WTF is going on at K Health?
Short summary
A journalist interviews K Health's CEO about why the company quietly shut down its consumer symptom checker (which ranked last in independent testing) and what PatientGPT actually is. The CEO dodges the core question while defending K Health's move toward health system partnerships.
- K Health's consumer symptom checker scored 21.5% accuracy in a 2025 systematic review of 40 symptom checkers—dead last. GPT-4 scored 71.3%, and laypeople with no app scored 47-62%.
- K Health discontinued its consumer platform December 31, 2025 with no announcement. When asked if poor performance drove the shutdown, the CEO gave a 6-minute answer that didn't address the question.
- PatientGPT's architecture remains vague: the CEO describes "about a dozen agents" running on a mix of open-source and unnamed "other models," with costs determining which models get swapped in. Three different descriptions exist across seven months.
- K Health's own studies (where K Health funded or co-authored) show 77-85% accuracy, while independent research shows 21.5%—a pattern the journalist notes repeats across the AI health industry.
Full summary
K Health, a telehealth startup that's been around longer than most people remember, raised $59 million in 2026 and got called out for it on Twitter. The core problem: their AI symptom checker was genuinely bad. A 2025 systematic review ranked K Health dead last among 40 symptom checkers at 21.5% accuracy—worse than untrained laypeople (47-62%), way worse than Doctorlink (90%), and embarrassing next to GPT-4 (71%). Earlier studies had similar results. K Health quietly killed the consumer platform on December 31, 2025, with no announcement. When pressed directly on whether poor performance drove the shutdown, the CEO gave a 6-minute non-answer about finite resources and needing medical records. The author, who runs his own healthcare AI startup, tested the old checker himself and wasn't impressed.
The real mess is K Health's studies versus independent ones. When K Health funded research or had employees as authors, the AI looked great—85% accuracy here, 84% clinician agreement there, 77% optimal recommendations there. When independent researchers tested it, K Health ranked last. The CEO pushed back hard, insisting the researchers were independent and he had no editorial control. Fair point on the passion, but the journals require financial disclosures for a reason. The author's blunt take: when K Health sponsors the study, K Health wins; when Ada Health's employees write it, Ada wins. It's not a K Health problem—it's how the game works.
The author actually agrees with K Health on some things. Bayesian and decision-tree models are better than LLMs for medical AI because they can say "I don't know" without hedging. K Health's approach of training the AI "quite specifically not to guess" beats what Babylon tried to pull off with spreadsheet decision trees. Neural networks have done solid work on complex conditions. But the article cuts off mid-sentence on generative AI's fundamental design flaw, leaving the broader argument incomplete.