Gradient Zero

I used ChatGPT to research my new heart condition. And no, I didn't fire my doctor. It started recently while I was training for a marathon. Like many patients, I wanted to understand the underlying mechanics of my symptoms before my appointment. So I turned to an LLM. Not to replace medical advice (the actual diagnosis and treatment plan obviously came from my cardiologist). Instead, I used it to translate dense medical terminology, understand the physiology, and prepare a highly specific list of questions. But from my experience, when used as an educational layer, these tools don't replace clinical expertise. They help patients become health-literate, informed partners in their own care before they even walk into the consultation room. To the clinicians and healthcare innovators in my network: how are you navigating this shift? Are you finding that AI-informed patients lead to more productive, nuanced consultations, or is it creating new friction in the clinic?

Healthcare has a Shadow AI problem, and most people running IT already know it. When clinicians can't get to an approved AI tool, a lot of them just open ChatGPT on their phone instead. Not to break the rules, but because they've got a discharge summary to write and twelve minutes to do it. Banning AI doesn't kill the demand, it just moves the demand somewhere you can't see it - off the network, onto personal devices, out of any audit log. Now you've got patient data sitting in a consumer chatbot and no idea it happened. That's the actual risk. Not that clinicians want AI, but that they're using it in the one place you have zero governance. So the fix isn't a stricter policy or another training module telling people to stop. It's giving them a secure tool that's good enough that they don't go looking for the consumer one. If the approved option is slower or clunkier than ChatGPT, people will keep using ChatGPT. Every time. Curious how others are handling this. Are you trying to lock it down, or building a version of AI your staff will actually choose to use?

My perspective on healthcare AI changed the day I became the patient. Sitting at home with an ECG report feels very different from discussing AI in a boardroom. You suddenly have questions. You want to understand what's happening, not just wait days for every answer. Uploading the terminology from my report to a standard LLM is where AI became genuinely useful for me - not by replacing my doctor, but by helping me understand medical information, prepare better questions, and become a more informed patient. That experience reinforced something I already believed - The future of healthcare AI isn't replacing clinicians. It's helping patients have better conversations with them.

A few months ago, many people treated AI dependency as a theoretical risk. Then it became real. When Anthropic's Fable model was suddenly restricted following a U.S. government decision, it wasn't just another AI headline. It exposed something much bigger. Many European companies are building products and internal workflows on models they don't control. If access changes overnight (because of regulation, geopolitics, or a provider's own decision), those businesses suddenly have a problem they can't solve with better prompting. For me, the biggest lesson wasn't about Anthropic. It was about dependency. This isn't an argument against using frontier models. They're incredibly powerful tools. But if your entire product or business depends on a single provider remaining available forever, that's not an AI strategy. It's an assumption. The conversation in Europe is slowly shifting from "Which model is best?" to "How do we stay operational if the landscape changes?" That's a much healthier question to ask.

Europe doesn't have an AI talent problem. It has a deployment problem. I'm actually optimistic about companies like Mistral. European AI models, sovereign infrastructure and on-device AI all matter. Especially in healthcare. But we shouldn't confuse good models with production-ready healthcare systems. The difficult part isn't training another model. It's integrating that model into hospitals, clinical workflows and existing infrastructure. Healthcare AI won't be won by whoever has the biggest model. It will be won by whoever earns the most trust.

Sovereign AI doesn't mean building your own foundation model. It means not being held hostage by your vendors. The term gets thrown around a lot, but for most organizations, it's really about maintaining control over the parts that matter. Where does your data live? Can you choose which model to use, or switch providers without rebuilding everything? Can your systems continue operating if one vendor changes pricing, access, or strategy? Those questions have much more impact on long-term resilience than whether you're using the latest benchmark leader. The conversation shouldn't start with "Which model should we use?" It should start with "Which parts of our AI stack do we actually control?" That's what sovereignty looks like in practice. What's the hardest part of your AI stack to keep control over right now?