How to De-Identify Client Notes Before Using AI (A Step-by-Step Guide)

The single most important skill for any therapist using AI is de-identification: removing protected health information from your input before it ever reaches the tool. Get this right and AI becomes a safe drafting assistant. Get it wrong and you have a privacy incident. This guide walks through a practical workflow you can use before every prompt.

Step 1: Know the 18 identifiers

Under HIPAA, protected health information includes 18 identifier types — names, dates (birth, session, discharge), contact details, record numbers, device identifiers, biometric data, photos, and anything else that could reasonably identify the person. The rule of thumb: if a detail points to a real individual, it stays out of the chatbot.

Keep a printed copy of the identifier list near your desk. The de-identification guide in a purpose-built prompt pack includes one as a reference for exactly this reason.

Step 2: Convert facts to roles and structures

Take your raw note bullets and rewrite them with roles instead of people. "Maria, 34, teacher, panic attacks since her divorce in March" becomes "adult client, occupational stress, panic symptoms following a major life transition." Preserve the clinical structure you need (symptom pattern, session focus, plan format) while dropping everything identifying.

This is a skill that improves fast with practice. After a dozen notes, the conversion takes under a minute — and a checklist-driven prompt library reminds you to do it every single time with a pre-prompt check.

Step 3: Prompt about format, not people

Frame your prompts around the document, not the client. Instead of "write a progress note about my client's trauma," use "draft a progress-note outline with Subjective, Objective, Assessment, Plan headings from these de-identified bullets, leaving bracketed fields for me to complete." The AI handles structure and phrasing; all clinical content and decisions remain yours.

The same principle applies to treatment plans, homework sheets, and letters: prompt for scaffolding and professional language, then fill in and verify every clinical fact yourself.

Step 4: Review, verify, then file

Every AI draft gets a full clinician review before it enters any record. Check for invented details (AI tools sometimes add plausible-sounding specifics you never provided), confirm the tone matches your records, and make sure nothing identifying slipped in on your side either.

Never paste AI output containing anything sensitive back and forth unnecessarily, and keep your practice's broader compliance posture in mind: your licensing board, your EHR's policies, and whether any tool you use has a proper business associate agreement if PHI would ever be involved. The default should always be that it is not.

How this book helps

ChatGPT Prompts for Therapists (available here) is built around this workflow: 100 documentation prompts designed to work with de-identified input, a de-identification guide with an 18-identifier reference, an ethics primer on appropriate AI use, and a pre-prompt checklist so safe practice becomes habit rather than memory.

For the human side of a heavy caseload, the Burnout Recovery Pathway offers an 8-week structured workbook on energy and boundaries. And for clients who struggle between appointments, the Anti-Anxiety Affirmation Card Deck gives them 60 short, plain-language cards plus grounding exercises.

Educational content only, not legal or compliance advice. Consult your board and a qualified advisor for your jurisdiction's requirements.

Related posts

• Should Therapists Use AI Prompts for Documentation?
• How a Prompt Library Cuts Therapy Admin Time Without Cutting Corners