https://youtu.be/zKMWLs3S9yg?si=qqgh74nQYdmPgOHm
Dear friend, Robert Wachter’s central argument in A Giant Leap: How AI Is Transforming Healthcare and What That Means for Our Future (2026) is surprisingly pragmatic:
AI does not have to become a perfect doctor. It only has to be meaningfully better than the expensive, exhausted, fragmented healthcare system we already have.
He is cautiously but distinctly optimistic. Unlike the first digital transformation—the electronic health record, which often turned physicians into “expensive, grumpy data-entry clerks”—generative AI may finally add intelligence to the information that medicine has spent decades digitizing.
1. From electronic filing cabinet to clinical intelligence
For Wachter, electronic health records accomplished only the first stage: they converted paper into data. They made laboratory results and notes retrievable, but did little to help clinicians understand a complicated patient.
Generative AI is different because it can read the narrative portions of the chart, synthesize heterogeneous information, and respond to a complex clinical question in ordinary language. Instead of merely locating documents, it can potentially connect:
- symptoms;
- medications;
- laboratory trends;
- imaging;
- previous notes;
- guidelines and medical literature.
Wachter describes this as moving from a large digital filing cabinet toward something resembling a clinical consultant. For a general physician, AI may provide near-instant access to subspecialty-level knowledge that previously required locating and consulting several different experts.
The shift, therefore, is roughly:
EHR: stores what happened.
AI: helps interpret what it might mean and what might be done next.
2. AI will initially transform the work around medicine
Wachter does not think the earliest and most consequential changes will necessarily be spectacular autonomous diagnoses. They are more likely to occur in the mundane work that currently consumes clinical life:
- ambient documentation and AI scribes;
- summarizing long charts;
- drafting discharge instructions and patient messages;
- prior authorization and administrative correspondence;
- retrieving relevant evidence;
- preparing a structured account before the clinical encounter.
These applications matter because contemporary medicine is buckling under documentation, messaging and bureaucratic workload. AI scribes may allow physicians to look at the patient rather than the screen, while clinical copilots can reduce the cognitive labour of searching through an enormous chart.
Thus, his first promise of AI is not necessarily superhuman medicine, but the recovery of attention:
The doctor may once again be able to face the patient.
There is, however, an irony here: AI might restore human contact precisely by taking over some linguistic work formerly regarded as belonging to the physician.
3. The medical encounter will become tripartite
Wachter anticipates that the future consultation will no longer consist simply of doctor and patient. It will increasingly include an AI that has:
- questioned the patient before the visit;
- organized the history;
- identified missing information;
- proposed possibilities;
- documented the encounter;
- helped communicate the plan afterwards.
The patient might converse with an AI for much longer than the doctor could ever afford to do. It has effectively unlimited time to inquire about symptoms, explain laboratory findings and answer repetitive questions. The clinician then receives a condensed account and tests, corrects or contextualizes the AI’s suggestions.
The consultation therefore becomes:
patient–AI–physician, rather than simply patient–physician.
Wachter does not yet know where authority will finally reside in this triangle. But he thinks the division of labour will be radically renegotiated.
4. Patient-facing AI cannot simply be the doctor’s AI handed to the patient
One of Wachter’s more perceptive points is that clinical expertise lies partly in knowing which facts matter.
A patient may have fifty symptoms, experiences, medications and historical details, but may not know which five belong in the prompt. Nor can a patient necessarily recognize that three plausible-looking diagnoses are sensible while the fourth is absurd.
Therefore, generic chatbots are not adequate patient tools merely because they can answer medical questions. Proper patient-facing systems must behave more like clinicians:
- ask follow-up questions;
- elicit missing red flags;
- distinguish urgent from routine problems;
- indicate uncertainty;
- know when to escalate to a human.
Wachter consequently argues that AI for professionals and AI for laypersons must be designed differently. Current patients are already using general-purpose models, sometimes very successfully, but sometimes without the knowledge needed to judge the answer.
5. The deepest safety problem is not an obviously stupid AI
An AI that is wrong half the time would be ignored. An AI that is right every time could simply be trusted.
The truly difficult system is one that is:
right often enough to become indispensable, but wrong often enough to remain dangerous.
That is the world Wachter expects.
After a clinician has checked an AI-generated note forty-nine times and found it accurate, the fiftieth error may glide past unnoticed. Humans are poor at sustaining vigilance over a highly reliable automated system. This is the classic problem of automation bias, but generative AI makes it more powerful because its prose sounds intelligent, fluent and human.
Wachter puts the danger particularly well: previous machines were clearly a what; generative AI is the first “what” that behaves like a who. We assign trust partly on the basis of fluency, responsiveness and apparent understanding. AI possesses these cues without necessarily possessing dependable judgment.
So the risk is not only hallucination. It is:
the human tendency to fall asleep beside a machine that has earned almost—but not entirely—our trust.
6. “Human in the loop” may be less stable than it sounds
Medical organizations frequently reassure themselves that safety is preserved because a doctor remains “in the loop.” Wachter is sceptical about how meaningful this will remain.
When AI performs most of the cognitive work, the nominally supervising physician may:
- approve rather than independently reason;
- lose diagnostic skill;
- cease generating an initial formulation;
- become unable to detect the unusual error;
- carry legal responsibility without exercising genuine control.
He suggests one possible safeguard: trainees and clinicians might be required to make their own provisional judgment before seeing the AI’s answer. Otherwise education becomes the teaching of how to endorse or edit machine output rather than how to think clinically.
But he is candid that there is no easy psychological solution to automation bias. Ultimately, he suspects that one AI may have to audit another AI, rather than relying indefinitely upon uninterrupted human vigilance.
This is an important admission. It means that AI will not simply be a tool added beneath medical authority. It may reorganize where competence and oversight actually reside.
7. Medical expertise will change rather than simply disappear
Wachter does not predict the immediate elimination of physicians. Instead, the value of different medical abilities will change.
Memorizing facts and rapidly retrieving standard guidelines will become less distinctive. Greater value may attach to:
- framing the problem correctly;
- recognizing when the data are incomplete;
- judging whether the AI’s answer fits this particular patient;
- managing ambiguity and competing values;
- explaining uncertainty;
- assuming responsibility;
- negotiating among several medically reasonable choices.
Some specialties and tasks will be more affected than others, but he regards simplistic predictions such as the disappearance of radiologists as having underestimated the institutional, relational and workflow complexity of medicine.
In other words, medicine moves from knowing the answer toward:
knowing whether this answer belongs to this person, at this moment.
8. AI may expose the weakness of the existing healthcare system—or magnify it
Wachter’s optimism is conditional. New technologies are usually inserted into existing incentives, workflows and power structures. That is what happened with electronic records: a technology intended to support care became entangled with billing, regulatory documentation and institutional surveillance.
AI could similarly be used to:
- improve care;
- increase throughput;
- generate more documentation;
- deny claims more efficiently;
- intensify monitoring of clinicians;
- widen disparities between well-resourced and poor systems.
The technology alone does not determine the outcome. Governance, reimbursement, liability, workflow redesign and organizational culture do. He therefore argues that healthcare must not merely insert AI into present arrangements; it must reconsider why those arrangements exist.
This may be the lesson he draws from the electronic-health-record disaster:
A general-purpose technology produces its real benefits only when institutions reorganize themselves around its new possibilities.
9. What happens to empathy?
Wachter does not simply assert that empathy will remain uniquely human. He is more intellectually honest—and more unsettling.
AI can already produce responses that patients rate as more empathic than hurried clinicians’ replies. It has time, patience and access to an inexhaustible repertoire of validating language. He therefore asks whether human empathy will retain its superiority when the machine is more available, more articulate and less irritable.
His answer remains open. Machines may simulate many behavioural signs of empathy successfully enough that patients derive genuine comfort from them. Yet clinical care still involves embodiment, responsibility, relationship and the capacity to act in the world.
I would formulate the unresolved distinction this way:
AI may become excellent at the language of care.
It remains uncertain whether that amounts to caring, or whether the distinction will continue to matter to patients.
My reading of Wachter
Wachter is neither saying “AI will replace doctors” nor merely “AI will assist doctors.”
His stronger claim is:
AI will dissolve the present definition of what a doctor is.
The doctor has traditionally combined several functions:
- recorder;
- information retriever;
- diagnostician;
- interpreter;
- adviser;
- witness;
- responsible agent.
AI can take over large portions of the first five. What may remain especially human is not simply warmth or bedside manner, but situated responsibility: someone has to say, “Given this person’s history, wishes, fragility and circumstances, this is what I recommend—and I will remain here for what follows.”
From the perspective of our recent discussion of ethics, Wachter’s account stops just before the decisive question. AI may see more data, remember more possibilities and produce more rational recommendations. But can it be the subject of an ethical decision? Can it regret the person not saved, bear responsibility for an omission, or remain haunted by the road not taken?
Medicine may therefore divide into two layers:
AI performs increasingly much of the epistemology—what is probably happening.
The clinician remains responsible for the ethics—what shall we do, and who will answer for it.
The danger is that healthcare institutions will retain the physician merely as the legal bearer of responsibility after much of the actual judgment has migrated to the machine. That, I think, is the darker implication inside Wachter’s generally hopeful book.
