Skip to main content

Nate Gross, VP Health OpenAI

Grand Rounds, Oct 07 2026. Summary by ChatGPT 5.5.


The transcript specifically supports Gross’s three-part framing of patient, professional, and science applications; his criticism of exam-style benchmarks; and the need to test models in realistic clinical scenarios. His later sections develop the workflow/RAG argument, patient-context and proactive-agent vision, and concerns around liability, trust, physician disintermediation, and de-skilling.


Core thesis​

Gross’s central argument was that AI may be unusual among medical technologies because it has the potential to restore proximity rather than create distance between clinicians and patients.

For roughly two centuries, new technologies have often inserted themselves between doctor and patient: first instruments such as the stethoscope, and eventually computers and EHRs that can leave physicians sitting across the room looking at a screen. Gross sees AI as potentially reversing that trend by taking work away from clinicians rather than adding another interface.

The goal, in his framing, is not simply to make medicine more automated. It is to give people back time and attention: more time with patients, more time thinking carefully about difficult cases, and potentially more time outside work.

OpenAI’s health work is organized around three broad areas:

  1. Patients
  2. Health professionals and health-care organizations
  3. Science and biomedical research

Across all three, the challenge is moving from AI that can answer questions to AI that can perform useful work safely in real-world environments.


From “smart models” to useful medical AI​

Gross emphasized that conventional medical benchmarks are inadequate.

A model scoring extremely well on a medical licensing exam does not necessarily mean it is good at medicine. Multiple-choice questions may resemble material in the model’s training data, and actual clinical care requires much more than selecting the correct diagnosis.

OpenAI therefore evaluates models using simulated clinical encounters resembling OSCEs. Physicians construct scenarios and detailed scoring rubrics covering not only diagnosis but considerations such as:

  • dangerous diagnoses that should not be missed
  • social determinants of health
  • appropriate communication
  • completeness of reasoning
  • safety
  • contextual factors affecting care

The ideal is something closer to asking:

What would an excellent, fully attentive physician with unlimited time think through in this situation?

Gross argued that useful benchmarks should remain unsaturated: instead of celebrating movement from 97% to 98%, researchers should identify capabilities where models perform poorly and try to move them substantially upward.

He also described an important shift in evaluation from knowledge toward getting meaningful work done. Knowing the answer is different from knowing how to execute the sequence of actions required to solve a real problem.


Opportunities​

1. Giving clinicians time and attention back​

Perhaps the largest opportunity is using AI to remove low-value cognitive and administrative work.

Examples include:

  • documentation and ambient scribing
  • preparing for patient visits
  • synthesizing records
  • answering clinical questions
  • prior-authorization and administrative correspondence
  • coordinating schedules
  • finding relevant protocols
  • organizing patient information
  • assisting with routine decisions

Gross’s desired endpoint is clinicians practicing closer to the top of their license, rather than spending large portions of their day documenting, searching, clicking, or rubber-stamping routine processes.

The important question is therefore not simply, “Can AI do this task?”

It is:

If AI does this task, what does the clinician get to spend their time doing instead?


2. Clinicians can become builders​

Gross argued that physicians and other domain experts no longer need to leave medicine and become computer scientists to build useful technology.

Modern AI tools substantially reduce the technical barrier to creating prototypes. A clinician who deeply understands a workflow problem can increasingly build a rough system themselves, demonstrate its value, and then work with engineers to make it safe and production-ready.

This could change health innovation from:

clinician has idea → committee → requirements → engineering → pilot

to something closer to:

clinician has problem → builds prototype → demonstrates value → develops it properly

The prototype itself should not immediately be used on patients. But rapid prototyping gives clinicians far more agency in defining what eventually gets built.


3. Workflow-aware clinical decision support​

Gross described work with Penda Health in Kenya where an AI copilot watched information entered into the EHR and produced green, yellow, or red alerts.

The important lesson was not simply that the model could detect possible problems. The product had to be redesigned repeatedly before clinicians actually used it.

Eventually, the intervention was placed at the right point in the workflow and made disruptive enough to attract attention without becoming another ignored alarm.

Gross reported that the deployed system was associated with a roughly mid-teens percentage reduction in diagnostic and treatment errors.

The broader opportunity is therefore behavior-changing AI embedded at the right moment, rather than another dashboard or recommendation feed.


4. Retrieval-augmented clinical AI​

Because models may have outdated knowledge and can hallucinate, Gross emphasized retrieval-augmented generation (RAG).

Instead of asking the model to answer entirely from what it learned during training, the system can retrieve relevant information first and ground its response in it.

Possible sources include:

  • current journal articles
  • clinical guidelines
  • institutional protocols
  • local care pathways
  • electronic health records
  • laboratory results
  • previous clinical notes

This allows a model to say not only:

“Here is my answer.”

but:

“Here is my answer, and here is the guideline, lab result, or clinical note it came from.”

That creates opportunities for both greater accuracy and easier fact-checking.


5. Localizing AI to the actual care environment​

The “correct” medical recommendation can depend on local circumstances.

A recommendation for an MRI may be very different if the nearest MRI is 200 miles away. Hospitals may also have their own treatment protocols, referral pathways, formulary rules, and operational constraints.

RAG therefore creates the possibility of combining general medical knowledge with:

  • institutional protocols
  • geographic constraints
  • available services
  • patient-specific information
  • organizational permissions

AI could become less like a generic medical encyclopedia and more like a system that understands how medicine is actually practiced here.


6. AI grounded in the patient’s full context​

Gross described health AI progressing from generic answers toward answers grounded in information such as:

  • medical records
  • physician notes
  • laboratory results
  • wearable data
  • prior conversations
  • potentially, eventually, molecular or genomic data

The argument is that context changes the quality of an answer enormously. Advice for a healthy 20-year-old should not be identical to advice for a 70-year-old with multiple chronic conditions.

This shifts AI from general health information toward something much more individualized.


7. Patient preparation before visits​

AI may improve the small amount of time patients and clinicians actually spend together.

Before an appointment, it could help patients identify:

  • symptoms worth mentioning
  • changes detected through wearable data
  • important questions
  • information they have not yet told their doctor
  • the two or three issues that matter most

This is especially useful for the opposite problem: patients who arrive with many concerns and struggle to prioritize them within a short visit.

Rather than replacing the visit, AI could make the visit substantially better prepared.


8. Support between visits​

Gross argued that health care currently focuses heavily on the brief moments when patients interact with clinicians while largely ignoring the rest of their lives.

AI could provide reinforcement during the other days of the year.

Examples include:

  • explaining terminology from a visit
  • reviewing what the doctor recommended
  • helping interpret dietary advice
  • connecting everyday choices to treatment goals
  • helping patients remember or follow care plans

A patient could, for example, photograph a restaurant menu and ask which options best fit dietary guidance discussed with their clinician.

In this framing, consumer AI could potentially increase adherence to professional care rather than compete with it.


9. Health-care logistics and navigation​

A major opportunity lies outside diagnosis.

Seriously ill patients frequently become their own care coordinators: arranging appointments, communicating between specialists, tracking prescriptions, arranging transportation, handling insurance, and managing family obligations.

AI agents could potentially help coordinate these tasks.

Gross described possibilities such as:

  • arranging transportation
  • identifying an affordable pharmacy along someone’s existing travel route
  • checking whether a prescription has been picked up
  • coordinating multiple clinicians
  • helping patients navigate clinical trials
  • connecting information across systems

Much of the value of health AI may therefore come from solving navigation failures rather than diagnostic failures.


10. Proactive health AI​

Gross described an evolution through roughly three stages:

Context​

AI knows enough about the patient to provide a relevant answer.

Action​

AI can actually help accomplish tasks rather than merely explain them.

Proactivity​

AI recognizes something that needs attention and initiates the interaction itself.

This last phase could be especially important for prevention.

Rather than sending another generic reminder saying, “You are due for a colonoscopy,” an AI system might recognize that screening has not occurred and help determine why:

  • fear
  • cost
  • access
  • uncertainty
  • scheduling
  • transportation
  • misunderstanding

It could then help address the specific barrier.

That transforms reminders from notifications into interventions.


11. AI as an orchestrator of scientific tools​

Gross does not expect one general model to become the best tool for every scientific task.

Instead, general models may increasingly act as orchestrators.

For example, an AI system might:

  1. understand the scientist’s goal
  2. retrieve the relevant data
  3. determine which specialized tools are needed
  4. operate those tools
  5. collect the outputs
  6. move the workflow to its next stage

A specialized system such as AlphaFold may remain better at its particular task. The general model’s role would be to know when and how to use it.

This could shorten the cycle between hypothesis generation, analysis, experimentation, and interpretation.


12. Computer use could unlock legacy health systems​

Many health-care systems lack modern APIs.

Rather than waiting for every hospital system to become interoperable, AI may increasingly interact with software through the same graphical interfaces humans use.

A model could potentially:

  • inspect a scheduling interface
  • understand what is on screen
  • click and type
  • rearrange appointments
  • move information among systems

Paradoxically, this could extend the useful life of old technology.

Gross even suggested that AI could make the fax machine somewhat more useful: although faxed information is unstructured and difficult for conventional software to process, multimodal AI can increasingly read and interpret it.


13. Lower-cost intelligence​

Gross emphasized not only making models better but making comparable intelligence dramatically cheaper.

This matters because AI that works in a demonstration but costs too much to operate continuously will not transform health care.

The real target is therefore:

higher capability + lower cost

Ideally, AI would eventually become deflationary for health-care spending rather than simply adding another expensive technology layer.


14. Personalized medical education​

AI creates an opportunity not merely to prevent clinician de-skilling but potentially to improve clinician knowledge continuously.

Instead of requiring someone to deliberately enter “study mode,” a system could recognize where a clinician has a knowledge gap.

For concepts the clinician already understands, it could stay fast and concise.

For concepts where the clinician seems uncertain, it could slow down and explain or teach.

That creates the possibility of continuous, personalized medical education embedded in work itself.


Pitfalls and Open Questions​

1. Benchmark performance is not clinical competence​

Excellent performance on medical exams or benchmark questions can produce false confidence.

Real medicine involves:

  • incomplete information
  • competing priorities
  • workflow constraints
  • unusual patients
  • communication
  • uncertainty
  • follow-up
  • consequences over time

A model that performs impressively on a benchmark may still fail when deployed.

Gross repeatedly returned to the idea:

Everybody has a plan until they get punched in the face.

Deployment exposes problems that laboratory evaluation misses.


2. Human + AI is not automatically better than AI​

A common assumption is that combining clinicians and AI must outperform either alone.

Gross presented evaluations in which adding a physician eventually produced little improvement and, by the benchmark’s scoring criteria, sometimes made the answer worse.

That does not imply that physicians are unnecessary. The human may be adding contextualization, regionalization, relationship knowledge, or preferences that the benchmark fails to capture.

But it does challenge the simplistic assumption that:

AI + human = automatically superior system

Human-AI collaboration itself must be designed and evaluated.


3. Health-care AI changes faster than conventional evidence systems​

Gross identified a major structural problem for clinical research and regulation.

A conventional process might involve:

  1. studying a model
  2. conducting a trial
  3. writing a paper
  4. peer review
  5. publication
  6. regulatory assessment
  7. integration into guidelines

That can take years.

Meanwhile, the model being studied may be replaced within weeks or months.

This creates a serious mismatch between the speed of AI improvement and the speed of medical evidence, publication, regulation, and guideline development.

Existing evaluation systems may therefore need substantial redesign.


4. Workflow integration can determine success or failure​

A technically impressive system can fail because it asks clinicians to:

  • open another application
  • perform another step
  • interpret another alert
  • leave their existing workflow

Health-care software generally needs to save time rather than create more work.

But embedding everything directly inside the EHR creates another problem: it may constrain what the AI can do.

This creates a central design tension:

When should AI flow inside the existing workflow, and when is it worth creating a new interaction model?

Gross described this as one of the major design decisions facing health AI.


5. Alert fatigue​

An AI system that constantly interrupts clinicians will rapidly become useless.

A recommendation must occur:

  • at the right moment
  • with the right level of urgency
  • at the right frequency
  • with enough evidence to justify interruption

Otherwise AI becomes another ICU alarm: technically informative but functionally ignored.


6. Model knowledge goes stale​

Training sophisticated models takes time.

As a result, a model’s internal knowledge can lag current medical science by months or longer.

Medicine itself changes quickly.

This is one reason RAG, current guidelines, institutional knowledge bases, and explicit citations are so important.


7. Hallucination and uncertainty​

A particularly dangerous failure mode is not simply being wrong but being confidently wrong.

Gross said considerable effort has gone into making models recognize uncertainty and request additional information rather than forcing an answer.

In health care, a good system must sometimes say:

  • I need another laboratory value.
  • I need more history.
  • There is not enough information.
  • You should speak with a clinician.
  • This may require urgent evaluation.

Knowing when not to answer is itself a clinical capability.


8. Escalation is difficult​

Patient-facing AI needs to recognize when a conversation should move to a human clinician or emergency service.

Some situations are obvious, such as classic symptoms of a medical emergency.

Others are much subtler, particularly in mental health.

Over-escalation is also harmful. If every uncertain situation results in “go to the emergency department,” AI can overwhelm health systems and lose usefulness.

The challenge is therefore appropriate escalation, not maximal escalation.


9. Privacy becomes more consequential as personalization improves​

The more useful an AI health assistant becomes, the more intimate information it may contain:

  • medical records
  • medications
  • laboratory results
  • wearable data
  • family information
  • behavioral patterns
  • conversations that patients have not shared with anyone else

That creates an uncomfortable relationship:

The same context that makes AI more useful also makes failures of privacy or trust more consequential.

Gross argued that health conversations should receive heightened privacy protections and should not be treated like ordinary conversational data.


10. Trust cannot depend entirely on the company building the model​

Gross argued that developers should not be the only organizations defining whether their own models are safe.

Independent evaluation will be important.

Potential participants include:

  • universities
  • regulators
  • independent research organizations
  • professional societies
  • competing AI labs

The broader principle is that trust requires external scrutiny, not simply assurances from the company deploying the technology.


11. Liability is unresolved​

Current clinical systems generally leave responsibility with the clinician.

That arrangement becomes increasingly awkward as AI becomes more autonomous.

If an AI system independently:

  • renews a medication
  • initiates therapy
  • adjusts a dose
  • triages a patient
  • misses a dangerous diagnosis

who is responsible?

Possibilities include:

  • the clinician
  • health system
  • AI vendor
  • software developer
  • insurer
  • some entirely new liability structure

Gross expects new insurance and liability mechanisms to emerge, but this remains unresolved.


12. The safety standard for AI is unclear​

Humans make medical errors every day.

AI will also make errors.

But society is unlikely to tolerate AI errors at the same rate it tolerates human errors.

That raises a difficult question:

What level of safety should an AI system have to achieve before it is acceptable?

Human-level performance may be too low.

Perfect performance may be impossible.

Determining the appropriate standard will be a major regulatory and ethical challenge.


13. Automation could restore the doctor-patient relationship—or destroy it​

The talk began with the possibility that AI could restore proximity between clinicians and patients.

But the opposite outcome is plausible.

Financial incentives could encourage health systems to automate increasingly large portions of care and reduce physician interaction.

Instead of:

AI handles administrative work so doctors can spend more time with patients

the system could become:

AI handles most care so organizations need fewer minutes of physician time.

Gross acknowledged that this is largely uncharted territory.

The outcome will depend heavily on incentives and system design.


14. Physicians could become “rubber stamps”​

One particularly undesirable future is one in which an AI system generates the diagnosis and treatment plan and the clinician’s primary role becomes signing it.

Gross rejected that as the desirable model.

He expects valuable physician work to shift toward areas such as:

  • navigation
  • ethics
  • relationships
  • judgment
  • contextual interpretation
  • handling meaningful variability
  • generating new ideas

The important question is not simply which tasks AI can perform, but which human capabilities medicine must deliberately preserve.


15. De-skilling​

If AI performs a task repeatedly, humans may gradually lose their ability to perform it themselves.

That creates difficult trade-offs.

Some de-skilling may be perfectly rational. Clinicians do not need to preserve every administrative skill that technology can reliably perform.

But clinical reasoning skills are different.

Questions include:

  • Which capabilities must clinicians retain?
  • When should trainees be permitted to use AI?
  • How do we know whether AI is assisting thought or replacing it?
  • Should some examinations remain AI-free?
  • Can AI itself detect knowledge gaps and actively teach the clinician?

Gross suggested that the answer may involve using AI as both assistant and teacher.


16. Automation may simply accelerate both sides of dysfunctional systems​

Gross used prior authorization as an example.

Insurers can use AI to process claims and denials more efficiently.

Hospitals can use AI to generate appeals more efficiently.

Both sides become faster, but the underlying system remains unchanged.

That creates an AI arms race rather than meaningful efficiency.

The larger opportunity would be redesigning the interaction itself—for example, creating a trusted shared arbiter or new protocol through which both sides can reach decisions with less administrative conflict.


17. More personalization means greater dependence​

If an AI system knows:

  • your medical history
  • your calendar
  • your medications
  • your clinicians
  • your habits
  • your location and transportation constraints
  • your preferences
  • your previous conversations

it can become enormously helpful.

It can also become infrastructural.

That raises broader questions about:

  • dependence
  • portability
  • vendor power
  • data control
  • long-term trust
  • what happens when the system is wrong

The usefulness and the risk grow together.


The likely evolution of health AI​

Gross’s vision can be summarized as a progression:

Knowledge → Context → Action → Proactivity

Knowledge​

AI can answer medical questions.

Context​

AI can answer those questions using the patient’s records, local protocols, current literature, and circumstances.

Action​

AI can help carry out tasks: scheduling, coordination, medication logistics, trial matching, communication, and administrative work.

Proactivity​

AI can notice something before the user asks and intervene appropriately.

The largest transformation may come from the final two stages.

A chatbot that answers medical questions is useful.

An agent that understands a patient’s circumstances, recognizes a problem, determines what needs to happen, coordinates the relevant systems, and knows when a human needs to intervene is a fundamentally different kind of technology.


Central design principle​

The most important question in health AI may not be:

What can the model do?

but:

Where should the model act, when should it act, what context should it have, and when should it get out of the way?

Gross repeatedly emphasized that technical capability alone will not determine whether AI improves medicine.

The outcome will depend on:

  • workflow design
  • evaluation
  • cost
  • privacy
  • trust
  • liability
  • appropriate human oversight
  • escalation
  • incentives
  • whether clinicians remain in meaningful control

Bottom line​

Gross presented a future in which AI becomes less of a chatbot and more of an intelligent layer surrounding health care.

It could:

  • prepare patients and clinicians for encounters
  • retrieve and synthesize medical information
  • reduce administrative work
  • coordinate care
  • support decisions
  • operate existing software
  • connect specialized scientific tools
  • personalize education
  • help patients between appointments
  • eventually identify and act on problems proactively

But the same capabilities create serious risks.

AI could reduce clinician burden—or enable organizations to reduce clinician involvement.
It could strengthen clinical expertise—or cause de-skilling.
It could personalize care—or create unprecedented concentrations of sensitive information.
It could reduce administrative waste—or simply automate an existing arms race.
It could restore the doctor-patient relationship—or insert an even larger technological intermediary between them.

The technology is advancing faster than the institutions responsible for evaluating, regulating, and integrating it.

Gross’s strongest implicit message was therefore that health professionals need to participate in designing these systems now. The key decisions are increasingly not about whether AI will enter health care, but about what roles it should occupy, what should remain human, and whose interests the resulting systems will serve.