I’ve said before that AI won’t replace doctors. But what happens when patients replace doctors with AI? AI hasn’t suddenly smarter, but more and more patients are using it as their first stop for medical advice. They are getting labs from third-party companies and using LLMs to understand the results. We are moving into a world where people arrive already carrying an AI generated differential diagnosis in their pocket, sometimes helpful, sometimes terrifying.

So what do we do as clinicians? The challenge is not fighting this trend. It is learning how to guide it responsibly. Clinicians have the foundation and context that LLMs lack. Just like WebMD, while LLMs are good tools to understand symptoms and interpret labs, they tend to predict worst case scenarios. The physicians who understand how these systems think will be better positioned to calm fears, correct inaccuracies, and help patients use AI without spiraling into unnecessary testing or anxiety.

Inside this week:

Plus: Evidence grounded AI search built specifically for physicians.

Let’s dive in.

🤖 Patients Are Skipping the GP for Chatbots: A new UK study found that roughly one in seven people would rather ask an AI chatbot than see a physician for health advice. That number honestly feels believable already. Patients increasingly want instant answers, reassurance, and explanation, especially when access barriers are high. The concerning part was that some users delayed seeking care because the chatbot reassured them incorrectly.
Why this Matters: AI is rapidly becoming the first layer of triage whether we like it or not, and clinicians need to understand what patients are hearing before they walk into the room.

🧠 AI Outperformed Physicians in Emergency Triage Simulations: On the flip side, a Harvard study showed reasoning models outperforming physicians on text based emergency triage scenarios. This is where the nuance matters. AI is becoming remarkably good at pattern recognition and differential generation, especially when clean information is provided. But real patients are messy, emotional, incomplete, contradictory, and contextual.
Why this Matters: The future is likely physician plus AI, not physician versus AI. The clinicians who learn to supervise AI intelligently will probably practice better medicine.

🩺 AI Chatbots Struggle With Subtle Mental Health Signals: A new clinician led benchmark study found that major AI chatbots often missed nuanced warning signs of worsening mental health, especially when distress was indirect rather than explicit. The systems performed reasonably well with obvious suicide related language but frequently failed when conversations involved gradual emotional decline, ambiguity, or passive hopelessness. What really stood out was how “agreeable” many models remained even in potentially dangerous conversations.
Why this Matters: Patients increasingly use AI for emotional support, but current systems still struggle with clinical nuance and hidden risk signals that experienced clinicians recognize intuitively.

🧬 LLMs Are Highly Vulnerable to Patient Self Diagnosis Bias: One of the most important papers I read this week explored how LLMs can be steered toward incorrect diagnoses when patients frame symptoms around their own fears. If a patient strongly suggests cancer, neurologic disease, or another feared diagnosis, the model often reinforces that direction instead of correcting it. This mirrors exactly what many clinicians are now seeing in practice.
Key Takeaway: AI does not just answer questions, it amplifies the framing patients bring into the conversation.

🩻 Multimodal Conversational Diagnostic AI Advances Further: A new Nature Medicine paper explored conversational diagnostic systems that combine LLM reasoning with multimodal patient data, including imaging, labs, and symptom dialogue. The paper focused heavily on explainability, adaptive questioning, and dynamic interaction rather than static diagnosis generation.
Key Takeaway: The future of clinical AI will likely depend less on “the answer” and more on how well systems explain uncertainty and reasoning.

⚖️ OpenAI Faces Lawsuit Over Harmful Medical Advice: A wrongful death lawsuit this week alleged that ChatGPT provided dangerous drug related recommendations to a teenager who later died from overdose complications. This case will likely become a major inflection point in defining accountability for consumer AI medical guidance.
Why this Matters: Patients increasingly treat conversational AI like a trusted authority even when the system was never designed for autonomous medical decision making.

🧾 Healthcare Experts Warn About AI Becoming an “Unregulated Front Door”: Several policy experts are now describing AI chatbots as an emerging parallel healthcare access system operating outside traditional regulation. That framing feels accurate. Patients are already using AI for triage, reassurance, symptom interpretation, and medication questions before ever interacting with us.
Why this Matters: Healthcare regulation still assumes physicians are the primary gatekeepers of medical advice, but that assumption is rapidly changing.

I’ve talked about OpenEvidence before. But its more important than ever. I increasingly use this to check rapidly evolving evidence before clinic or between consults. What I like is that it behaves more like a clinical literature assistant than a generic chatbot. It creates charts, graphs, and visuals that are easy for me to understand. I also use it to anticipate what patients may have already read online about a diagnosis before I walk into the room.

The most important skill in healthcare AI may not be prompt engineering. It may be contextualization.

Patients do not just need answers anymore. They need help understanding which answers matter, which possibilities are unlikely, and when uncertainty is normal. AI can surface possibilities incredibly fast, but clinicians remain responsible for weighing probability, context, and human nuance.

The physicians who learn how patients are using AI will be far better positioned to reduce fear instead of escalating it.

Forward this to a colleague who keeps hearing, “Well ChatGPT told me…”