September 21, 2026

@michaelokun

Can we trust AI chatbots for Parkinson’s disease information? Generative AI chatbots use artificial intelligence to create conversational answers to questions, but in health care a convincing answer is not necessarily a correct or safe answer. Lange and colleagues describe in a new paper in The Lancet Regional Health Europe what happened when a Parkinson’s-specific AI chatbot called jAImes was evaluated during real-world use. Key points: - The chatbot handled 2,035 conversations and 6,146 messages over 129 days, and automated screening classified 88.6% of conversations as good. - Human review uncovered clinically important errors that automated monitoring missed, including incorrect medication information and a failure to appropriately escalate suicidal ideation. - The authors propose CARE-LLM, a monitoring framework combining automated screening, expert human review, random sampling and a feedback loop to identify and correct failures. My take: AI is coming to Parkinson’s care, whether we are ready or not. The exciting part is its potential to put useful information into the hands of folks and families 24/7. The sobering part is that an AI answer can sound terrific and still be dangerously wrong. This study reinforces for me that we should not simply ask whether an AI performs well. We should ask what happens when it fails, whether we can detect the failure and how quickly a human can get into the loop. Here are 5 points that resonated w/ me: 1- A Parkinson’s-specific AI grounded in curated information can perform well across thousands of real-world conversations, but well is not the same as perfectly safe. 2- Automated AI monitoring can itself miss important AI mistakes, which is why human expert review remains critical. 3- Medication advice and mental health emergencies are areas where getting an answer almost right may not be good enough. 4- AI systems used in health care need surveillance after deployment, not simply testing before launch. 5- The future should not be AI versus health care providers. It should be carefully designed AI w/ health care providers, rigorous monitoring and the person w/ disease remaining at the center. https://www.thelancet.com/