4  Centaur or cyborg: what you hand to the machine, and what you keep

AIM-5 · AIM-X · AIM-4 · about 90 minutes

NoteWhat you’ll learn

By the end of this chapter you will be able to:

1. Sort the tasks on an intern’s list into three kinds of use. For each, give a reason that names the task’s risk and whether any tool has been validated for it.
2. Tell deskilling from never-skilling from mis-skilling, and name one task in your own coming internship at risk from each.
3. Rewrite a prompt that asks for an answer into one that asks for a hint, using the Who / Where / What pattern, and say what the rewrite made you do that the original did not.
4. State what the randomized trials of physicians working with a language model found, what the randomized studies of learners found, and name two techniques that counter the second finding.
5. Say what you document, and whom you tell, when a machine drafted something that carries your signature.
6. Write down which part of clinical judgment you will not hand over, and why.

Time. About 30 minutes of reading, 25 minutes of listening, 30 minutes of doing, and five minutes at the end that you should not skip. You need one general-purpose chatbot for the second half of the Do; your institutional ChatGPT Edu or Microsoft Copilot account is enough, and so is any free general-purpose chatbot. The first half of the Do needs only a pen. There is nothing to install.

4.1 The first list

It is 6:10 on the first Monday of intern year. You have six patients, a badge that opens more doors than it should, and a list.

Mon · day 1 · before rounds

1. Pre-round: beds 4, 7, 9, 12, 14, 15. Overnight events, vitals, labs.
2. Progress notes x6 before 8:00.
3. Discharge summary, bed 12 (admitted 11 days ago; three services involved).
4. Portal inbox: four patient messages from the weekend.
5. Call nephrology about bed 9. What exactly are we asking?
6. Journal club slides, Thursday. Have not picked the paper.

Illustrative intern task list, made up for this chapter. No patient on it is real.

Your senior resident has been an intern for exactly one year. She looks at the list and says what every senior says now: “Have it draft the notes. Have it draft the discharge summary. That’s what it’s for.” She means the assistant built into the chart, the one with the sparkle icon, or the institutional chatbot in the other tab. She is right that it can. Last week, on your sub-internship, you watched an attending send an AI-drafted reply to a patient message without reading it. Nobody said whether that was fine.

Look at the list again with one question. For each line, if you hand it to the machine, what do you get back, and what do you lose? The notes are the easy case. Even there the answer splits in two. A note that is drafted for you and a note that is written by you are different things that share one template. The discharge summary is the hard case. It is the only document on the list that forces you to read eleven days of chart and decide what happened to this person. That reading is most of what you will know about them when their primary care doctor calls you next month. The consult question is a different kind of hard. The machine will produce a very good question. But the reason you are calling is that you do not yet know what to ask.

This chapter is about that sorting. It does not ask whether you should use these tools at all. You will meet them in the chart regardless, and “never” is one of the verdicts open to you for any single line. It asks what kind of use, for which task, at what risk. And it asks you to write your answer down.

4.2 Why this matters

In three months, the list above is yours. The tool will be in the chart whether or not you have decided how to use it. Health systems are building these assistants into systems you do not control. You will use them more often, with less supervision, and with more freedom, in roughly that order. Two of this course’s faculty, quoted from its prior session, describe the two ends of this. One put it bluntly at a national medical-education meeting: you can choose to avoid AI, but is that really an option? (Chad Stickrath, MD). The other, a pediatric hospitalist, said what this chapter must not forget: we should be scared, because what we do is scary (Christina Olson, MD). Both are right.

There is a second reason, and it is specific to your stage of training. You are about to learn clinical reasoning for the first time, on real patients. A tool will be available that produces a plausible answer to every question you were supposed to struggle with. Someone who has practiced for fifteen years and starts using these tools risks losing a skill. You risk never gaining it, or gaining it wrong. Those are three different problems. The most useful point in this chapter is that they need three different defenses.

4.3 How it works

4.3.1 Two ways to work with a machine

the appraisal chapter ends with a randomized trial and a puzzle, and this chapter starts there. Fifty physicians were each given six diagnostic cases. Half also had a language model: the kind of AI behind ChatGPT and Copilot, a program that produces text in reply to text. The physicians with the model did no better than the physicians without it. The adjusted difference was 2 points on a 100-point reasoning score (95% CI −4 to 8). The model working alone scored 16 points above the physicians (95% CI 2 to 30) (Goh et al., 2024).

Figure 4.1: Median diagnostic-reasoning score per case, as a percentage of the rubric, in the 2024 randomized trial. The model-alone bar is drawn at 74 + 16, because the paper reports the difference, not a median. Redrawn from the published numbers (Goh et al., 2024); data are not copyrightable.

The tool was capable. The physicians using it did not get the benefit. The authors read the gap as a training problem. They call for “technology and workforce development to realize the potential of physician–artificial intelligence collaboration.” Workforce development means you.

The same group ran a second trial a year later, this time on management rather than diagnosis. Ninety-two physicians worked through five cases, making treatment and testing decisions as each case unfolded. This time the physicians with the model scored 6.5 points higher than those without (95% CI 2.7 to 10.2). They spent about two minutes longer per case. Their scores could not be told apart from the model alone (Goh et al., 2025). Two trials, same team, same kind of tool, opposite result. The thing that changed was not the model. It was the task, and how the physicians worked with it.

There are names for those ways of working. They come from a 2023 field experiment in which several hundred management consultants were given GPT-4, the model behind ChatGPT at the time. The researchers watched how the successful ones worked (Dell’Acqua et al., 2023). Two patterns appeared, and the names are still used:

• A centaur divides the work. Like the creature, there is a clean line between the human half and the horse half. You give the machine one clearly marked piece of the task. You keep the rest. You check what comes back before you use it anywhere.
• A cyborg blends the work. You move back and forth with the machine inside a single piece of work: drafting a sentence, asking for three alternatives, taking one, rewriting it, asking what is missing.

Neither way is the correct one. Chess players were the first to use the word centaur, after Kasparov’s human-plus-computer exhibition matches in the late 1990s. Within a few years they found the same thing: in a 2005 online tournament, two amateurs with three computers and a good process beat grandmasters with a worse one.¹ The skill is knowing which way of working a task calls for, and switching. The 2025 NEJM review this chapter is based on gives a simple rule (Abdulnour et al., 2025). Work as a centaur for high-stakes tasks, and for tools that have not been validated for the use you are putting them to. Work as a cyborg where the task is low-risk or creative, or where the tool is well validated for exactly this. Figure 4.2 turns that rule into two questions.

flowchart TD
  accTitle: Which way of working a task calls for
  accDescr: Decision flowchart. The task in front of you leads to the question, if the output is wrong, who is harmed and how badly. If a patient, seriously, ask whether any tool is validated for this exact use at your institution. No leads to centaur or not at all, hand over one clear piece and check every line. Yes leads to centaur, hand over the one piece and check before signing. If mostly me or nobody, ask whether the tool is validated for this use. No leads to cyborg is reasonable, stay involved and check what matters. Yes leads to cyborg, go back and forth freely.
  T["The task in front of you"] --> R{"If the output is wrong,<br/>who is harmed, and how badly?"}
  R -- "A patient, seriously" --> V1{"Is any tool validated<br/>for this exact use,<br/>at your institution?"}
  R -- "Mostly me, or nobody" --> V2{"Is the tool validated<br/>for this use?"}
  V1 -- No --> C1["Centaur, or not at all.<br/>Hand over one clear piece.<br/>Check every line."]
  V1 -- Yes --> C2["Centaur.<br/>Hand over the one piece.<br/>Check before signing."]
  V2 -- No --> C3["Cyborg is reasonable.<br/>Stay involved.<br/>Check what matters."]
  V2 -- Yes --> C4["Cyborg.<br/>Go back and forth freely."]

Figure 4.2: Which way of working a task calls for. Two questions to ask before you open the tool. Adapted from the framework in Abdulnour et al. (2025).

the appraisal chapter spends two hours on the phrase “validated for this use.” Here it means one thing. Someone has measured how the tool performs on this task, on patients like yours, and you know the number. Take an ambient scribe, the tool that listens to a visit and drafts the note. With three randomized trials behind it, it is validated for drafting a visit note. The same product is not validated for writing your discharge summary. The chatbot in the other tab is validated for nothing on your list. The validation column will look different in twelve months. The sorting will not.

4.3.2 Three ways to lose a skill

The 2025 review names three perils. They are worth learning as three words, because the defenses differ (Abdulnour et al., 2025):

• Deskilling: losing a reasoning skill you had.
• Never-skilling: never learning a skill at all, because the tool arrived before the skill did.
• Mis-skilling: learning a wrong pattern, because the tool’s plausible wrong answers became your template.

None of these is hypothetical. Two of them have been measured.

Deskilling has a number. At four endoscopy centers in Poland, computer-aided polyp detection was switched on at the end of 2021. In the three months before, endoscopists doing ordinary colonoscopy without the tool found adenomas in 28.4% of patients. In the three months after, doing the same colonoscopy on days the tool was off, they found them in 22.4%. That is an absolute drop of 6 points (95% CI −10.5 to −1.6). It happened in a skill these were experts at, after a few months of having a machine draw the boxes for them (Budzyń et al., 2025). The study is observational and the authors say “might.” But the direction is the one every clinician who has stopped reading their own ECGs would predict.

Never-skilling has a trial. Nearly a thousand high-school students in Turkey were randomized during math practice sessions to one of three groups: no AI, a plain GPT-4 chat, or a GPT-4 “tutor” prompted to give hints rather than answers. While they had the tool, both AI groups did far better on practice problems: 48% higher scores with the plain chat, 127% with the tutor. Then the tool was taken away for the exam. The plain-chat students scored 17% worse than the students who had never had access. The tutor group’s losses were much smaller (Bastani et al., 2025). The authors’ word for what the plain chat became is crutch. The finding is not that the tool failed. The thing it improved and the thing the students needed were two different things.

Mis-skilling has no trial yet. That is why the chapter draws it instead. Below is a progress note drafted by an assistant from a made-up chart. The patient is invented; the note is the kind you will be handed. Read the plan.

✨ Draft progress note · hospital day 3Accept · Edit · Discard

S: Pt reports breathing "a little better." Slept flat x2 pillows. No CP. O: T 37.1 HR 88 BP 118/70 RR 18 SpO2 94% RA. Wt 84.2 kg (admit 87.0). JVP 8 cm. Bibasilar crackles, improved. 1+ pedal edema. Na 133 K 3.4 Cr 1.6 (baseline 1.0; yesterday 1.3) BUN 41. BNP trending down. CXR: improving vascular congestion. A/P: 1. Acute decompensated HFrEF, improving. Continue furosemide 80 mg IV BID. Daily weights, strict I/O. Goal net negative 1-1.5 L. 2. AKI, likely prerenal in setting of poor PO intake and diuresis. Continue IV fluids at 75 mL/h. Trend Cr. Hold nephrotoxins. 3. Hypokalemia. Replete K to >4.0. 4. HTN. Continue home lisinopril. Dispo: continue inpatient diuresis.

Illustrative draft, made up for this chapter, for a fictional patient. Not the output of any real product.

Problem 2 is fluent and correctly formatted. It also contradicts problem 1. The plan removes fluid from the patient and gives fluid to the patient in the same note. A rising creatinine on day 3 of diuresis is the expected cost of getting fluid off a congested kidney. The decision an intern is supposed to make here, and to be quizzed on at rounds, is whether to hold the diuretic, slow it, or continue it. The draft made that decision by template. If you accept it, you have a note with a wrong plan. If you accept it on ten patients, you have a pattern: creatinine up, add fluids. That is mis-skilling. The danger is that nothing about the sentence looks wrong. It looks like every other note.

Table 4.1: Three perils, three defenses. The middle row is where a fourth-year is most exposed.

Peril	What it is	The intern task most exposed	The defense
Deskilling	A skill you had fades when you stop using it	Reading your own images, ECGs, and strips before the machine’s read	Do it first, then look. Keep a count of where you disagreed and who was right.
Never-skilling	A skill you never build because the tool answered first	The differential; the consult question; the discharge summary’s narrative	Start on your own. Ask for hints. Take the tool away before the test, because residency will.
Mis-skilling	A wrong pattern learned from plausible output	Any drafted assessment and plan	Read the reasoning, not the conclusion. Ask why this and not that. Never accept a plan you could not have written.

4.3.3 The learners’ study, and six techniques

The Turkish trial measured grades. A separate randomized study measured what learners do. It compared students working with a chatbot, with a human expert, and with a checklist tool. It recorded their behavior as well as their scores. The chatbot group did better on the immediate task. They did not gain more knowledge, more motivation, or more ability to apply what they learned to a new problem. The authors’ name for the behavior behind that gap is metacognitive laziness: handing the job of monitoring and checking your own thinking to the tool (Fan et al., 2024). Metacognition is the ordinary word for noticing what you know and do not know. Laziness here is not a moral judgment. It is what any of us does when there is a faster way.

The same authors’ answer to “should learners use these tools?” is probably, yes. The course’s prior session turned their recommendations into five techniques and added a sixth. They are the defense against never-skilling in Table 4.1, made specific:

1. Start on your own. Build your differential before you ask anything. The struggle is the part that does the learning. Ask for help when you are truly stuck, not when you are merely unsure.
2. Ask for hints, not answers. Not “give me the differential for this presentation.” Instead: “I have settled on a cardiac cause and I think I am missing something in the abdomen. What categories should I be considering that I have not named?”
3. Teach back. State your reasoning in your own words and ask whether it is right. Do not ask for the reasoning to be supplied.
4. Focus on the why. Why this drug and not that one, why this cutoff, why first-line. The reason is what lets you apply the knowledge to the next patient.
5. Use the four patterns on purpose. Explanation of something you already read. Debugging your own reasoning after it failed. Clearing up a concept. Learning a better approach after you have solved the problem yourself.
6. Ask it to quiz you. When the tool tests you, it is least able to make you dependent on it. It also turns around the relationship in the opening quote of the prior session, where a student described comparing their own thinking against the machine’s “as if the program is a mentor or learner.”

4.3.4 Who, where, what

Half of the difference between a hint and an answer is in the prompt. If English is not your first language, or you are new to these tools, the difference is not obvious. The pattern the course uses has three parts:

• Who you are: “I am a fourth-year medical student.”
• Where you are: “I am on an inpatient medicine rotation, day 3 of a patient admitted with a heart-failure exacerbation.”
• What you want, phrased as the thing you want to understand rather than the thing you want produced: “I want to understand how to think about a rising creatinine during diuresis, and I want you to ask me what I think first.”

Here is the same request both ways, for the made-up patient above.

NoteAnswer-seeking

67M with HFrEF exacerbation, day 3 of IV furosemide, creatinine 1.0 → 1.3 → 1.6. What should I do about the AKI?

TipHint-seeking

I am a fourth-year medical student on an inpatient medicine rotation. My patient was admitted three days ago with a heart-failure exacerbation and is being diuresed; weight is down 3 kg, still has crackles and edema, and creatinine has gone from 1.0 to 1.6. I think this is expected cardiorenal physiology and I want to keep diuresing, but I am not confident. Do not tell me what to do. Ask me two questions that would change your mind, then tell me what I have missed.

The first prompt gets a plan, and sometimes the plan will be the one in the draft note. The second gets a conversation. In it you have to say what you think before the tool says anything. That is the only arrangement in which being wrong teaches you something.

4.3.5 The practical rules, in the order you will need them

This section is for the reader who skipped the explanations above. Each rule has a reason and a limit.

You sign the note. Whatever drafted it, the signature is yours, and so is the responsibility. In every chart review, malpractice case, and quality-committee process that exists today, the question is what the signing clinician knew and did, not what the software suggested. The limit: a tool your institution licenses and runs inside the EHR (the electronic health record) is covered by a business associate agreement. That is the contract that makes a vendor legally responsible for protecting patient data. The tool is also governed by an institutional policy, which may say how to document its use. Ask what that policy says in your first week. A consumer chatbot in the other tab has no such agreement. At CU Anschutz, ChatGPT Edu and Microsoft Copilot are covered when you sign in with your university credentials. The same tools through a personal account are not.

No patient information in a consumer tool, ever. A consumer tool is a chatbot you reach through a personal account: ChatGPT, Claude, or Gemini on your own login. Not a name. Not a paraphrase of a case with the name removed. Not a screenshot. The temptation you will feel is time pressure at 6:10 a.m. The rule holds. A made-up case that makes the same teaching point is always available, and this chapter gives you one.

Document what you did, in the words your institution uses. Some systems add a line to AI-drafted notes and messages automatically. Some require the clinician to attest, that is, to add a signed statement. Some say nothing yet. VERIFY: current attestation and disclosure language at the course institution; policies differ by system and change yearly The lasting rule is this. You should be able to answer “did a tool draft this?” honestly and without hesitation, to a patient, an attending, or a lawyer. The answer should already be in the record if your institution has a place for it.

Telling the patient is a separate question from documenting. The attending who sent the unread draft made two decisions: not to read it, and not to tell the patient. The second is the one the patient would care about. It is the subject of the patient-communication chapter. Here, one number. When a large academic system turned on AI-drafted replies for 162 clinicians, the drafts were used for about 20% of replies. Reply time did not change. Clinicians’ task load and exhaustion fell significantly (Garcia et al., 2024). The tool reduced burden without saving time. That is what you would expect if the clinicians were reading and editing the drafts. That is the centaur version of the task. Sending unread is not a way of working. It is the absence of one.

The one sentence for a co-intern. If you could not have written it, do not sign it. That sentence covers the drafted note, the discharge summary, the consult question, and the portal reply. It is the working form of the third row in Table 4.1.

4.3.6 Your supervisors are learning this too

One more thing from the NEJM review, and it changes how you should read your own supervisors in three months. The authors note that supervisors often have less experience with these tools than their learners. They recommend that supervisor and learner look into the question together, rather than the supervisor acting as the authority (Abdulnour et al., 2025). The structure they propose for that conversation is a debrief. You will be the one being debriefed, so learn it now.

NoteDEFT-AI: the debrief

• Diagnosis, or discussion: what was your reasoning, and how did the tool assist?
• Evidence: how did you verify?
• Feedback: what could improve?
• Teaching: what did you learn, about the tool and about reasoning?
• AI recommendation: what are the next steps for safe use?

Five questions. The second is the one that matters. If the honest answer to “how did you verify?” is “I did not,” the rest of the debrief follows from that.

The review names one more thing the perils above depend on, and calls it the foundation: adaptive expertise. That is the ability to shift between fast, routine practice and slower, creative problem-solving. Critical thinking, in their account, has to be taught, modeled, and assessed during AI use rather than assumed to survive it. That is the reason this course exists. It is the reason this chapter ends with you writing a policy rather than reading one.

4.4 Watch or listen

Podcast. NEJM AI Grand Rounds, episode 35, “Medicine, Machines, and Magic: Dr. Jonathan Chen on Medical AI” (October 15, 2025; 48 min). https://ai-podcast.nejm.org/e/medicine-machines-and-magic-dr-jonathan-chen-on-medical-ai/

Chen is the senior author of both randomized trials above. Budget 25 minutes. Listen for three things. First, his own account of why the model beat the physicians in the first trial and why the second trial came out differently. Second, what he says about automation anxiety, which is the feeling the perils section is about. Third, the claim in the episode’s own summary that empathy, rather than memorization, may become the most valuable clinical skill. Then ask the question this chapter asks: which of the tasks on the 6:10 a.m. list does that claim actually change, and which does it leave exactly where it was?

Alternative. NEJM AI Grand Rounds, “AI and the Evolution of Medical Thought with Dr. Adam Rodman” (June 19, 2024). https://ai-podcast.nejm.org/e/ai-and-the-evolution-of-medical-thought-with-dr-adam-rodman Rodman is a co-author on both trials and a historian of clinical reasoning. The episode takes the historical view: what clinicians have handed to instruments before, and what happened to the skill each time. The episode is 53 minutes.

4.5 Do

WarningNo patient information, ever

Part A needs no tool. Part B uses the made-up patient from this chapter, or a made-up case you write yourself. Do not use a patient from your sub-internship, a classmate’s case, or a de-identified real chart. Consumer chatbots have no business associate agreement. Your university ChatGPT Edu and Copilot accounts have one, but it covers patient care, and this is coursework. Free tiers may also use what you type to train the next model.

What counts as patient information, and why taking out the name is not enough: the patient information page.

4.5.1 Part A: your policy (20 minutes)

This is the piece of work you will keep from this chapter. It is a set of positions, not evaluations. You will be asked to reread it in the final session of the course.

Step 1. Sort the eight tasks. These are the tasks from the course’s prior session. They are the 6:10 a.m. list with the phone call and the pre-rounding added. For each one, write a verdict, a risk, whether you know of a validated tool for it, the peril it exposes you to most, and one sentence of justification. Copy the grid into a document.

Task                              Kind of use (centaur / cyborg / not at all)
                                  Risk if wrong (patient / me / nobody)
                                  Validated tool at my institution? (Y / N / don't know)
                                  Peril (deskill / never-skill / mis-skill)
                                  One-sentence rule

1. Drafting a progress note       ______  ______  ______  ______  ______________________
2. Writing a progress note        ______  ______  ______  ______  ______________________
3. Literature search              ______  ______  ______  ______  ______________________
4. Building a differential        ______  ______  ______  ______  ______________________
5. Patient education handout      ______  ______  ______  ______  ______________________
6. Journal club slides            ______  ______  ______  ______  ______________________
7. Summarizing a paper            ______  ______  ______  ______  ______________________
8. Answering a patient message    ______  ______  ______  ______  ______________________

The task on this list I would add: _________________________________
Its row: ______  ______  ______  ______  ______________________

The first two rows are deliberately a pair. If you gave them the same verdict, stop. Ask what the difference is between a task where a human is already writing and one where nobody is. Most of the learning in this exercise is in that pair.

For row 3, you may want a real search to sort rather than an imagined one. Use this pattern with your specialty as the variable, and take the first paper that fits your journal-club slot:

("large language model" OR "artificial intelligence") AND <your specialty>
AND (validation OR "external validation")

Filter to the last three years. The paper counts toward this week’s resource log.

Step 2. Add two columns. For every row you marked centaur or cyborg, write what you would document and whom you would tell. “Nothing” and “nobody” are acceptable answers if you can say why. For row 8, answer for the patient specifically.

Step 3. Find your disagreement. Pick the row you are least sure of. Write the strongest case for the opposite verdict, in two sentences, as if a co-intern were making it. Then say whether you have changed your mind, and on what grounds. An unresolved disagreement, written down with both positions, is a better result than a forced answer.

4.5.2 Part B: the prompt (10 minutes)

Step 1. Take the task you rated most suited to cyborg use. Write the prompt you would actually type: the fast version, the one you would type at 6:10 a.m. Keep it.

Step 2. Rewrite it with Who / Where / What. Make it ask for a hint, a question, or a check on your reasoning rather than the finished product. Use the made-up heart-failure patient above if the task needs a case, or write your own in three lines with no real person in it.

Step 3. Run the second prompt. Then, in the same conversation, type: “Now quiz me on this. Three questions, one at a time, and tell me when I am wrong.” Answer them.

Step 4. Record it.

Task: ____________________________________________

Fast prompt (as I would have typed it):
______________________________________________________________

Rewritten prompt (Who / Where / What, hint-seeking):
______________________________________________________________
______________________________________________________________

One thing the rewritten prompt made me do that the fast one would not have:
______________________________________________________________

Of the three quiz questions, the one I got wrong, and why:
______________________________________________________________

If the tool answered the rewritten prompt with a finished plan anyway, note that. It happens. It is the reason “ask for hints” is a technique rather than a setting.

TipIf you have ten minutes and no appetite for the chatbot

Do Step 1 of Part A and nothing else. The sort is a judgment task, not a tool task. A completed grid with honest “don’t know” entries in the validation column is enough. The tool can wait until you have decided what you would let it do.

4.6 Check

Score yourself against this before moving on.

You have	Meets	Falls short
Eight sorted tasks	Each verdict names the risk and the validation status, and the draft/write pair has different verdicts or a stated reason they do not	Verdicts without risk or validation; the pair sorted together without comment
A peril per row	Deskilling, never-skilling, and mis-skilling each appear at least once, on a task where that peril fits	One word used for every row
Documentation and disclosure columns	Row 8 answers for the patient, and “nothing” is justified where it appears	Columns blank or “N/A”
A written disagreement	The opposing case is one a co-intern could actually make, and your verdict states its grounds	A weak version no co-intern would make, or “I still think I’m right”
Two prompts	The rewrite says who, where, and what, and asks for a hint or a question rather than the product	The rewrite is the fast prompt with “please” added
The two findings	You can say what the two physician trials found, what the two learner studies found, and why they are not in conflict	“AI is better than doctors” or “AI hurts learning” alone
Two techniques	Two of the six, in your own words, with the peril each defends against	“Be careful”
The one sentence	If you could not have written it, do not sign it, and you can say which task on the list it applies to most	Only the slogan

4.7 What this changes for you

What does this change about how you’ll practice? Go back to 6:10 a.m. The senior resident has said “have it draft the notes,” and she is waiting. Your grid is your answer. But say it in two sentences she could repeat: which lines on the list go to the machine, in which way, and which one line does not go at all.

Then the sentence this course collects and rereads in the final session. Which part of clinical judgment will you not hand over, and why? Write it, date it, and keep a copy where you will find it in four weeks and, ideally, in four years. Not a category (“patient care”) and not a slogan. A task, and a reason. If you cannot yet name one, write that down instead. It is an honest answer, and it is the one this chapter most wants you to revisit.

4.8 Summary

• The question is not whether to use the tool. It is what kind of use, for which task, at what risk. A centaur divides the work and checks it; a cyborg blends the work and goes back and forth. The task’s risk and the tool’s validation decide which.
• In one randomized trial, physicians with a model did no better than physicians without it, while the model alone did. In a second trial, on management tasks, they did better. The model was not the variable. How they worked with it was.
• Deskilling, never-skilling, and mis-skilling are three problems with three defenses. A fourth-year is most exposed to the middle one.
• Randomized studies of learners find the tool improves the task and not the learning. Asking for hints, teaching back, and being quizzed are the techniques that reduce that gap.
• You sign the note. Institutional tools are covered by an agreement and a policy; the chatbot in the other tab is covered by neither.
• If you could not have written it, do not sign it.
• The list at 6:10 a.m. is yours in three months. The sort is yours now.

4.9 Go deeper

Papers

• The review this chapter is based on, with the three perils, adaptive expertise, and DEFT-AI in full (Abdulnour et al., 2025).
• The two physician trials: diagnosis (Goh et al., 2024) and management (Goh et al., 2025). Read the discussion sections for what the physicians actually did with the tool.
• Never-skilling in a randomized field experiment (Bastani et al., 2025), and deskilling in a multicenter observational study (Budzyń et al., 2025).
• The learner study that named metacognitive laziness (Fan et al., 2024); the implications section is written for teachers and is the source of the six techniques.
• The consultants’ experiment that gave centaur and cyborg their current meaning (Dell’Acqua et al., 2023). Not medicine, but worth reading for one finding: the tool helped most on tasks inside its competence and hurt on tasks just outside it. The authors call that edge the jagged frontier.
• AI-drafted patient replies at scale: no time saved, burden reduced, drafts read and edited (Garcia et al., 2024).

Talks

• Prompt engineering: ten rules with clinical examples throughout. Who / Where / What is a shortened form of the first three. the prompting lab is based on this deck.
• AI in medical education: attribution and academic integrity, for the moment one of these tools drafts something with your name on it outside the hospital.

Abdulnour, R.-E. E., Gin, B., & Boscardin, C. K. (2025). Educational Strategies for Clinical Supervision of Artificial Intelligence Use. The New England Journal of Medicine, 393(8), 786–797. https://doi.org/10.1056/nejmra2503232

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences of the United States of America, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122

Budzyń, K., Romańczyk, M., Kitala, D., Kołodziej, P., Bugajski, M., Adami, H. O., Blom, J., Buszkiewicz, M., Halvorsen, N., Hassan, C., Romańczyk, T., Holme, Ø., Jarus, K., Fielding, S., Kunar, M., Pellise, M., Pilonis, N., Kamiński, M. F., Kalager, M., … Mori, Y. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: A multicentre, observational study. The Lancet. Gastroenterology & Hepatology, 10(10), 896–903. https://doi.org/10.1016/s2468-1253(25)00133-5

Dell’Acqua, F., McFowland III, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality. Elsevier BV. https://doi.org/10.2139/ssrn.4573321

Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2024). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544

Garcia, P., Ma, S. P., Shah, S., Smith, M., Jeong, Y., Devon-Sand, A., Tai-Seale, M., Takazawa, K., Clutter, D., Vogt, K., Lugtu, C., Rojo, M., Lin, S., Shanafelt, T., Pfeffer, M. A., & Sharp, C. (2024). Artificial Intelligence-Generated Draft Replies to Patient Inbox Messages. JAMA Network Open, 7(3), e243201. https://doi.org/10.1001/jamanetworkopen.2024.3201

Goh, E., Gallo, R. J., Strong, E., Weng, Y., Kerman, H., Freed, J. A., Cool, J. A., Kanjee, Z., Lane, K. P., Parsons, A. S., Ahuja, N., Horvitz, E., Yang, D., Milstein, A., Olson, A. P. J., Hom, J., Chen, J. H., & Rodman, A. (2025). GPT-4 assistance for improvement of physician performance on patient care tasks: A randomized controlled trial. Nature Medicine, 31(4), 1233–1238. https://doi.org/10.1038/s41591-024-03456-y

Goh, E., Gallo, R., Hom, J., Strong, E., Weng, Y., Kerman, H., Cool, J. A., Kanjee, Z., Parsons, A. S., Ahuja, N., Horvitz, E., Yang, D., Milstein, A., Olson, A. P. J., Rodman, A., & Chen, J. H. (2024). Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial. JAMA Network Open, 7(10), e2440969. https://doi.org/10.1001/jamanetworkopen.2024.40969

---

1. Garry Kasparov, “The Chess Master and the Computer,” New York Review of Books, 11 February 2010, https://www.nybooks.com/articles/2010/02/11/the-chess-master-and-the-computer/. Kasparov’s own account; an essay, not a study.