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

1. Key Themes

Specialized Training Inside a General Model Beats Forking a Healthcare Model

OpenAI deliberately does not build a separate medical model. Singhal, who previously built specialized medical models, explains that those fall behind the frontier quickly, so OpenAI injects healthcare data into every training stage of the general model instead. "What we do for ChatGPT and all the models that we ship in ChatGPT, across every stage of model training for ChatGPT, we actually have specialized healthcare data added to each phase of training since all of our models in the last year." 00:07:23 His rationale: "If you kind of forked and said, 'Hey, we're going to train a kind of smaller specialized model that is just for healthcare,' like I used to and like many people have, what you'd find is that model would just kind of fall behind the frontier relatively quickly, because the pace of the frontier and the general model improvements is so fast." 00:07:50 For investors, this is a direct warning about thin vertical-model plays in health.

Health Intelligence Is Being Commoditized and Distributed for Free

OpenAI's explicit strategy is abundance, not scarcity. Singhal says the goal is to "give people access to trustworthy, abundant, reliable health intelligence for free or nearly free" 00:00:09, and notes "of the kind of 300 million weekly active health users, the vast majority of them are free users... and most of them are outside the US as well." 00:10:05 On cost curves: GPT-6 Luna "performed better than 5.6 Sol, which was literally our best model like a month ago... it performs better despite being 34 times cheaper." 00:44:26 He frames the bet: "the cost of health intelligence will plummet at a given level of health intelligence." 00:44:50 If the intelligence layer is commoditized, value migrates to integration, workflow, distribution, and trust.

The Real Bottleneck Is Knowledge Diffusion, Not Knowledge Creation

Singhal highlights that evidence-based guidelines take years to reach practice. "There are these studies on like, how long does it take for clinical guidelines to penetrate the practice of everyday doctors across the world? And in the past, people have found that this takes, depending on the study, like 5 to 17 years." 00:21:38 OpenAI's goal is to compress that "down to weeks, or days, or even hours" 00:22:09 by bringing the latest literature into the moment of care. He notes that for many questions "there is medical consensus, but it's not being practiced." 00:22:39

Trust Is Built Through Distributed Expert Consensus, Not a Few In-House Doctors

OpenAI built a global network of hundreds of physicians to avoid single-expert bias. "It turns out if you ask five physicians or something... any question, you'll get like eight different answers. And so... build this kind of global physician network of hundreds of experts that we work with kind of across the globe, across, you know, dozens of specialties and languages and countries of practice." 00:17:49 When user preference and clinical judgment conflict, the rule is explicit: "It's basically a prioritization of safety." 00:18:39 They also build shared industry benchmarks (HealthBench, HealthBench Professional, MentalHealthBench) to create "this shared ruler, or measuring stick." 00:29:12

OpenAI as the "Front Door" and Quarterback, With Partners Delivering Care

Singhal is clear about the boundary of OpenAI's ambition: "There's no future where, you know, any one company is delivering all the medical care, right? And so our role here is really to think about, what does the last mile of medical care really look like for people?... so that that last mile... just looks like the last 100 m for all of our partners." 00:31:25 OpenAI can "act as a quarterback for your health" 00:32:07 but will rely on "all of our startup and care provider organizations and partners." 00:33:07 This implies a large opportunity for companies that own the last-mile care delivery.

AI as a Safety Net for Both Patients and Clinicians

The framing is augmentation, not replacement. "I love this idea of a kind of safety net for both patients and clinicians that is just kind of there in the care journey... really thinking about the ways in which it can augment the existing experience." 00:11:20 The Penda Health clinician copilot study is the evidence: the AI "came in and proactively suggested like, 'Hey, this might be off... there might be this medication interaction that you might not be tracking.'" 00:11:50 Jay's anecdote from Greg Brockman, in which a doctor thanked a patient for catching a drug interaction she "didn't have enough time" to find 00:10:35, illustrates the same dynamic.

Removing "Activation Energy" and Integration Friction Is the Real Product Work

Singhal frames the product problem as reducing barriers: "You can have a model that's really good, but if the activation energy to use the model to improve health is too high, then people won't do it." 00:24:47 OpenAI invested roughly two years in HIPAA compliance and privacy, then integrations such as Epic: "you can use Epic in ChatGPT... and also for ChatGPT to be useful in Epic itself... a little sidebar." 00:27:06 He also notes "healthcare is also one of those industries where you shouldn't expect everything to have an MCP connector tomorrow" and that many run "primary EHR software... on like a VM," which "actually makes things easier in some ways for AI agents." 00:14:12

The Move 37 Thesis: Raising the Ceiling, Not Just the Floor

Beyond access, OpenAI wants superhuman medical insight. "One thing that's been motivating our team... is this notion of bringing to bear a kind of Move 37 for health." 00:38:39 Early evidence: in the Penda Health study, clinicians were "prone to overuse of antibiotics," and "a very common pattern that the models flagged is just kind of like avoiding this overuse of antibiotics." 00:40:07 The larger vision is models integrating "a longitudinal history of a user over 10 years across all their data... making predictions that seemed impossible before." 00:40:25

Agents Unlock Value in Dormant Healthcare Data

"One of the most interesting things about agents is they kind of solve a lot of the bottlenecks in analyzing data that already exists... EHR data, longitudinal histories, claims... it is kind of just stored there, and the insights from it are not fully mined." 00:41:33 He adds that "I'm also very optimistic about a future in which AI helps accelerate a lot of the bottlenecks in validation, which I think will end up being the bottlenecks, as opposed to kind of proposing new ideas or treatments." 00:42:13

Cost Compression Comes From Workflow Integration, Not Intelligence Alone

Singhal is notably restrained on macro healthcare spend: "I don't think all of that is about health intelligence. And I don't think like health intelligence has had yet a macro effect on the problem." 00:45:24 The savings are in "chronic care management, or documentation, or administration... back office workflows" and "I don't think that just happens from making health intelligence abundant. I think that happens from integration and workflows and deployment." 00:45:54

2. Contrarian Perspectives

Don't Build a Specialized Medical Model, Even Though the Instinct Says You Should

The conventional startup and investor assumption is that vertical domains need vertical models. Singhal, who built them, argues the opposite. "If you bet on scale, what we've found as OpenAI, as an industry, is that things will get better." 00:07:50 His practical advice to entrepreneurs: "Bet on the models getting better... many people in healthcare are on some kind of spectrum between betting against the models, betting neutrally for the models, and betting with the models... I would say, you know, bet on the models getting better." 00:48:44

Intelligence Abundance Alone Will Not Fix Healthcare Costs

Despite being the person building the cheapest medical intelligence ever, Singhal pushes back on the hype that cheap AI automatically collapses the $5.3 trillion spend. "I don't think like health intelligence has had yet a macro effect on the problem that you're describing." 00:45:24 The leverage, he says, is in deployment and workflow integration, which is where application-layer companies win.

The Payer-Provider AI Arms Race Is a "Short-Term Distraction"

When Jay cites headlines that billing bots on both sides increased spend by about a billion dollars, Singhal doesn't treat it as a structural problem: "There's no reason this has to exist, right?... the ideal world, in my view, is a world in which providers and payers kind of have a handshake like, 'Hey, here's our ground truth of what happened in this visit.'... And then an AI helps adjudicate that situation... I view that as a bit of a short-term distraction." 00:46:53 This suggests a shared-ground-truth layer (ambient scribing plus adjudication) as a thesis.

Distribution to Free Users Outside the US, Not Premium US Patients, Is the Main Event

Most assume health AI is a premium US product. Singhal says the actual usage skews global and free: "the vast majority of them are free users, and... most of them are outside the US as well. And so we take that kind of question of distribution here really, really centrally." 00:10:05

Procedural Care and Robotics Are Not a Near-Term Story

Against robotics hype, Singhal says the near-term macro lever is prevention: "the biggest macro factor in the next, you know, year or two will just be better kind of proactive preventative stuff, right? So you don't have to actually go in later and get that procedure." On robotics: "My expectation is that it's not like a next few years kind of thing." 00:51:03

3. Companies Identified

OpenAI: Creator of ChatGPT and GPT models. Why mentioned: Singhal leads its health AI research; ChatGPT has over 300 million weekly users asking health questions. Quote: "There's more than 300 million people weekly using ChatGPT for health and wellness types of questions, and we take that as an incredibly important responsibility to get that right, and that's been rapidly growing over time." 00:03:32 — Karan Singhal

Penda Health: Healthcare provider where OpenAI ran clinician-copilot studies. Why mentioned: Evidence that AI as a safety net improves care and surfaces patterns like antibiotic overuse. Quote: "We did this work... in a series of experiments that we ran with Penda Health, where we studied this idea of like a clinician copilot that was deployed... it was a safety net for the clinician as they were doing their work." 00:11:34 — Karan Singhal

Epic: Dominant EHR software vendor. Why mentioned: Deep integration with ChatGPT in both directions, removing workflow fragmentation. Quote: "There's now a way to use Epic in ChatGPT, if you're a provider or a healthcare professional, and also for ChatGPT to be useful in Epic itself... a little sidebar where you can actually use ChatGPT, as well." 00:27:06 — Karan Singhal

UCSF: Academic health system. Why mentioned: Early adopter of the Epic/ChatGPT connection, power users for chart review. Quote: "They've been, you know, one of our earliest adopters for this connection between Epic and the EHR and ChatGPT... clinicians using it for chart review." 00:34:01 — Karan Singhal

AdventHealth: Large US health system. Why mentioned: A leader in deploying agents across the workspace with strong results. Quote: "They've been some of our power users in thinking about deploying agents across the workspace... A lot of health systems, I think, should look to them." 00:34:01 — Karan Singhal

ChatGPT Health / ChatGPT for Healthcare / ChatGPT for Clinicians: OpenAI's consumer, enterprise, and clinician offerings. Why mentioned: Singhal says clinician retention is a pleasant surprise. Quote: "We've had lots of users using ChatGPT for Clinicians and really enjoying it. It's been the retention on the offerings has been like a very pleasant surprise." 00:33:44 — Karan Singhal

The Cognitive Revolution (podcast): Podcast hosted by Nathan Labenz where he told the story of his son's leukemia diagnosis. Why mentioned: Source of the patient story. Quote: "He hosts The Cognitive Revolution podcast and has kind of told his story in that podcast." 00:05:56 — Karan Singhal

AlphaGo / DeepMind: Go-playing AI system. Why mentioned: Source of the "Move 37" analogy. Quote: "This idea of a Move 37 is it comes from AlphaGo... a game that the AlphaGo system... played with Lee Sedol." 00:38:39 — Karan Singhal

Blue Cross Blue Shield Association: Insurer association. Why mentioned: Analysis showing provider and payer billing bots increased spend by about a billion dollars. Quote: "There was some Blue Cross Blue Shield Association analysis that showed that... the respective billing bots and paperwork bots effectively increased spend by a billion dollars or something like that." 00:46:20 — Jay Rughani

New York Times: Mentioned as the source of the article on provider AIs and payer AIs fighting. Quote: "There's a New York Times article that basically provider AIs and payer AIs are fighting." 00:46:20 — Jay Rughani

OpenAI Dots / personal agent products (Instinct, Muse, Grokbot): Emerging personal-agent apps named as the model for future proactive health assistants. Quote: "Personal agents, in general, have kind of emerged on the scene... Instinct, Muse, Grokbot, OpenAI's Dots." 00:12:22 — Jay Rughani

4. People Identified

Karan Singhal: Leads health AI research at OpenAI; previously built specialized medical models. Why mentioned: Guest and architect of OpenAI's health strategy. Quote: "I used to work on specialized models actually in a previous role, and so definitely have kind of experience on both sides of this question." 00:07:04

Jay Rughani: a16z partner investing in healthcare and biotechnology; host. Why mentioned: Frames the investment-relevant questions. Quote: "I'm Jay Rughani, a partner at a16z, leading investments in healthcare and biotechnology companies." 00:02:14

Nathan Labenz: Host of The Cognitive Revolution podcast. Why mentioned: His son's leukemia journey illustrates AI's patient-side impact. Quote: "AI helped them kind of navigate the process of realizing that this was like a pretty serious thing while they were on vacation... And then it turns out he was diagnosed with an aggressive form of leukemia." 00:06:14 — Karan Singhal

Greg Brockman: OpenAI co-founder. Why mentioned: Told a story on the podcast about a friend catching a drug interaction via ChatGPT. Quote: "When Greg Brockman came on the podcast, he spoke a little bit about this situation he had with a friend... her physician was about to administer a certain medication. She double checked that medication in ChatGPT, realized based on her medical history that she would have some drug interaction." 00:10:35 — Jay Rughani

Lee Sedol: World-class Go player who played AlphaGo. Why mentioned: His match produced Move 37, the analogy for AI-driven insight in health. Quote: "A game that the AlphaGo system... played with Lee Sedol, where there's this specific move that just kind of in the moment seemed like a really dumb move." 00:38:39 — Karan Singhal

The global physician network (hundreds of experts): Unnamed physicians across specialties and countries. Why mentioned: The aggregated expert base behind OpenAI's evaluations. Quote: "A global physician network of hundreds of experts that we work with kind of across the globe, across... dozens of specialties and languages and countries of practice." 00:17:49 — Karan Singhal

5. Operating Insights

Aggregate Many Experts Instead of Relying on a Few

When ground truth is contested, don't centralize judgment in a small in-house team. "It turns out if you ask five physicians or something... any question, you'll get like eight different answers." 00:17:49 OpenAI's response was a network of hundreds of experts across specialties and geographies, which makes evaluation data more robust and defensible.

Sequence Investments by the Adoption Bottleneck You Actually See

Singhal describes a clear operating cadence: identify the blocker, then fix it. "A lot of the way we design these offerings is basically to think about like, what are the barriers and bottlenecks and activation energies for people to be able to use the models?" 00:25:08 For enterprise healthcare, that meant roughly two years of security, HIPAA, and privacy work first: "we need to make sure that we're able to offer HIPAA compliance." 00:25:27 Enter regulated verticals by clearing compliance before scaling features.

Learn the Customer's Language Before Selling Intelligence

"Even if the models get really smart, if nobody's like really trusting them or using them, that's not really achieving our mission." 00:26:07 OpenAI invested in "how to speak their language, and how to kind of bring them along this journey." In trust-gated industries, communication and partnership motion are part of the product.

Build Public Benchmarks to Set the Ruler for the Industry

Rather than keeping evaluations proprietary, OpenAI publishes them to define the standard. "Can we create this benchmark that kind of raises the ceiling for everybody, and kind of creates across all model developers and the healthcare industry and mental health experts like this shared ruler, or measuring stick?" 00:29:12 Owning the measurement standard is a strategic position, and it builds trust.

Make Models Explicitly Handle Persona and Uncertainty

Evaluate behavior by user type and train for calibrated hedging. The original HealthBench had "a theme... focused on this distinction of how the model behaved and tailored its level of use of technical jargon... to whether a person was a lay user or clinician" 00:15:54, and includes "evaluations on hedging, for example, when there is this kind of uncertainty." 00:23:16 Product teams in high-stakes domains should evaluate persona-fit and uncertainty-expression, not just accuracy.

6. Overlooked Insights

The Ceiling May Be Limited by Validation, Not Idea Generation

In one sentence, Singhal reframes where drug discovery value accrues: "I'm also very optimistic about a future in which AI helps accelerate a lot of the bottlenecks in validation, which I think will end up being the bottlenecks, as opposed to kind of proposing new ideas or treatments." 00:42:13 He reinforces it later: "the number of molecules and drugs that people want to suggest to be used... is way more than the number of slots of available clinical trials." 00:54:01 If AI floods the funnel with candidate molecules, the scarce resource becomes trial capacity and prioritization, which suggests investable opportunities in trial design, patient recruitment, and validation infrastructure rather than molecule generation.

OpenAI's Real Leverage Is the "Front Door" to Health, and It Is Not Trying to Own Care Delivery

Buried in a longer answer is an explicit strategic statement: "I think there is stuff that OpenAI can do relatively uniquely here because of the front door to health that it has." 00:32:07 Combined with Singhal's statements that OpenAI "won't be in the business of doing all the things" like managing chronic care or direct clinical delivery, this signals that OpenAI intends to be a demand-aggregation and routing layer, and that care-delivery startups plugged into that front door (scheduling, prescribing, diagnostics, chronic care) may receive significant referral flow. The investment question becomes which delivery businesses are best positioned to be the "last 100 meters" partners.