How to Use AI as a Doctor: Uses, Limits, Risks

A doctor's week splits into two different kinds of work filed under one job title: producing and processing information, and forming a clinical judgement about a specific patient. A general AI tool is genuinely good at the first — a first-pass visit note, a plain-language rewrite of a treatment plan, a summary of a long article before a journal club. It has no access to the second, because a diagnosis comes from an exam, a history and a trained judgement call a chat window never sees. Knowing which half of the week you are handing to a tool, and which half stays yours, is most of what makes this useful rather than risky.

Will AI replace doctors covers that split at the level of the whole profession — what is genuinely shrinking and what is not. This is the practical version: what to actually hand over this week, and where a fluent-sounding draft still needs a check before it reaches a patient or a chart.

Start with what makes medicine different from most office work

Two things narrow the options before a tool is even open. Patient health information is protected under federal law, and pasting identifiable patient details into a consumer chat tool with no business associate agreement behind it is not a grey area — it is the kind of mistake that gets reported. Clinical judgement: a physician is accountable for a diagnosis and a treatment decision in a way no tool can be, and nothing here changes who answers for that call.

The practical fix is the same tiered approach that works in every regulated role on this site: information you would never put in a general consumer tool, information that is fine inside a vetted, institution-approved system, and clinical judgement that stays on you regardless of which tool touched a draft first. That tiered thinking traces back to the NIST AI Risk Management Framework, which treats "what happens if this specific use goes wrong" as the question to answer before adoption, not after.

Where AI genuinely helps

The pattern across everything below is the same: the tool drafts text from information you already have and have checked, and you supply the clinical judgement and the final read before anything goes in a chart or to a patient.

1. 01Turning a visit's worth of spoken notes into a structured first-draft note, so you are editing a draft rather than writing from a blank field at the end of a long clinic day.
2. 02Summarising a long journal article or a set of guidelines against a specific clinical question, so you can go straight to the relevant section rather than rereading the whole paper from the top.
3. 03Turning an already-decided treatment plan into plain language a specific patient can actually follow, adjusted for their reading level.
4. 04Drafting a first-pass reply to a routine patient portal message, which you then review and approve before it goes out under your name.
5. 05Drafting routine internal correspondence — a referral letter, a standard administrative note — that still gets your own review before it is sent.

A worked example

Say a patient has just been started on a new medication, with the dose and the plan already decided in the visit, and they ask for the instructions written down in plainer terms before they leave.

The bad version: paste the patient's symptoms into a general chat tool and ask what might be going on, or ask it to adjust the plan on the fly. Either one invites a plausible-sounding answer that was never reviewed by anyone clinically responsible for it, delivered with the same fluent confidence whether it is right or wrong — a property of how these systems generate text, not a sign they have checked anything against this patient's actual case.

The version that works: take the plan you have already decided and ask the tool only to restate it in plain language at a lower reading level, with nothing added or changed from what you actually decided. That is the same instinct OpenAI and Anthropic both give in their own prompting guidance: name the exact task and the exact constraint, rather than asking an open question and trusting whatever synthesis comes back. You read the result against your own plan before handing it over, every time.

A model can rewrite a treatment plan in plainer language faster than you can retype it. It cannot be the one accountable for what the plan actually says.

Where it will quietly get you into trouble

• —Any question framed as "what could this symptom mean" or "is this the right dose". A model can produce a fluent, confident clinical-sounding answer that is simply wrong — a widely cited survey of hallucination in large language models covers why fluent and accurate are not the same property in these systems, and a wrong answer here is not a stylistic slip.
• —Identifiable patient information pasted into a consumer tool with no institutional agreement behind it. Responsible AI examples that actually come up at work covers the redact-first habit that applies before anything else does, and for anything clinical the simpler answer is not to paste the real record in at all.
• —Treating a rewritten patient letter as equivalent to the plan you actually decided, rather than a plain-language version of it — the decided plan still governs, and any discrepancy gets corrected against it, not around it.
• —Letting a cleanly worded draft substitute for the actual check. Fluent, well-formatted text is not the same signal as clinically correct text, and a patient will not know the difference until it matters. How a worked method catches an added detail in a summary is the general version of the check worth running on anything a tool drafts before it goes out under your name.

Getting started without a project plan

1. 01Check your institution's policy on AI tools and patient data before using any tool on real patient information, even informally. A short written policy beats guessing, and most institutions already have one.
2. 02Pick one recurring, non-clinical task first — a visit-note first draft or a plain-language rewrite of an already-decided plan are the usual right answers, because the output is instantly checkable against what you actually decided.
3. 03Run that task for a few clinic days and keep the time actually saved, not a general impression that it helped.
4. 04Draw one hard line and keep it: nothing that touches a diagnosis or a treatment decision goes near a general-purpose tool. How to use AI as a nurse covers the same hard line one level over on the same care team, for anyone comparing notes across roles.

The part that actually protects the role

The tasks these tools genuinely speed up — summarising, rewording, first-draft notes — are the ones furthest from the exam, the history and the judgement that make up the diagnosis itself, which is also why medicine holds up well in the research on this. Will AI replace pharmacists covers the same tiered-risk approach one step further down the prescribing chain, where the licensed sign-off plays the role the diagnosis plays here.

The useful stance is neither refusing the tools nor leaning on them for anything clinical. It is being the doctor who gets a visit note or a patient letter drafted in a fraction of the usual time, and never once asks a chat tool to think for them about a patient.

This also explains why the regulated clinical-decision tools that do touch diagnosis or risk scoring look nothing like a general chat tool in how they reach a clinic. They go through validation, oversight and a defined scope before anyone uses them on a real patient, and they are deployed alongside a physician's own judgement rather than in place of it. A consumer chatbot has none of that behind it, which is the real reason the line between "drafts text you review" and "makes a clinical call" is not a formality — it is the entire difference between a tool that helps and one that creates a new kind of risk.

The short version

Use AI for visit-note drafts, literature summaries, plain-language rewrites of already-decided plans and routine correspondence, on an institution-approved tool with nothing identifiable pasted into a consumer product. Never ask it a clinical question and never let it near a diagnosis or a treatment decision. Check your institution's policy before you start, pick one checkable task, and keep the hard line exactly where it is.

That checking habit is also the one the labour market has started pricing in more broadly: PwC has measured a real wage premium for workers with demonstrable AI skills across the wider economy — 62%, up from 57% the year before — even as Challenger, Gray & Christmas tracks AI-attributed cuts concentrated almost entirely in document-processing roles medicine is not. Coursium teaches the practical layer underneath that — using the tools well on the paperwork, and knowing exactly where to stop. Stay ahead of AI by learning them on your phone.