The AI Model Race Is Over. The Data Race in Healthcare Is Just Starting (Robert Tovornik, Better)

AI models are "eager to please" — and in healthcare, that's a liability. So what should LLMs never be allowed to do in clinical software? Three years after GPT-3 reached the public, frontier models have largely converged in capability. In this episode, Robert Tovornik, Innovation Lead at Better — the healthcare IT company building on openEHR — explains why the real differentiator in healthcare AI is no longer the model but the data layer underneath it. He makes the case for keeping clinical coding and terminology in deterministic systems, confining LLMs to retrieval and orchestration, and validating AI the way you'd come to trust a colleague: through experience, not certification. Guest: Robert Tovornik, Innovation Lead, Better What the conversation covers: - Why frontier LLMs are converging — and why context now matters more than model capability - What AI should never do in clinical software: inference vs retrieval - Why ICD-10 and SNOMED coding should stay in deterministic systems, not ChatGPT - How to validate non-deterministic AI systems when unit tests no longer work - Automation bias: what happens when users stop checking AI outputs - Conversational EHRs — solving the "missing button" problem in clinical interfaces - Vibe coding vs regulated clinical software: why one iterates in hours and the other in years - An ambient AI scribe built in two weeks — deployed in India, stalled in Europe - EU AI Act, data residency laws, and the cost of compliance - Digital twins, ambient AI, and what hospitals should invest in before deploying AI Faces of Digital Health explores how healthcare systems around the world adopt digital technologies and AI. 🔗 Website: https://www.facesofdigitalhealth.com 🎧 Spotify: https://open.spotify.com/show/4cElKJHrauyP6QJQaCkvdY 🎧 Apple Podcasts: https://podcasts.apple.com/gb/podcast/faces-of-digital-health/id1194284040 📰 Newsletter: https://fodh.substack.com 💼 LinkedIn: https://www.linkedin.com/company/faces-of-digital-health #healthcareAI #openEHR #digitalhealth #healthIT #EHR #clinicalAI #healthcareinnovation 02:20 Three years of GPT: from model capability back to data and context 04:26 How AI changed software development inside an openEHR vendor 06:13 Clients now arrive with AI-informed (and misinformed) requirements 09:05 Why clinical coding belongs in deterministic systems, not ChatGPT 11:27 "Eager to please": why LLMs shouldn't be trusted with inference 14:03 Validating non-deterministic AI when unit tests no longer work 16:50 How do you trust AI? The same way you trust a colleague 18:03 Regulation, compliance costs, and the automation bias problem 20:09 Conversational EHRs and the missing-button problem 23:02 Vibe coding vs iterating regulated clinical software 25:23 Users are building AI experience faster than health systems 27:50 An ambient scribe built in two weeks — adopted in India, stalled in Europe 29:16 Data quality as the differentiator between good and bad AI systems 31:10 Ambient AI, operation prep, and the digital twin horizon