Talks at GoogleDr. Robert Wachter | A Giant Leap: How AI Is Transforming Healthcare... | Talks at Google
Talks at Google
Introducing Bob Wachter 0:06
Michael Howell, Google's chief health officer, introduces Bob Wachter, a professor and chair of the Department of Medicine at UCSF, who coined the term hospitalist and has written five books. The conversation centers on his newest book, A Giant Leap: How AI Is Transforming Health, though his earlier book, The Digital Doctor, also comes up for comparison.
A system in need of transformation 2:31
Wachter reads a passage from his book describing the US health system as caring and well trained on the surface but bogged down in red tape, calling it a Kafkaesque nightmare in desperate need of transformation. He explains that the day before ChatGPT launched, nobody thought Google Search was flawless, yet nobody thinks the health care system is fine either. He argues that accepting the current system's flaws as normal is not a neutral choice but an actively negative one.
The numbers behind the crisis 5:01
Wachter lays out the statistics that justify his alarm. Evidence based therapy is delivered only about half the time, and it takes an average of seventeen years for proven treatments to become standard practice. Medical mistakes kill between fifty thousand and possibly a couple hundred thousand people a year, comparable to a large airplane crashing every day. Access to primary care and mental health is poor even for the well insured, equity gaps mean the poor and minorities live decades less than the wealthy and white, and health care costs about six trillion dollars a year, roughly twenty percent of GDP, with about a third of that spent on administrative paperwork rather than patient care.
Miracles and mundane failures 8:31
Wachter notes the paradox that the same system capable of curing HIV complications and delivering lung transplants that patients would travel the world for still fails at basic access, like getting a primary care appointment. He points out that Americans spend more on health care than on food or housing, and that cities and counties are laying off teachers to cover health costs, which frames why he sees such large room for improvement.
From paper charts to electronic records 9:32
Wachter recounts how American hospitals barely digitized before 2010, with fewer than one in ten having electronic health records in 2008. Before that, doctors scribbled orders on paper, used little colored wheels on charts to signal nurses, sent triplicate carbon paper forms through pneumatic tubes or fax machines, and relied on a single physical chart binder. Digitization only happened because of thirty billion dollars in federal incentive payments tied to 2008 stimulus money, not because hospitals chose it themselves.
Why The Digital Doctor was grumpy 11:30
Wachter explains that his earlier book, The Digital Doctor, was pessimistic because the shift to electronic records, despite promising to end illegible handwriting and give everyone access to the same chart, left nearly everyone he interviewed hating their system. He came to believe the electronic record was only the necessary foundation, not the solution, since most patient information remained unstructured text that computers could not really use until large language models and generative AI arrived.
Turning cautiously optimistic 17:31
Writing the new book, Wachter again interviewed around 110 people, and this time the picture shifted. He describes realizing how generative AI could finally deliver real computerized decision support and let doctors look patients in the eye instead of typing into a chart, restoring some of the humanism of medicine that electronic records had stripped away.
Lessons from NYU on training 18:30
Wachter describes visiting NYU, which he found had thought most deeply about combining AI with medical education, alongside Mayo Clinic's broader approach. He raises a governance problem specific to academic medical centers, which uniquely combine research, education, and patient care under one roof, so decisions to deploy AI tools for patient benefit rarely consider whether trainees should learn core skills, like writing a note or reviewing a chart, before relying on AI scribes or summarizing tools.
The human in the loop problem 21:00
Wachter frames a central tension of the book: if AI were only right half the time it would be worthless, and if it were right all the time a human in the loop would be pointless, but for now it sits in between, right often enough to help and wrong often enough that a human must remain the final check, especially in life or death settings like the ICU. He notes the difficulty that humans are poor at staying vigilant once a tool has proven reliable many times in a row, since trust erodes attention.
How much should trainees still learn 22:00
Wachter argues it would be a mistake to strip too much cognitive work from medical training, since expert judgment about which of two hundred patient data points matter, and how to evaluate an AI's response, still depends on deep clinical knowledge. He mentions a Macy Foundation conference on medical education where the only consensus item safe to remove from the curriculum was the Krebs cycle, the biochemistry pathway students dread memorizing.
Experts versus novices with AI tools 24:00
Drawing on a conversation with Peter Lee of Microsoft, Wachter explains that expert and novice use of AI tools differs fundamentally, since experts know what information to input and how to interpret the output. He cites an Oxford study where patients given a scripted worst headache scenario, a classic sign of subarachnoid hemorrhage, failed to phrase it correctly to a chatbot and received wrong, reassuring advice, showing how much interpretive expertise still matters even when the underlying AI tool works correctly.
Skills that fade with new tools 26:06
Wachter closes this segment by comparing today's debates to past shifts in medical training, noting that doctors once needed to perform emergency open chest heart massage, a skill few use now, and that physical exam skills have declined as portable ultrasound and imaging take over roles once filled by listening to the lungs with a stethoscope. He observes that every generation laments the loss of old skills as new tools take their place.
Deskilling and its tensions 27:01
Wachter admits he has deskilled on things like reading maps and remembering phone numbers, and says that is mostly fine. The real tension for medical trainees is that some deskilling matters, because a young doctor might end up somewhere without the tools they leaned on, so society has to discover over time which skills can safely fade and which cannot.
Where AI already helps 28:02
Wachter uses an AI scribe on rounds that lets him look patients in the eye instead of typing, and it also helps him digest a seven hundred page chart he could never read in a three minute visit. He and colleagues rely heavily on OpenEvidence, a tool built like GPT or Gemini but trained on medical literature, journals, and specialty guidelines rather than the open internet. He compares it to a curbside consult, the quick hallway question doctors ask a specialist, saying it now lets him type a messy real case, such as a seventy two year old with CLL and a fever, and get a trustworthy, sourced answer in a way older tools like UpToDate never allowed.
Trust, jobs, and judgment 33:00
He worries the human-AI partnership looks safer in clean studies than in messy real clinics, and that humans may either over-trust or under-trust the tool, especially since people tend to trust a who more than a what. On jobs, he expects clinician roles to survive for now because unmet need is huge, pointing to Geoffrey Hinton's 2016 prediction that radiology would vanish, which instead saw applications crash then rebound once students realized the field was more complex than sorting images. His deeper worry is that as AI gets embedded in records to suggest diagnoses and treatments, doctors may stop exercising the judgment that medicine's cost, risk, and ethical tradeoffs require. He also flags patients using AI themselves as net positive for access, though tools must ask careful questions first, and doctor-patient trust will need renegotiating when patients arrive with answers pulled from searches. He closes by citing Mayo Clinic's CEO, who said the risk of moving too slowly outweighs the risk of moving too fast, since the status quo itself is unacceptable.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
