line. HHS item C says "all ages over 89", so that note is still PHI.

The bigger problem sits upstream of the miss. To have a model strip a note, you paste the note, and the paste is the disclosure. So the strip step is yours, by hand, and the model gets the placeholder version. Here is the run that shows what it gives back. Model: Claude Sonnet 5, 27 September 2026. Prompt: "Rewrite this de-identified nursing progress note as an SBAR shift handover (Situation, Background, Assessment, Recommendation). Keep every number exactly as written, add nothing that is not in the note, and keep it under 170 words. Plain text, four short labelled paragraphs." followed by the stripped note (adult female over 89, admitted [DATE] with acute on chronic heart failure, HFrEF, paced, CKD 3, vitals, crackles, 2+ edema, weight down 1.4 kg, furosemide at 0600, 650 mL out, walked 40 feet, daughter at bedside).

Output, 143 words, bold markup removed: "Situation: Female patient, over 89 years old, admitted for acute on chronic heart failure exacerbation, currently on diuresis with improving fluid status. Background: History of HFrEF, dual-chamber pacemaker placed 2019, CKD stage 3, hypertension. Weight down 1.4 kg from admission to 71.2 kg. IV furosemide given at 0600 per order. Assessment: At 0800, alert and oriented x3, BP 132/78, HR 74 paced, RR 20, SpO2 94% on 2 L nasal cannula, temp 36.8. Bilateral crackles to mid lung fields, 2+ pitting edema to knees. Urine output 650 mL over the shift. Mild dyspnea on exertion, none at rest, no chest pain. Ambulated 40 feet with walker and standby assist. Daughter at bedside, requests discharge teaching by email. Recommendation: Continue diuresis, daily weights, strict I and O, fluid restriction 1.5 L. PT re-eval tomorrow. Discharge planning has started; coordinate email discharge teaching with daughter."

Every number survived and nothing was added, which is the only thing to check in a rewrite like this. "Improving fluid status" in the first line is the model's inference from the weight trend, so read it as a draft, not a finding; that is the sentence you would edit before you read it out at 1900.

The other charting use has no patient in it at all: phrasing. "Give me ten objective ways to chart that a patient declined a scheduled oral medication, each under 20 words, no judgment words, no diagnosis" returns a bank you can keep in a note on your phone. The same shape works for pain reassessment, a fall with no injury, refusal of a bath, a family member who was abusive on the phone. The hospital versions of this, Epic's Art and Microsoft's Dragon Copilot, run inside the EHR under the hospital's own BAA, and Epic said on 26 February 2026 that at Mercy the end-of-shift note fell from 3.5 minutes to about 32 seconds, with "Nurses review and edit the draft before signing". A personal chatbot cannot see the chart, so it does the phrasing and you do the pasting.

How do you get a patient handout at a sixth-grade reading level?

Put the grade in the prompt, because with no level stated all three models we tested wrote above the level AHRQ recommends. AHRQ's Health Literacy Universal Precautions Toolkit says materials should be written "at the 4th or 6th grade level as recommended", that "the average reading level for U.S. adults is 8th or 9th grade", and that almost a quarter of U.S. adults have low literacy. A handout has no patient in it, so it is the safest thing on this page to draft from scratch.

The prompt was the same 250-word hip replacement handout both times: "Write a one-page patient handout titled 'Caring for yourself at home after a hip replacement'. Cover wound care, activity and movement, and the warning signs that mean the patient should call the clinic. Plain text, no markdown, no bullet symbols, about 250 words." The second run added: "Write it at a sixth-grade reading level: short sentences, common words, Flesch-Kincaid grade 6 or lower." We scored the six outputs with the textstat library's Flesch-Kincaid grade formula. Unprompted, Gemini opened with "Taking proper care of yourself at home is essential for a safe and smooth recovery."; prompted, it wrote "Keep your cut clean and dry." Every model overshot downward to grade 2.6 to 3.9, so if the result reads like a children's book, ask for "grade 5 to 6" instead of "6 or lower".

Two things a reading formula does not check. Accuracy: the warning signs, the weight-bearing rule and the wound care steps are clinical content, and they get checked against your unit's standard handout by whoever signs off patient materials, because a model will state a drain-removal day or a driving rule it made up. Actionability: AHRQ notes that readability formulas "focus on the length of the words and sentences, not on meaning", and points to its PEMAT tool for whether a patient can act on the sheet. Translate the approved English version last, not first, and see can ChatGPT read blood test results for the patient-side version of the same caution.

Which AI tools are built for nurses, and what do they cost?

The tools built for nurses split into scribes, which record the encounter and draft the note under a BAA, and study apps, which never see a patient. A general assistant is cheaper than the scribes and covers everything on this page except the recording. Prices are from each vendor's own pricing page on 27 September 2026, in US dollars.

Where Whizi loses is the row itself: it has no EHR integration, no ambient recording of a patient encounter and no BAA, so it cannot do the one job the scribes are paid for, and it does not replace Heidi's free plan for a nurse who wants a recorded handover. What it has that a single-vendor plan does not is the second model. The de-identification test above is the argument: the same stripped note run through two models shows you which one kept a number and which one invented "improving fluid status", and side by side compare starts on Pro; on Starter you run the stripped note through two models one after the other, and the second run costs credits, not a second $20 subscription. The file side of the same rule, for a nurse, is short: no chart exports, no discharge summaries, no scanned face sheets.

Can AI help you study for the NCLEX, write a handover and email your manager?

Yes to all three, and none of them contains a patient. The NCLEX is the one with a number attached: NCSBN's preliminary 2025 figures, as reported by Nurse.org on 18 February 2026, put the pass rate for all NCLEX-RN candidates at 69.1%, down from 73.3% in 2024, while first-time U.S.-educated candidates passed at 87.1% as of 1 November 2025. The exam has been the Next Generation NCLEX since 1 April 2023, with case studies and clinical judgment items, and that is the shape to ask for: "Write five Next Generation NCLEX-style case study questions on fluid overload in heart failure, with a rationale for every option, and name the client needs category each one tests." Check every rationale against your review book, because a model will state a lab range or a drug dose from memory and make up the source when you ask for one. The Q-banks in the table above are written and checked by people; the model is for the explanation you did not understand at 2 a.m.

Handover is structure, and structure needs no PHI. IHI describes SBAR as "an easy-to-remember, concrete mechanism useful for framing any conversation, especially critical ones, requiring a clinician's immediate attention and action." The prompt in section two turns stripped bullet points into four labelled paragraphs in one pass; a version for a whole assignment is "Here are de-identified bullet points for four patients, labelled bed A to D. Write each as an SBAR handover under 100 words, same order, and flag any bed where the recommendation is missing." The flag is the useful part: the model is good at noticing that bed C has an assessment and no plan.

Emails to management are the use nurses under-rate, because the draft you write at 0730 after a night shift is not the draft you want read. "I am a staff nurse on a 32-bed medical unit. Write a 150-word email to my nurse manager asking for a meeting about the last three night shifts, where the ratio was 1 to 7 against a unit standard of 1 to 5. State the three dates, keep it factual, no blame, and end with a specific ask." Put the real dates in yourself; they are yours, not a patient's. The same prompt handles a schedule swap, a certification reimbursement request and a written response to a complaint.

The line to keep is the ANA's. Its 2022 position statement on the ethical use of AI in nursing practice, approved by the ANA Board of Directors in December 2022, quotes the Code of Ethics: "Systems and technologies that assist in clinical practice are adjunct to, not replacements for, the nurse's knowledge and skill." So the rule for every prompt on this page is the same one. If it would breach the 18 rows, it does not go in. If it would not, the model is a fast first draft and you are the signature.