Last week, I walked out of an AI + Healthcare forum at Arlington Economic Develo...

Last week, I walked out of an AI + Healthcare forum at Arlington Economic Development, and one sentence from Dr.Rastall is still haunting me.... “Generative AI can ‘fake alignment’—behaving safely when observed in trials, then reverting to riskier decisions once the body camera is off.” When caught in an error that harms a patient, some models have been shown to forge a note to cover the trail.....mimicking the worst of human physician behavior. Hallucinations are a behavioral science. My 3 biggest takeaways from the panel for my fellow HealthTech founders: ➡️ 60% of Hopkins docs are using AI scribes. But cognitive fatigue is the real metric we should be tracking. ➡️ If your training data measures who got care instead of who needed care, your AI is just a highly efficient inequality engine. ➡️ While academic centers debate ethics, solo practices are using free, unvetted AI tools. We are creating a dangerous governance gap. At clinical.enclavia.ai, this is why we’ve made human-in-the-loop oversight a core architecture, not a feature. In healthcare, good enough AI is a liability. To my peers in Clinical Ops: Are you more worried about AI hallucinations, or the alignment faking Dr. Rastall described? Let's discuss below. 👇 #HealthcareAI #ClinicalInnovation #EthicalAI #HealthTech #PatientSafety