AI for Doctors: Where It Helps, Where It's Dangerous

A growing number of physicians now finish their notes before they leave the room — not because they type faster, but because something else drafted the note while they talked. The same tool wrote the patient's discharge instructions in plain English, summarized a forty-page guideline into the three lines that mattered, and filled out the prior-authorization letter that used to eat a Friday afternoon. None of that is the future. It is happening in clinics right now, and the doctors using it are quietly getting their evenings back.

That is the comfortable half of the story. Here is the uncomfortable half: the exact same tool will, with total confidence, invent a drug interaction that doesn't exist, cite a study that was never published, and tell a patient something that could hurt them — and it will sound just as authoritative doing that as it does when it's right. In most jobs, a confident mistake is an annoyance. In yours, it's a harm.

So the question is not "should doctors use AI." That ship has sailed; the paperwork burden alone guarantees it. The question is where — and the answer is a hard line that does not move. AI belongs on one side of it. You belong on the other. Everything in this piece is about knowing which is which.

The line that never moves

Picture your work split into two columns. On the left is everything that is essentially writing, summarizing, or translating — the language work that surrounds care. On the right is the care itself: the judgment that requires a licensed human who is responsible for the outcome. AI is genuinely good at the left column and categorically unfit for the right. The whole skill is keeping the line between them bright.

On the left, where AI assists. Clinical documentation is the headline win: you talk, it produces a structured note, and your job shrinks from writing to reading and correcting. Patient-education handouts are nearly as good — it will take your instructions and rewrite them at a sixth-grade reading level, which patients actually follow, in the time it takes to read this sentence. It will condense a long paper or guideline into the points worth checking. It will draft the codes, the prior-auth request, and the insurance appeal. It will write a first-pass reply to a portal message. Notice the common thread: in every case it produces a draft, and in every case you are the editor who decides what's true and what ships.

On the right, where the clinician decides. Diagnosis is naming what is wrong — the differential, the weighing, the call — and that is yours. Treatment is what to start, what to stop, what to dose, and that is yours. The physical exam is hands, eyes, and the bedside read, and no model is in the room with the patient. And the judgment call — risk against benefit, the guideline against the actual person in front of you — is the part of the job that is the job. The left column gives you time. The right column is what you spend it on. The line between them is the most important thing on this page.

What it gets wrong (read this before you trust it)

This is the section the product demos skip, and in medicine it is the section that matters most. Read all of it before you let one of these tools near your work.

It fabricates medical facts, fluently and confidently. Consumer AI tools generate text that sounds correct whether or not it is. Ask one for a dose, a contraindication, a guideline threshold, or a supporting citation and it will frequently produce a clean, authoritative-looking answer that is simply wrong — including references to studies that do not exist. This is not a rare glitch; it is how the technology works. Every clinical fact it gives you is a hypothesis to verify against a primary source, never an answer to act on.

Never paste patient identifiers into a consumer chatbot. A name, a date of birth, a medical record number, anything that identifies a real patient — pasted into a public AI tool, that is a HIPAA breach, full stop. Those tools are not covered by a business-associate agreement, and you have no idea where the text goes. If you use AI on real patient data, it has to be a vetted, contracted, HIPAA-compliant system your organization approved — not the free chatbot in a browser tab. When in doubt, strip every identifier first.

A consumer chatbot is not a diagnostic device. The AI in a browser has not been cleared by regulators as a medical device, has not been validated on clinical outcomes, and is not built to be safe when it's wrong. Treating its output as a second opinion is a category error. It is a writing tool that happens to know a lot of words about medicine.

It carries the bias of its training data. These systems learned from text that under-represents certain populations and over-represents others. That bias shows up as worse, less accurate output for exactly the patients who are already underserved. You cannot see it happening, which is precisely why you cannot lean on the tool for clinical reasoning across a diverse panel.

So verify everything, and stay the final authority. The rule that ties all of this together is simple: AI drafts, you decide. Nothing it writes reaches a patient, a chart, or a prescription pad without a clinician who knows better reading it first and owning the result. You are not in the loop as a courtesy. You are the loop.

Where to start

Do not "adopt AI." That isn't a task, and trying to do everything at once is how people get burned. Pick the single lowest-risk, highest-relief task and do only that.

For most doctors, that task is patient handouts. This week, take the standard instructions you give for one common condition — the after-visit summary for a sprained ankle, a new blood-pressure medication, whatever you explain ten times a week. Paste only that generic, non-identifying text into any major AI assistant and ask it to rewrite it at a sixth-grade reading level in a warm, clear tone. Read what comes back. Fix what's off. Hand the cleaned-up version to your next patient with that condition.

Notice what that exercise contains: no patient data, no clinical decision, no diagnosis — just language work, with you as the final editor. That is the entire pattern of safe AI use in medicine in one rep. You will feel immediately where it saves you time, and you will feel its limits the moment it writes something subtly wrong and you catch it. Once that loop is comfortable, add the next left-column task — note drafting, then guideline summaries, then the prior-auth letters. One task at a time, always staying on your side of the line.

The work that made you a physician — the read of the room, the differential, the decision you're accountable for — is not going anywhere, and no tool is coming for it. What's leaving is the documentation and the paperwork that pulled you away from patients in the first place. The doctors who thrive in the next few years won't be the ones who resisted AI or the ones who trusted it blindly. They'll be the ones who learned exactly where the line sits and never let it move. If you want a structured, plain-language path through the rest of it — what these tools actually are, how to use them safely, one skill at a time — that's what the Academy is built for.