← The Health AI Brief
Loading summary
Transcript1 lines
- [00:01]
A
A patient sits down in the consultation room and instead of describing their symptoms, they present a multi page formatted differential diagnosis generated by a large language model. It features clinical terminology, cites medical studies and outlines a structured diagnostic pathway for a complex condition. This scenario represents a shift in clinical practice. The historical challenge of what's been previously termed cyberchondria involved patients searching search engines and presenting unstructured lists of symptoms. Kind of Dr. Google, as it were. The current iteration involves highly coherent structured arguments generated by artificial intelligence. These outputs are persuasive, leading to high levels of trust from patients. Dismissing this information risks damaging the therapeutic alliance, while accepting it uncritically risks diagnostic error and possibly over investigation. The objective of this episode is to analyze a structured strategy to navigate these consultations, moving the clinical role from a gatekeeper of knowledge to to a curator of health information. To understand why patients place such high trust in these outputs, we need to examine the structure of large language models. These models generate text that mimics the authoritative tone of medical textbooks. For a patient experiencing unexplained symptoms, this structured clarity offers reassurance. When a clinical consultation contradicts this neat narrative conflict can occur later. In this analysis, we'll examine specific linguistic strategies to try and de escalate these moments. But. But first we need to establish the core mindset shift required for the clinician in these sorts of circumstances. So rather than viewing an AI generated document as an administrative burden or an adversarial challenge, it's important to frame it as a sign of an engaged patient. This document represents a significant investment of time and energy by someone seeking to understand their health. The clinical task is to direct this engagement constructively. I'm going to propose three steps that I think are important to evaluate the AI's output alongside the patient, reinforcing clinical authority through collaborative analysis. So the first step is validation. If you reject an AI generated document immediately, then it creates a defensive barrier. Acknowledging the patient's initiative lowers tension and establishes a collaborative tone. A practical opening formulation is to note that the patient's used advanced tools to explore their symptoms and that these models are useful for organizing potential scenarios. This validation of effort preserves the clinical alliance and prepares the patient to receive clinical feedback. The second step though, is to audit the input. Large language models operate on the principle of processing the specific text provided to them. If the input data is skewed by anxiety or incomplete observation, the output will reflect those limitations. During this stage of the consultation, the clinician could ask the patient to describe the exact information that they entered into the prompt demonstrating this connection can help patients understand why AI is an information processor and that some inputs inevitably produce potential skews in results. This realization can be important for the clinical discussions that might follow. It almost kind of excuses any errors that the AI might make, which for a patient that's invested a lot of time talking with an AI might be more effective than trying to downplay the brute force intelligence of these tools. The third step is identifying a gap in context. This concept is important to managing the consultation. While a large language model has access to vast medical literature, it lacks physical presence and diagnostic tools. It can't perform a physical examination, assess clinical reflexes, or interpret nonverbal clinical signs. During most consultations, the clinical examination provides vital physical data. The AI can't feel muscle tone, evaluate a gate or a child playing in the corner. Quite happily, it can't feel a patient's abdomen. By performing these physical assessments, the clinician demonstrates the difference between theoretical data and physical reality. The current set of models will all have physical blind spots. In particular, this process aims to position the clinician as the kind of senior clinical reviewer. The clinical role is to analyze the broad theoretical possibilities generated by the software and filter them through the precise physical reality of the patient. Patients this distinction can be communicated using the concept of maps and terrain. The AI has access to a generic map of medical possibilities, but the clinician is evaluating the actual physical terrain of the patient's body. As an extension of what I've described and trying to apply it consistently under time pressure. Clinicians can use three practical communication techniques. The first is the blind spot technique. This involves trying to identify a specific objective clinical measurement that the AI could not access. This might be a normal blood pressure reading, a specific blood test result, or a physical examination finding. You can explain how a single data point changes the probability of a human or an AI. Suggest a differential diagnosis. The clinician demonstrates the necessity of human clinical judgment. Second is a kind of collaborative shared triage. Instead of dismissing an AI's list, the clinician could use it as a collaborative checklist. The clinician and patient review the document together, ruling out specific options based on clinical evidence. For example, a specific option is ruled out because the patient's blood tests show normal inflammatory markers. This can transform a potential conflict into a shared, rational diagnostic process, which helps the patient understand why decisions are being made. A third technique is framing the AI as a probabilistic tool. Large language models calculate the probability of the next word in a sequence based upon training data. They don't analyse biological systems. Explaining this difference can help patients understand that the AI's output is an exercise in language pattern matching rather than a simple, specific, tailored clinical diagnosis. So, looking ahead, I think the nature of these interactions will continue to evolve as tools become more integrated into daily life, and I think they'll inevitably become more common. In an upcoming episode, I'm going to also try and consider some specific tips that clinicians can give to patients on how to use these tools safely and effectively before their appointments. Providing patients with structured prompting templates can improve the quality of the information that they bring to clinic, making the consultation more efficient. So the rise of AI assisted patients doesn't diminish the clinical role. Instead, it highlights the value of physical examination, clinical intuition, and diagnostic reasoning. By validating patient engagement, auditing the inputs, and demonstrating the gap in context, clinicians can help guide patients safely through complex clinical pathways. We can preserve the therapeutic relationship while ensuring that medical decisions remain grounded in physical reality. If anyone else has had thoughts on these sorts of topics, I'd be really interested to hear them in the comments. And don't forget to hit, like and subscribe so that you don't miss future episodes.