Science Friday explores AI in medicine with Dr. Jonathan Chen

Episode snapshot

In this Science Friday episode Flora Lichtman speaks with Dr. Jonathan Chen about the rapid adoption of artificial intelligence in medical practice. Doctors describe using AI for everyday tasks such as generating replies in electronic charts and assisting with curbside consultations, while acknowledging limitations like hallucinations and bias. The discussion covers tools from general chat interfaces to medical specific platforms, the need for source back‑ups, training gaps for clinicians, hospital policies, and privacy and liability concerns. The host and guest also compare AI uses in radiology and the broader health system, emphasizing responsible use and physician judgment.

Overview

The podcast centers on how doctors are integrating artificial intelligence into routine clinical work. Flora Lichtman interviews Dr. Jonathan Chen, an associate professor of medicine and director for medical education in AI at Stanford, to unpack what is happening in real world settings. The conversation traverses the pace of change over the last few years, the variety of AI tools being used in medicine, and the core idea that AI is a powerful assistive technology rather than a replacement for professional judgment. Throughout, the discussion highlights both practical benefits and significant risks that require careful governance, training, and safeguards.

Adoption and impact in clinical practice

The speakers explain that usage of AI among doctors has grown rapidly in the United States. Many clinicians are using AI without always recognizing it, for example through AI‑assisted search responses or in electronic charts to draft patient replies. They note that the most compelling AI tools in daily workflow are ambient scribe features that transcribe conversations and provide quick answers, functioning as a second set of eyes. Over time, AI has moved from an experimental stage to integration in diagnosis, documentation, and staying current with patient data. Yet Chen cautions that AI is not a universal oracle and must be used with robust verification and source attribution.

Tools and platforms

The episode surveys a spectrum of AI tools used by clinicians. General chat platforms like ChatGPT, Claude, and Gemini are common, while medical‑specific interfaces are rising in prominence with products from Amboss, Glasshel, Doximity, and OpenEvidence. The latter show retrieval augmented generation that pairs a language model with access to guidelines or articles so the system can point to sources rather than fabricating citations. Chen emphasizes the importance of rag retrieval augmented generation and the ability to click through to the original sources to verify information.

Applications in daily practice

Doctors describe AI as a powerful assistant for organizing information and answering questions in real time. Ambient scribe tools can listen to conversations and generate notes, while AI can provide quick second consultations that help clinicians quickly assess guidelines and best practices. The interview contrasts this with traditional approaches, such as looking up articles in a medical encyclopedia and reading through long texts to determine applicability to a patient. The improved accessibility of AI enables faster, more comprehensive responses but requires clinicians to interpret and apply information in context.

Challenges and risks

A central concern is hallucinations or confident miscitations. Although AI is getting more accurate, it can still produce plausible but wrong information or misinterpret studies, especially if the literature base is misread. Chen highlights the danger of treating AI as an oracle and stresses the need for human judgment and the ability to verify claims by checking sources. He also notes different risks depending on what the question is and the quality of supporting evidence. Retrieval augmented generation helps mitigate these issues by linking to relevant sources.

Training, education and policy

Conversation turns to how clinicians are trained to use AI tools. The consensus is that formal training is lacking and that many clinicians learn by doing rather than through structured curricula. Chen describes his role as Director for Medical Education AI at Stanford as a response to this gap, promoting education on how AI systems work, their caveats, and guardrails. The discussion then broadens to hospital policies and regulatory frameworks, noting that policies for AI use are evolving rapidly and may lag behind technology. Privacy, HIPAA compliance, and liability are identified as major policy challenges, with debates about who bears responsibility when AI advice leads to harm.

Radiology and AI

The conversation touches computer vision in radiology as a relatively mature area of AI in medicine. AI tool use in imaging can improve detection and interpretation, but the shift also reshapes roles toward synthesis and coordination of results. Geoffrey Hinton’s early predictions that radiologists would be replaced are cited as an example of how expectations about automation evolve. The takeaway is that AI supports clinicians in radiology, enabling more efficient workflows, while human expertise remains essential for final interpretation and clinical decision making.

Patients, privacy and liability

Chen and Lichtman discuss patient perspectives, noting that patients may welcome AI assistance but require transparency about how AI is used in care, what questions AI can answer, and how their data is handled. Privacy concerns are prominent, with warnings against inputting real patient information into consumer AI tools, given the risk of data leakage and potential malpractice liability. The panel also considers the regulatory landscape, including FDA oversight and evolving liability norms for both clinicians and technology providers.

Education and the future of medical training

Medical students and trainees are discussed as both beneficiaries and potential risks. There is hope that AI can function as an excellent tutor, but concerns remain about overreliance eroding clinical reasoning. The Stanford example describes how students performed differently on open book vs closed book assessments when guided to use AI with guardrails. The overall message is that AI should be integrated with clear guardrails, ethical considerations, and ongoing evaluation of outcomes in training programs.

Conclusion

The podcast closes by reiterating that AI in medicine is a dynamic, rapidly changing space. The guest emphasizes responsible use, source verification, and maintaining professional judgment as central to benefiting patients. This conversation outlines a path toward broader adoption of AI in health care that preserves patient safety, privacy, and trust while accelerating clinicians' capabilities to provide up to date, evidence based care.