The AI Doctor Is Here — Here Is Exactly What It Can and Cannot Do

AI systems are now handling triage, diagnosis support, and patient management in real clinical settings. Here is an honest assessment of what these systems do well and where they fail dangerously.

The healthcare AI landscape in 2026 encompasses several distinct categories of AI system whose clinical capability and appropriate use cases vary dramatically, and whose conflation in media coverage creates both unwarranted optimism and unwarranted fear. Understanding the categories is the prerequisite for understanding what AI actually does in healthcare settings.

Diagnostic imaging AI — systems that analyse X-rays, CT scans, MRIs, and pathology slides for specific findings — is the most clinically validated category. Multiple randomised controlled trials have demonstrated specific AI systems achieving radiologist-level sensitivity for specific conditions: lung nodule detection, diabetic retinopathy, skin lesion malignancy, and colonic polyp identification have all been demonstrated at clinical benchmark performance. These systems are deployed at scale in several countries as radiologist extenders — screening large volumes of images to identify those requiring urgent review.

Clinical decision support AI — systems that analyse electronic health record data to identify patients at risk of specific conditions (deterioration, readmission, sepsis) — has shown promising results in studies and is deployed in several healthcare systems, but has also produced several well-documented cases of disparate performance by race and ethnicity — systems performing better for populations well-represented in their training data and worse for underrepresented groups.

Conversational AI for patient triage and symptom assessment — the 'AI doctor' that patients interact with directly — is the most publicly visible and most controversial category. UK NHS trials of conversational AI for initial triage, Israeli health system deployments, and US health system chatbot rollouts have produced cautious assessments: these systems appropriately handle straightforward presentations and effectively direct patients to appropriate care levels, but have produced safety concerns when encountering presentations where the most serious diagnosis is atypical or where the history the patient provides is incomplete.

For the overall assessment: AI is becoming a genuine tool in healthcare that improves specific outcomes in specific applications. It is not replacing physician judgment in the complex, contextual, relationship-dependent aspects of clinical medicine that constitute most of healthcare's value.