Three rules for starting with AI
The three rules
Clinicians already protect the chart, call consults with a clear question and recheck a surprising lab value. AI competence applies the same three habits to a new tool, in the order you use it.
Before you type
1. Keep the patient out
No names, record numbers, dates or places in a personal AI account. Personal accounts, free or paid, carry no business associate agreement.
While you type
2. Ask like a consult
Give a role, de-identified context, one clear question, the format you need, and ask the model to separate evidence from debate and flag what to verify.
After it answers
3. Verify before you use
Every answer is a draft. Check numbers, doses, citations and guideline statements against the source. The AI drafts, you verify, you sign.
Transcript
Scene by scene, all on-screen text
- 0:00 Title. Three rules for starting with AI: how clinicians become AI competent.
- 0:08 You already have the habits that matter. You protect a patient's chart. You call a consult with a clear question. You recheck a surprising lab value before you act. AI competence is the same three habits, applied to a new tool.
- 0:24 Three rules, in the order you use them. Before you type: keep the patient out. While you type: ask like a consult. After it answers: verify before you use.
- 0:32 Rule 1. Keep the patient out. No names, record numbers, dates or places in a personal AI account.
- 0:40 Whatever you paste leaves your hospital. It travels to the vendor's servers. What protects it there is a contract, not a setting.
- 0:52 No business associate agreement, no patient data. Under HIPAA, an outside company may handle protected health information for you only under a business associate agreement (BAA). A personal account, free or paid, has no BAA; turning off model training is a privacy setting, not a BAA. An institutional enterprise account can have one, and your privacy office decides which tools and features are covered.
- 1:04 De-identify before you paste (illustrative example, invented patient). Name, record number, exact date, hospital and town are struck out, leaving: 34-year-old, G2P1 at 39 weeks, BP 148/96, headache since morning. HIPAA's Safe Harbor method lists 18 identifiers, including names, record numbers, dates more specific than the year, and places smaller than a state.
- 1:20 When in doubt, leave it out. Which tools are cleared for patient data is your institution's decision, not the tool's marketing.
- 1:28 Rule 2. Ask like you are calling a consult. The model answers the question it is given, and nothing more.
- 1:36 A good consult request carries the question, the context (the relevant facts, de-identified) and what you need back (format, depth and urgency). A good prompt carries the same three things.
- 1:46 A short prompt leaves the model to guess. "Is induction at 39 weeks better?" Better for whom? Compared with what? Which outcome? The model fills those gaps with a general answer for no one in particular.
- 1:58 The same question, asked like a consult (illustrative prompt). Role: I am an obstetrician preparing to counsel a patient. Context: healthy nulliparous woman at 39 weeks, no medical indication for delivery. Question: compare induction with expectant management for cesarean delivery and perinatal outcomes. Format: one page, name the key trials, end with a plain-language summary. Uncertainty: separate established evidence from debate, and tell me what I should verify. That last line is the heart of the Clinical Evidence Prompt (CEP) taught in AIM4Med 101.
- 2:16 Why this works. A language model predicts the most plausible next words. Your context changes what is plausible (schematic, not measured probabilities).
- 2:26 Rule 3. Verify before you use. Every answer is a draft. You are the final reviewer.
- 2:34 A reference can look perfect and not exist (schematic; every element is a placeholder). The model builds citation-shaped text from patterns in its training. Unless it searched, nothing was looked up. A fluent, confident tone tells you nothing about accuracy. If it did search, each source is a lead: open it and read it.
- 2:50 What to check, every time. Always verify numbers, risks and rates; drug doses; citations, DOIs and PMIDs; statements about guidelines. Lower risk: rewording text you wrote and already trust; formatting and structure; plain-language rewrites, then confirm the meaning held.
- 3:04 The same reflex as a surprising lab value: recheck before you act. The AI drafts. You verify. You sign. Responsibility for anything that reaches a patient stays with the clinician.
- 3:16 Recap. Keep the patient out. Ask like a consult. Verify before you use.
- 3:28 One thing to try today. Pick a clinical question whose answer you already know well. Ask it with no patient identifiers, written like a consult. Check every number and every reference against the source. You will see where the tool helps you, and where it needs you.
- 3:42 End card. AIM4Med 101, an independent interactive course for clinicians, at aim4med.com/aim4med101.
Sources
- Business associate contracts. 45 CFR § 164.502(e) and § 164.504(e).
- De-identification of protected health information, Safe Harbor method. 45 CFR § 164.514(b)(2).
- Vendor BAA documentation for consumer and enterprise AI plans, checked October 2026. Coverage changes; confirm with your institution's privacy office.
The patient, prompt, citation and probability bars shown in the video are illustrative and labeled as such on screen.