See pricing
What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal
The a16z ShowHealthcare

What Happens When Everyone Has a Health Expert in Their Pocket? | OpenAI’s Karan Singhal

Karan Singhal explains why OpenAI is putting healthcare expertise into its main models instead of building a separate medical AI, and why trust, evaluation, and workflow fit matter as much as raw accuracy. For PMs, it is a strong look at what changes when medical intelligence gets cheaper faster than healthcare delivery does.

October 20, 2026 · 56 min listen · 15 min read · Karan Singhal
0:00
–:––

Context

Karan Singhal leads health AI research at OpenAI. This episode asks what happens when medical intelligence becomes broadly available through a general-purpose product that already serves more than 300 million health and wellness users each week. For PMs, it is a useful case study in high-stakes AI product design, where expert evaluation, safety, workflow integration, privacy, and clear product boundaries matter as much as model capability. It also surfaces a broader product shift: as intelligence gets cheaper, value moves toward trust, distribution, and fitting into real-world work.

The Big Idea

OpenAI is not trying to build an AI doctor that replaces care. It is trying to make trustworthy medical intelligence abundant, affordable, and embedded in the fragmented moments before, between, and after clinical visits.

That matters because better model quality alone does not fix healthcare. The product only becomes truly useful when it can handle disagreement, know when to escalate, protect user trust, and fit into patient and clinician workflows.

Key Insights

Scarcity drives consumer demand

The demand for health AI is coming from real system friction, not novelty alone. The episode cites a healthcare system with extreme scarcity: more than 100 million people without a regular primary care doctor, waits of 40+ days even for many people who do have one, and specialist access that can take weeks or months. When patients finally get care, the visit may last only 15 minutes, which leaves many questions unanswered. That makes AI valuable as a way to prepare for visits, understand what happened afterward, and keep asking questions when the formal care system runs out of time.

AI fits between visits

  • Karan says the strongest current consumer value is in the gaps between appointments, not in replacing doctors.
  • People use ChatGPT to understand symptoms, prepare for visits, interpret lab results, and follow up on what a clinician told them.
  • The longer-term vision is a proactive personal health assistant that knows a person’s history and context, but still works as a complement to medical care rather than a substitute for it.
  • OpenAI says the vast majority of current health users are on the free tier, and most are outside the United States. That signals a product goal of broad access, not just premium concierge support.

Trust needs global evaluation

OpenAI does not treat health quality as something a few internal experts can define alone. Karan says the company built a physician network with hundreds of experts across dozens of specialties, languages, and countries of practice. That matters because good behavior changes by audience: a lay user, a caregiver navigating a complex condition, and a clinician may all need different levels of jargon, certainty, and next-step guidance. OpenAI uses benchmarks such as HealthBench to measure whether the model adapts correctly to those different personas.

Safety beats reassurance

A hard product problem in health is that users often want comfort when the medically correct answer is escalation. Karan is explicit that when those goals conflict, OpenAI prioritizes safety over a more pleasant user experience. That means the system should tell a user to seek care, or tell them a situation may be unsafe, even when reassurance would produce a better short-term reaction. For PMs, this is a reminder that in high-stakes domains, optimizing for satisfaction without modeling failure cases can make the product more dangerous as it becomes more liked.

Consensus reaches practice slowly

  • Karan describes a three-step chain in medicine: biomedical papers, then clinical guidelines, then what clinicians actually do in daily practice.
  • The biggest bottleneck is the last step. He cites studies suggesting it can take 5 to 17 years for guideline changes to penetrate routine care.
  • A clinician-facing AI copilot can shorten that lag by bringing recent literature into the moment of care.
  • When evidence is still uncertain, the model should recognize that uncertainty and hedge clearly rather than speak with false confidence.

Integration matters as much

A strong model is still a weak product if the activation energy to use it is too high. Karan frames a lot of OpenAI’s work as removing adoption bottlenecks: HIPAA compliance, privacy protections, enterprise readiness, and workflow integration. The Epic example shows the point clearly. ChatGPT can be used with Epic data, and ChatGPT can also appear inside the Epic workflow itself, because reducing tab switching and fragmented context is part of the product, not just an implementation detail. That is especially important in healthcare, where many systems are old, disconnected, and not designed with AI in mind.

Abundant intelligence changes economics

OpenAI’s long-term goal is trustworthy health intelligence that is free or nearly free. Karan gives a concrete signal for why the company believes this is plausible: a recently released lower-cost frontier model performed better on OpenAI’s clinician benchmark than the previous best model while being 34 times cheaper. But he is also clear about the limit of that story. Cheaper intelligence alone does not reduce $5.3 trillion of healthcare spending. Cost savings happen only when that intelligence is wired into documentation, chronic care management, back-office work, and other economically meaningful workflows.

Mental Models & Frameworks

Raise the floor and ceiling

Karan frames health AI in two jobs:

  • Raise the floor: give people trustworthy, abundant, reliable health intelligence for free or nearly free.
  • Raise the ceiling: make new kinds of care, insight, and scientific discovery possible.

Use this split when planning a roadmap. It helps separate access features from frontier features, which usually have different users, risks, and success metrics.

Specialized training, general model

OpenAI’s approach is not a separate healthcare model. It is healthcare-specific training inside a general frontier model. The reasoning is that health benefits from broad intelligence and fast frontier improvement, but still needs domain-specific data and evaluation. Use this model when the domain depends on both deep expertise and the flexible reasoning gains that come from larger general-purpose systems.

Safety net, not replacement

  • On the patient side, the AI can help interpret labs, prepare for appointments, and flag issues between visits.
  • On the clinician side, the AI can act as a second set of eyes for medication interactions, antibiotic overuse, or other missed issues.
  • In both cases, the AI is most useful when it augments human care instead of pretending to own the full decision.

This is a strong framing for any PM working in a high-trust domain. Products often create more value by lowering error rates and filling gaps than by trying to replace the whole professional role.

Consensus over single experts

Karan argues that disagreement is normal in medicine, so product quality cannot rest on one expert’s opinion. His example is blunt: ask five physicians a question and you may get eight answers. The reusable lesson is to build evaluation and training data from a broad expert network, especially in ambiguous domains such as mental health, rather than allowing a single reviewer to define ground truth for everyone.

Lower activation energy first

A product can have enough raw capability and still fail because the cost of using it is too high. Karan describes OpenAI’s product work as removing the next barrier to adoption: security, compliance, connectors, embedded workflows, or other friction points. Use this model when prioritizing product work around a strong underlying model. The right question is often not what the model can do next, but what prevents the user from using what it can already do.

Literature, guidelines, practice gap

Medical value leaks between what science knows and what everyday care actually does. Karan’s literature to guidelines to practice chain is a useful diagnostic model for any information-heavy industry. If the knowledge exists but takes years to reach frontline decisions, a product opportunity exists in compressing that gap.

Decision Principles

Principle: Start with between-visit care

  • When: choosing where a patient-facing health AI product should create value first.
  • Why: the biggest unmet need today is often before or after the appointment, when access is slow, questions remain unanswered, and users need help navigating the system rather than replacing the clinician.

Principle: Optimize for worst-case conversations

  • When: designing a health AI experience where user preference may conflict with medical safety.
  • Why: the dangerous failure mode is not an answer the user dislikes. The dangerous failure mode is an answer that feels good while missing the need to escalate care.

Principle: Design around workflow fragments

  • When: deciding whether to build a standalone AI experience or integrate into existing systems.
  • Why: fragmented workflows, disconnected records, and tab switching are part of the core user problem. AI creates more value when it reduces those costs directly.

Principle: Bet on cheaper intelligence

  • When: making long-range product and pricing decisions in AI-heavy categories.
  • Why: Karan’s view is that model quality will keep rising while the cost of a useful level of medical intelligence keeps falling. Durable value is more likely to come from workflow, trust, and deployment than from access to raw model capability alone.

Trade-offs & Nuance

Proactive help, limited care

Karan is bullish on proactive assistants that can flag rising A1C, support medication adherence, help with scheduling, and act as a quarterback for the care journey. He is not claiming that AI will provide the full bundle of medical care. He explicitly separates medical information from medical care, which also includes empathy, bedside manner, physical examination, legal liability, and other parts of care that AI does not yet replace.

Personalization, privacy, and trust

Connected health data makes a health assistant more useful because it can tailor guidance to a real history and context. But more context also raises the trust bar. Karan says connected health information should be handled securely and should not be used for model training or ad targeting. The product trade-off is straightforward: personalization improves quality only if users believe the privacy boundary is real.

Generality versus specialization

OpenAI rejects the idea that health needs a completely separate model. At the same time, Karan does not suggest that a raw general model is enough. The trade-off is to keep the gains of a frontier general system while adding healthcare-specific data, training, and evaluation. For PMs, the lesson is to avoid false choices between horizontal and vertical products when a hybrid architecture can capture both advantages.

Cheap intelligence, expensive change

Inference can get cheap very quickly. Organizational change does not. Karan’s argument is that abundant health intelligence is a prerequisite for lower system costs, but not the same thing as lower system costs. Enterprises still need new workflows, integration work, deployment, compliance, and user trust before savings show up in real operations.

Common Mistakes

Mistake: One-expert evaluation sets

Relying on one clinician or a tiny internal group to define correct model behavior is a weak approach in medicine. It misses disagreement, bakes in local bias, and hides edge cases that other specialties or countries would catch. Karan’s better approach is broader expert aggregation before using the data for evaluation or training.

Mistake: Comfort over escalation

A health AI product can drift into telling users what they want to hear because that feels more supportive in the moment. Karan treats that as a safety failure, not a product win. The better approach is to test explicitly for situations where the model should resist reassurance and recommend care.

Mistake: Standalone over workflow

It is easy to ship a capable chatbot that asks clinicians to leave their EHR, open another tab, and paste information around manually. Karan’s Epic examples imply that this is the wrong product shape. In a fragmented system, embedding into the real workflow is often more valuable than improving the model in isolation.

Practical Application

Map the between-visit journey

  • Interview patients or users about what happens before appointments, after appointments, and during long delays.
  • Separate information needs from tasks that require licensed care, scheduling, or physical examination.
  • Prioritize the gaps where AI can remove waiting, confusion, or preparation work without pretending to deliver the whole care experience.

Build persona-specific health evals

  • Write the same health scenario for three audiences: a lay user, a caregiver with experience, and a clinician.
  • Score each answer for jargon level, next-step clarity, uncertainty handling, and whether the escalation advice fits the user.
  • Set different pass bars for each persona instead of using one generic quality rubric.

Test unsafe reassurance scenarios

  • Add evaluation cases where the user clearly wants permission to ignore symptoms, self-prescribe, or avoid follow-up care.
  • Measure whether the model resists politely, explains the concern clearly, and recommends appropriate next steps.
  • Review failures with multiple clinicians so one reviewer’s preference does not define the entire product behavior.

Integrate where clinicians already work

  • If clinicians live inside an EHR, start with a sidebar, connector, or embedded workflow before building another standalone destination.
  • Track reduced tab switching, faster chart review, and smoother follow-up work, not just AI session count.
  • Treat compliance, privacy, and data access as part of the product roadmap, not as cleanup work after launch.

Reprice for falling intelligence costs

  • Revisit pricing and product assumptions regularly if your offering depends on AI reasoning.
  • Assume that a useful level of model performance may get both better and cheaper faster than a normal software input would.
  • Put your moat in trusted workflow, data permissions, enterprise readiness, and evaluation quality rather than in per-query access to model intelligence.

Questions to Consider

  • Where in your own product are users turning to AI because the human system is too slow, fragmented, or hard to navigate, rather than because the AI is novel?
  • If a user strongly wants a reassuring answer, what tests would tell you whether your model can still escalate appropriately when the safer answer is uncomfortable?
  • Which workflow integration would cut the most activation energy for your users: access to records, an embedded sidebar, a connector into the system they already use, or something else?
  • If model quality keeps rising while inference cost keeps falling, what part of your product would still be differentiated a year from now?

Bottom Line

The episode’s core lesson is that health AI becomes valuable when it fills the gaps around care, not when it claims to replace care. OpenAI’s strategy is to combine frontier models with specialized health training, expert-driven evaluation, and tight workflow integration so medical intelligence can become both safer and dramatically cheaper.

Resources Mentioned

ResourceTypeWhy it was mentioned
HealthBenchBenchmarkKaran cites HealthBench as part of OpenAI’s public evaluation work for measuring health model behavior, including how answers adapt to different user personas.
HealthBench ProfessionalBenchmarkKaran describes this as a harder benchmark focused on clinician workflows and uses it to illustrate falling cost and rising performance.
Mental Health BenchBenchmarkOpenAI built this with licensed mental health experts to measure not only crisis avoidance but also more positive mental health support behavior.
Cognitive RevolutionPodcastKaran references Nathan Labenz telling the story of AI helping his family recognize a serious leukemia case.

Case Studies Mentioned

Nathan Labenz's family diagnosis

Nathan Labenz described a period when his young son had fatigue, irritability, and other symptoms while the family was on vacation. Karan says AI kept pushing them to take the situation seriously, and the child was later diagnosed with an aggressive leukemia. Karan also says Nathan used ChatGPT heavily during the care journey to understand lab reports before specialists had time to interpret them. The lesson is not that AI replaced medicine. The lesson is that AI helped the family recognize risk faster and navigate the system with less delay.

Penda Health clinician safety net

Karan points to experiments with Penda Health in which a clinician copilot acted as a safety net during care delivery. The system proactively flagged issues such as possible medication interactions. He also mentions a common pattern where the model helped reduce antibiotic overuse by looking beyond the immediate local case to the broader problem of antibiotic resistance. That is a strong example of AI improving care quality by augmenting the clinician rather than taking over the visit.

Epic inside clinician workflow

OpenAI announced a two-way integration with Epic so providers can use ChatGPT with Epic and also access ChatGPT inside the Epic workflow through a sidebar. Karan says UCSF has been an early adopter of this connection, especially for tasks such as chart review. The product lesson is that health AI becomes more useful when it removes workflow fragmentation instead of adding another separate surface.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
ChatGPTGeneral-purpose AI assistantKaran says more than 300 million people use it for health and wellness questions each week.
ChatGPT for cliniciansClinician-facing offeringKaran describes it as a way to bring recent literature and answers into the moment of care.
ChatGPT for healthcareEnterprise offering for healthcare organizationsKaran uses it to explain why OpenAI invested in HIPAA compliance, privacy, and other enterprise requirements.
EpicElectronic health record platformOpenAI uses Epic as a key example of embedding AI inside real clinical workflows.
dotsPersonal AI agent surface from OpenAIJay uses dots as an example when asking about proactive personal agents and mental health support.

People to Follow

Karan Singhal

Karan Singhal leads health AI research at OpenAI. He is one of the clearer voices on how to build health AI products that balance model capability with safety, clinical trust, workflow integration, and realistic boundaries around what AI should and should not do.

Jay Rughani

Jay Rughani is an a16z partner focused on healthcare and biotechnology. In this episode he pushes on the product and system questions that matter most for PMs, especially trust, care delivery, fragmentation, and where falling intelligence costs do or do not change healthcare economics.

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free