AI is making real and significant contributions in several medical domains:

Diagnostic imaging: AI systems for radiology, pathology, and dermatology are achieving performance that matches or exceeds specialist physicians on specific, well-defined classification tasks — detecting diabetic retinopathy, classifying skin lesions, identifying pneumonia on chest X-rays. These are genuinely impressive results on narrow tasks. They are also being implemented as decision support tools for radiologists and pathologists — not replacements.

Clinical documentation: AI scribing tools (Nuance DAX, Abridge, Nabla) are reducing the documentation burden that has made physician burnout a systemic crisis. These tools listen to patient encounters and generate structured clinical notes, with studies showing meaningful reductions in after-hours documentation and clinician burnout, though time savings vary by tool and setting. This is unambiguously good for physicians — it returns time to patient care.

Drug discovery and research: AI is accelerating the identification of drug candidates, protein structure prediction, and clinical trial design in ways that will significantly change pharmaceutical research. This affects researchers more than practicing clinicians.

Predictive analytics: AI models analyzing EHR data can predict sepsis risk, readmission risk, and deterioration earlier than traditional monitoring. These tools enhance physician capability; they do not make physician judgment redundant.

The aspects of medical practice that remain most human-dependent:

The clinical encounter: A complete medical encounter requires taking a history (which depends on building enough trust that the patient discloses everything relevant), performing a physical examination, synthesizing findings in the context of the patient's whole situation, and communicating a diagnosis and plan in a way the patient can act on. This is not a discrete information-processing task — it is a complex, relational, embodied clinical skill that AI cannot perform.

Clinical judgment in complex or atypical cases: AI systems perform well on cases that resemble their training data. Medicine's most consequential moments are the atypical cases — the presentation that doesn't fit the pattern, the patient whose symptoms suggest multiple competing diagnoses, the case where standard treatment isn't working. These require the kind of flexible, contextual clinical reasoning that experienced physicians develop over years and that AI systems do not reliably produce.

Therapeutic relationship and communication: Patients facing serious illness need human presence — someone who will sit with them, explain clearly, answer questions, acknowledge uncertainty honestly, and help them make decisions aligned with their values. No AI system can provide this, and it is not a peripheral benefit of medicine — it is the core of what good medicine does.

Surgical and procedural skill: Physical procedural competence — surgical technique, catheterization, endoscopy, lumbar puncture — requires the kind of trained physical skill that current robotics cannot match across the full range of medical procedures in real clinical environments.

AI's impact is uneven across specialties:

Highest change: Radiology and pathology face the most direct pressure from AI diagnostic tools — though the evidence suggests augmentation rather than replacement, as AI catches things humans miss and humans catch things AI misclassifies. Both fields are adapting by emphasizing the clinical integration and communication skills AI cannot provide.

Moderate change: Primary care physicians will see documentation burden reduce substantially. Diagnostic decision support tools will become standard. The core clinical encounter — relationship, examination, complex judgment — remains human.

Lowest change: Surgeons, emergency medicine physicians, and proceduralists whose work is primarily physical and real-time. Psychiatrists whose therapeutic relationship is the treatment. Oncologists managing complex, emotionally charged treatment decisions. Pediatricians whose work depends on developmental knowledge and family relationship.

The physicians best positioned in the AI era:

Embrace AI tools that reduce administrative burden: Clinical documentation AI is the most direct opportunity — using it frees time for patient care and reduces burnout. Physicians who resist AI documentation tools are choosing to spend time on tasks that AI does well rather than on the clinical work that requires them.

Deepen the skills AI cannot replicate: Communication skills, physical examination skills, and the ability to manage complex, emotionally demanding patient relationships are the most AI-resilient clinical competencies. These are worth developing deliberately throughout training and practice.

Develop AI literacy: Understanding what clinical AI systems do well and where they fail is a patient safety competency. A radiologist who understands how AI classifies images can catch the cases where the AI is wrong. A primary care physician who understands predictive risk models can use them appropriately without over-relying on them.

For medical students: The physician skills most worth developing are the ones AI cannot replicate — the clinical encounter, complex reasoning in atypical cases, and therapeutic relationship. These are also, not coincidentally, the skills that make medicine meaningful to most physicians. The AI era is not a reason to avoid medicine; it's a reason to develop the distinctly human aspects of clinical care more deliberately.