What Trusting ChatGPT With Health Should Mean: Karan Singal on Assistance, Evidence, and Responsibility

1. Guest Background

Karan Singal joins this a16z conversation as OpenAI's health AI research lead. The interviewer speaking in the recording is Jay Rugani, who introduces himself as an a16z partner leading investments in healthcare and biotechnology. Their discussion examines what happens when conversational AI becomes a widely used entry point for health questions: how it should be trained, what trustworthy assistance requires, and where medical services and responsibility remain essential.

Singal supplies one particularly relevant piece of professional background. He previously worked on specialized models and says that experience informs his current preference for specialized healthcare training inside general frontier models. The interview supports that account of relevant technical experience; it does not establish a medical qualification, a clinical practice history, or a fuller career biography.

His perspective connects research with product development and institutional deployment. He frames OpenAI's health ambition in two parts: making trustworthy health intelligence available for free or nearly free, and developing capabilities that were previously unavailable. Those goals organize the conversation's movement between immediate assistance and more speculative discovery. They also explain why his account needs careful reading. He speaks as a leader involved in the work, offering useful detail about its design, while his claims about effectiveness and future impact still need to be assessed at the level of evidence actually presented.

2. What the Episode Covers

The interview locates an immediate opportunity in the gaps surrounding appointments. Rugani makes the problem tangible through his own dermatology visit: he prepared questions in Apple Notes, still forgot one, and then faced a four-month wait for another appointment. The story is not a measurement of healthcare performance. It illustrates why a short consultation can leave substantial informational work unfinished. Singal identifies those intervals as a particularly useful consumer setting for AI: preparing questions, understanding what a doctor said, and working out how to follow up.

He places that opportunity against reported demand. More than 300 million people, he says, use ChatGPT weekly for health and wellness questions; the vast majority use free access, and most are outside the United States. These are Singal's usage claims. They establish the scale of the responsibility he describes, rather than showing that the answers are clinically reliable or that users experience improved health.

The most affecting example is his account of Nathan and his young son. According to Singal, AI helped the family recognize that fatigue, irritability, and other symptoms during a vacation might be serious enough to warrant testing. The child was subsequently diagnosed with aggressive leukemia. Nathan then reportedly used GPT-5 Pro to understand laboratory reports and navigate treatment, sometimes before a specialist had time to explain them. The account illustrates both prompting a decision to seek help and supporting understanding during treatment. It supplies no comparative evidence from which to estimate diagnostic accuracy.

The conversation expands from responding to questions toward proactive personal assistance. Singal endorses the direction of an assistant that understands someone's history and preferences, while explicitly saying it does not replace a doctor. He also states that connected consumer health information is not used for training or advertising targeting. Scheduling, checking insurance coverage, supporting a prescribed post-discharge plan, and dealing with billing errors appear as possible agent tasks, rather than a demonstrated package of universally available services.

Making those tasks work requires access to a fragmented healthcare environment. Singal points to legacy software and warns against expecting every system to acquire a ready-made connector quickly. Computer use—agents operating existing interfaces—could offer another route. He consequently separates model capability, product design, and partnership deployment. A capable model can remain unused if the effort required to apply it is too high; enterprise adoption also requires security, privacy, and HIPAA-related compliance work. His account of Epic–ChatGPT connections and UCSF's early exploration of chart review makes that argument concrete: the product must enter the clinician's workflow, including reducing the need to switch between interfaces.

3. Core Views: Reasoning, Examples, and Limits

Singal's architectural argument is that medical specialization can be built into a general model. He says models released over the preceding year received healthcare data at every training stage, including data developed with physicians. His reasoning is that a separate, smaller medical model may fall behind rapid frontier improvements, whereas specialized training inside a frontier model retains gains in scale and reasoning. That is a coherent account of the intended mechanism. The interview does not provide a controlled comparison establishing that this approach outperforms every alternative medical architecture.

Trust, in his account, also involves more than answering medical knowledge questions correctly. He describes HealthBench evaluating whether explanations suit a layperson, an experienced patient managing complex illness, or a clinician. Technical precision has limited practical value if the intended reader cannot use the explanation. He also describes hundreds of experts across specialties, languages, and countries of practice contributing to evaluation and training. The rationale is to avoid making a small group of physicians the sole arbiters of disputed cases. Broader participation provides a method for handling disagreement; it does not establish that disagreement has disappeared.

The sharpest behavioral test arises when reassurance conflicts with safety. A user may want to hear that nothing is wrong, or seek an alternative treatment, while the medical assessment points toward seeking care. Singal says training and evaluation prioritize identifying danger even when the user does not welcome the answer. This makes the difference between user satisfaction and useful assistance concrete: a comforting response can fail the person it pleases. When a model and a doctor disagree, his proposed next step is another doctor's opinion. The model's contribution can therefore be to expose a question requiring review, without its fluency becoming a substitute for clinical adjudication.

Another important distinction concerns two sources of uncertainty. Singal describes a chain from biomedical papers to guidelines to everyday practice. Citing unnamed studies, he says guideline adoption can take five to seventeen years, and proposes point-of-care access to literature as a way to compress that delay. The interview establishes this as his account and ambition, not a demonstrated reduction in adoption time. Separately, some questions remain unresolved by medicine itself. Singal says a model should acknowledge insufficient evidence even after searching. Better retrieval can address missing access to knowledge; it cannot manufacture a scientific conclusion that the available evidence does not support.

Mental health makes the evaluation problem more demanding. Singal describes work with licensed experts worldwide to define measurable behavior in difficult situations, acknowledging greater disagreement than in physical health. He wants evaluation to cover positive assistance as well as a baseline safety floor. That expands the definition of quality, but the conversation does not supply enough methodological detail or outcome evidence to establish sustained therapeutic benefit. A shared benchmark is a tool for assessing development, rather than proof that every consequential use has been validated.

The clinical safety-net examples deserve similarly specific interpretation. Singal describes Panda Health experiments involving proactive medication-related warnings and flags for antibiotic overuse. In the latter example, the model reportedly drew attention to resistance beyond the immediate local case. The interesting mechanism is a supplementary check on considerations a clinician might overlook. Yet no sample size, quantified clinical effect, or burden from incorrect alerts is reported. A persuasive example of a warning cannot establish the net benefit of deploying the whole system.

The economic discussion is strongest where it preserves these distinctions. Singal reports roughly 80% reductions in several Advent Health administrative workflow costs and encouraging clinician-product retention, without detailed methods or retention figures. Using the transcript's model names, he also says GP6 Luna exceeded 5.6 Soul on HealthBench Professional while costing 34 times less, without providing scores or pricing details. Workflow expenditure, adoption feedback, and model input prices are different measurements. Singal explicitly says health intelligence has not yet had a macro effect on healthcare spending; integration, deployment, and scale must connect cheaper capability to actual savings.

Finally, information is only part of medical care. Rugani describes a bundle that includes diagnosis, triage, planning, empathy, and legal responsibility. Singal positions OpenAI as a navigator and partner enabler, rather than the supplier of the entire bundle. Even when patient and clinician both consult AI beforehand, the visit can still collect missing history, support investigation and physical interaction, and provide reassurance. A comprehensive-looking answer cannot establish that all relevant information has been gathered or that the corresponding medical service has been delivered.

4. Learning and Application

For an individual user, the most direct application is to organize the work between appointments. Questions about instructions, reports, and follow-up arrangements can become a clearer agenda for the next consultation. A practical approach is to distinguish an explanation from a judgment that needs clinical confirmation. If the model and doctor disagree, that disagreement can be formulated as a specific question or a request for another medical opinion. This follows the interview's navigation role without assuming that a compelling answer completes the medical task.

For proactive products, Singal's rising-A1c example offers a useful way to divide responsibilities. He imagines identifying a change and prompting medication review or attention to adherence. An evaluation should consequently distinguish detecting the change, delivering the prompt, obtaining appropriate review, and improving the eventual outcome. These are proposed evaluation steps derived from the discussion, not results reported in it. A reminder can initiate a process; it cannot automatically count as a completed diagnosis, a medication adjustment, or successful chronic disease management.

Healthcare organizations and builders can use the model–product–deployment framework to choose a narrower starting task. Chart review, scheduling, and support for discharge instructions each have different completion conditions. A connector or an agent's ability to operate an interface establishes a technical route, while practical evaluation still needs to examine review requirements, failure handoffs, and the burden on staff. Administrative savings also need a defined workflow, baseline, and cost boundary. Advent Health's reported percentage should prompt questions about those conditions rather than become an assumed reduction in an entire hospital's costs or a patient's bill.

The payer–provider discussion adds an institutional lesson. Rugani tentatively recounts a claim about billing bots increasing spending; Singal proposes shared encounter information, AI-assisted adjudication, and limited appeals instead of an automated arms race. A team considering that direction could begin by specifying the shared record, who may rely on it, and how disagreements can be challenged. The interview does not show that the parties have adopted the arrangement or that technical automation resolves their differing incentives.

Longer-term projects need an explicit separation between discovery and validation. Singal's health 'move 37' metaphor concerns unexpected insights, including predictions assembled from years of records and conversations. A useful research program would define the prediction, the information available when it is made, and the test by which it will be judged. The ability to summarize a decade of records should not itself be scored as reliable risk prediction. His advice to expect stronger models points builders toward care delivery and journey monitoring, while his forecasts about prevention and the slower development of medical robotics remain conditional on deployment, reimbursement, and integration.

The same discipline applies to drug research. Singal's five-to-ten-year forecast includes new targets or molecules, but also prioritization and smoother clinical validation. Because he describes candidate proposals exceeding trial capacity, the practical question becomes whether AI helps scarce validation resources test more worthwhile candidates. Counting generated ideas would miss that bottleneck. Across personal reminders, institutional workflows, and scientific discovery, the useful application of this episode is to examine each transition from information to action to outcome, and to identify which transition still lacks evidence.

Source

• Original episode: Should You Trust ChatGPT With Your Health?