Hippocratic AI says it is developing a language model for healthcare tasks, with safety and human review built into its approach. The company emerged from stealth with $50 million in seed financing and a valuation described as being in the “triple-digit millions.” Its plans raise a central question: how can a system designed to assist patients earn trust in a field where mistakes can carry serious consequences?
Support tasks, not diagnosis
Co-founder and CEO Munjal Shah said the model is not focused on diagnosing illness. Instead, Hippocratic describes consumer-facing uses such as explaining benefits and billing, offering dietary advice, sending medication reminders, answering pre-op questions, onboarding patients and delivering “negative” test results that indicate nothing is wrong.
These are varied responsibilities. Some involve routine explanations or reminders; others touch health decisions and patient expectations. Even when a system is not making a diagnosis, its wording and accuracy may shape what a person understands or does next.
That makes the company’s emphasis on safety important to evaluate. Shah said Hippocratic is building a model focused on safety, certifying it with healthcare professionals and working with the industry on data retention and privacy policies consistent with healthcare norms.
Claims of expertise need scrutiny
Hippocratic says its model outperforms leading language models, including GPT-4 and Claude, on more than 100 healthcare certifications. The company’s examples include the NCLEX-RN for nursing, the American Board of Urology exam and the registered dietitian exam.
Those claims sit alongside lower results in other areas. According to Hippocratic, the model scored 71% on the certified professional coder exam, covering medical billing and coding, and 72.7% on a hospital safety training compliance quiz. The results suggest that performance may vary by task, so broad claims about capability do not settle whether the model is ready for a particular role.
The company says each role, such as dietician, billing agent or genetic counselor, will be released only after people who do that work agree the model is ready. That review process could provide a meaningful checkpoint, though the article does not provide details about how readiness is judged or how the model performs in real patient interactions.
Human care is hard to measure
Hippocratic also says its system can detect tone and communicate empathy, drawing on the idea of bedside manner. Shah argues that interactions that leave patients with a sense of hope, even in grim circumstances, can affect health outcomes.
To assess this quality, Hippocratic created a benchmark for traits including “showing empathy” and “taking a personal interest in a patient’s life.” The company’s model scored highest across the categories it tested, including against GPT-4. But a benchmark result does not by itself show how patients will experience those exchanges, or whether a system can respond appropriately across the range of situations people bring to healthcare.
There is also a broader risk of bias. The source article points to a 2019 study that found an algorithm used by many hospitals to decide which patients needed care treated Black patients with less sensitivity than white patients. Models trained on biased medical records, studies and research may carry similar problems forward.
What remains unknown
At the time described in the article, Hippocratic had not made its model available. It had also not shared details about its partners or customers, or information about the data used to train the model or potentially used in future training. The company said it would use “de-identified” data.
Without access to the model and clearer information about its training and evaluation, outsiders have limited ability to assess its safety claims. That gap matters because people may trust automated responses even when they are wrong—a tendency known as automation bias. In healthcare, misplaced confidence can have high stakes.
Hippocratic’s funding, co-led by General Catalyst and Andreessen Horowitz, gives it resources to pursue its plan. Shah said most of the $50 million seed tranche would go toward talent, compute data and partnerships. Whether that investment produces a dependable healthcare tool will depend on evidence about how the model works in practice, how its limits are communicated and how it handles the risks its builders say they are addressing.