Educational disclaimer: This article is for general educational purposes only and is not legal, regulatory, employment, compensation, tax, or institutional compliance advice. Hospital policy, payer rules, documentation standards, and supervision requirements vary; use your program’s approved policies and consult qualified compliance, legal, privacy, or professional advisors when needed.
You’re on your fourth admission in six hours. The ED is calling again. Your senior wants the discharge summary cleaned up before sign-out. The inbox has radiology callbacks, nursing messages, and one family asking for “just a quick update.” Then somebody—often a chief, an enthusiastic attending, or a hospital admin who hasn’t written a note in years—says, “You should try this AI tool. It’ll save you time.”
That’s the pitch. Save time. Work smarter. Reduce burnout.
What they usually don’t tell you is the part that matters: if the note is wrong, if the med list is wrong, if the summary quietly invents a diagnosis, it’s still your problem. Not the vendor’s. Not the attending’s. Yours. Residents sit in the worst possible spot for sloppy AI adoption—enough responsibility to get burned, not always enough power to set the rules.
Here’s the tension. AI really can help. It can speed up note drafting, organize a chaotic chart, and give you a decent first pass at patient education or a broad differential. I’ve seen it shave real minutes off ugly workflows. But it also hallucinates, smooths over uncertainty, and says stupid things in a confident tone that makes tired people trust it more than they should. That’s the trap.
So if you’re being asked—or tempted—to pilot AI on the wards, this is the practical version. What to use it for. What not to touch. What your attendings may assume you already know. And how to stop before “helpful” turns into dangerous.
Most attendings who like AI mean well. They’re not trying to set you up. But they often skip the ugly middle step between “this tool is promising” and “this is safe in residency.”
AI is not a neutral shortcut. It is not a fancy autocomplete that only speeds up what you already know. It makes judgment calls. Bad ones, sometimes. It fills gaps with plausible nonsense. It misses nuance buried in a consultant’s note from three days ago. It compresses uncertainty into clean sentences that read better than they deserve to.
That’s the first thing people don’t say out loud enough.
The second thing: the danger is not just bad output. The danger is bad output at 5:47 p.m. when you’re behind, hungry, and trying to clear the list before sign-out. That’s when residents stop “auditing” and start “skimming.” I’ve seen this happen with AI-drafted notes that quietly imported old problem lists, wrong antibiotic durations, outdated code status, and fake rationale that sounded polished enough to sneak by.
The verification step is where mistakes happen. Everyone says, “Just review it.” Sure. In theory. In real life, review becomes a fast scroll, especially if the draft looks clean.
So here’s my rule: use AI only for tasks you can independently audit without fooling yourself.
Good uses:
- Drafting a progress note shell
- Turning a messy chart into a timeline you can compare against the record
- Brainstorming a differential you will edit yourself
- Creating a patient education handout in plain language
- Summarizing a guideline, then checking the source
Bad uses:
- Letting AI decide your assessment and plan
- Accepting med changes without line-by-line confirmation
- Using it as a triage brain when urgency is unclear
- Asking it to interpret patient-specific data in an unapproved system
- Treating it like a consultant instead of a draft assistant
If you can’t verify every meaningful fact, don’t let it near the chart. Simple.
How to Pilot AI Without Getting Burned on the Wards
If you want to test AI during residency, don’t start big. Don’t roll it into every note on day one. Don’t use it first on a crashing patient, a complicated discharge, or a consult where the entire point is nuance. That’s how people get embarrassed. Or worse.
Start with one low-risk workflow.
The best starter tasks are boring, repetitive, and easy to check:
- Drafting the non-interpretive parts of a note
- Rewriting patient instructions at a sixth-grade reading level
- Making a checklist for chest pain, AKI, COPD exacerbation, or delirium workup
- Summarizing a guideline update before you read the actual paper
- Organizing a chart timeline from information you already pulled
That’s where AI earns its keep. Structure. Speed. Cleanup. Not judgment.
Now build yourself a safety check. Not a vague “I’ll review it.” An actual mini-checklist you run every time. Mine would look like this:
Before using the tool
- Is this tool approved by my institution?
- Am I about to paste protected health information into something I shouldn’t?
- Do I know exactly what task I’m asking it to do?
Before using any output
- Are the meds correct? Dose, route, frequency, start/stop dates
- Are the labs and vitals real, current, and correctly interpreted?
- Are dates, hospital day, consultant recommendations, and diagnosis names accurate?
- Did the tool make the language cleaner while making the medicine worse?
- Would I still stand behind this if my attending asked me where every line came from?
That last question matters. A lot.
Because AI often “helps” by removing the friction that would have forced you to think. You used to pause before writing an assessment because you had to build the sentence yourself. Now the sentence appears instantly, polished and dangerously easy to accept. Less typing. Also less thinking, if you’re not careful.
Privacy is the next landmine. Never trust a vendor’s cheerful marketing page over your institution’s policy. “HIPAA-compliant” in a demo is not the same as “approved for your residency program to use with live patient data.” If the tool isn’t approved, do not paste PHI. No names. No MRNs. No full dates of birth. No room numbers tied to a case description. De-identify or don’t use it.
And ask directly what your program allows. Not casually. Ask your chief, attending, informatics lead, or compliance contact:
- Is this tool approved for patient-specific use?
- For what workflows?
- Can content generated from it enter the EHR?
- Do I need to disclose its use internally?
- Where is the data stored?
Residents get burned because they assume someone else has already sorted this out. Sometimes nobody has. That’s not rare. Hospitals love innovation right up until someone asks for the policy document.
Also track whether the tool truly saves time. Be honest. AI can save ten minutes and then steal twelve back in editing. That’s not efficiency. That’s admin theater.
Try this for one week:
- Pick one task
- Time how long it takes without AI
- Time it with AI
- Count how many factual errors or awkward rewrites you had to fix
- Decide based on reality, not vibes
If it consistently shortens a safe workflow, keep it. If it creates hidden cleanup, dump it.
One more thing nobody says enough: if you’re using AI to generate differential diagnoses, use it to widen your thinking, not replace it. Good residents ask, “What am I missing?” Bad AI use is asking, “What’s the answer?” Those are not the same question.
A rough rule:
- If the task is reversible, editable, and low stakes, AI can help.
- If the task changes management, affects safety, or requires nuance, you slow down and think with your own brain.
That’s the job. Still.
Related reading: 7 ways to use ambient AI scribes without getting burned in residency.
What to Say to Your Attending, Chief, or IT Team
If you want permission—or at least fewer raised eyebrows—don’t pitch AI like a techno-messiah. That’s annoying, and it makes you sound naïve.
Pitch it like this:
“I want to use it for a narrow task: drafting discharge instructions for common diagnoses. I’ll only use the institution-approved version, I won’t paste PHI into unapproved tools, and I’ll verify every medication, follow-up instruction, and return precaution before anything goes to the patient.”
That works because it answers the real concerns:
- What task?
- What benefit?
- What guardrails?
- Who is responsible?
If an attending pushes back, don’t argue that AI is smarter. It isn’t. Say this instead:
“I’m using it as a draft generator to reduce repetitive writing, not to replace my assessment or final review.”
That’s the right frame. Humble. Accurate. Safe.
If policy is unclear, escalate early. Don’t improvise with patient data just because everyone else is being casual. Bring in the chief, informatics, privacy office, or IT if:
- The tool gives inaccurate or fabricated output
- You don’t know whether it’s approved
- Someone pressures you to use a public tool with PHI
- There’s disagreement about what can go into the chart
And if your institution wants local documentation of AI-assisted drafting, do it cleanly. Something simple. Internal process note, approved disclosure language, or whatever policy requires. But don’t use “AI helped” as an excuse for bad work. The human clinician owns the final product. Always.
The Real Payoff: Using AI as a Resident to Learn Faster, Not Just Work Faster
If all you get from AI is faster note production, you’re thinking too small.
The best use of AI in residency is educational. Have it generate practice questions on hyponatremia workup. Ask it to build a comparison table for causes of anion gap metabolic acidosis. Use it to create a checklist for first-hour sepsis management, then compare that checklist against your hospital protocol. Ask for a plain-language summary of a new guideline, then go read the actual guideline and mark what the tool missed.
That’s useful. Really useful.
The strongest residents I’ve seen don’t hand their thinking to AI. They use it to expose gaps faster. To rehearse. To organize. To pressure-test their own understanding. Then they cross-check with primary sources, UpToDate, society guidelines, pharmacy references, and local pathways. That last part matters because AI is notoriously shaky on local practice patterns, formulary quirks, and the weird realities of your actual hospital.
And you should pay attention to where it fails. AI struggles most when the case is uncommon, symptoms are ambiguous, records conflict, or the decision is high stakes. In other words: exactly the moments when you most need a doctor, not a polished text generator.
So here’s the mindset I want you to keep.
Treat AI like a very eager junior helper. Fast. Useful. Sometimes impressive. Also wrong in ways that are smooth enough to fool tired people. You supervise it. You audit it. You decide when it’s helping and when it’s just creating slick garbage.
That’s the whole game as a resident: use the tool to learn faster and write faster, but never let it think for you when the stakes are real.
The residents who do this well aren’t the ones bragging about using AI for everything. They’re the ones using it quietly, narrowly, and safely—then turning in work that’s actually better.
Related reading: I’m not techy—will I fall behind in an AI-driven healthcare system?.
FAQ
1. Can I use ChatGPT or another public AI tool for patient notes?
Not with identifiable patient information unless your institution explicitly approves it. If there’s any doubt, stop right there. Don’t paste PHI into a public tool and hope policy catches up later. If you want to test prompts, use de-identified examples only and follow your hospital’s rules, not the vendor’s marketing.
2. What’s the safest way to start piloting AI on my rotation?
Pick one low-risk task, like drafting a patient education handout or summarizing a guideline. Then verify every fact yourself. Start small, measure whether it actually saves time, and be ruthless about stopping if it creates more editing, more confusion, or more risk than benefit.
3. How do I tell whether AI is giving me a bad answer?
Watch for fake confidence. If the output is oddly specific, too polished, inconsistent with the chart, or casually changes meds, diagnoses, severity, or urgency, treat it as suspect. The moment it affects management, you go back to source data and a trusted reference. No shortcuts.
4. Will using AI make me look lazy in front of my attending?
Not if you frame it correctly and your work is solid. Tell them you’re using it to draft or organize repetitive content, then explain how you reviewed and corrected the output yourself. Most attendings don’t care about the tool nearly as much as they care whether your final note, plan, and communication are accurate and safe.
