Medicine for artificial intelligence: applying a medical framework to AI anomalies

Area of Science:

• Artificial Intelligence
• Medical Informatics
• AI Safety

Background:

• Current approaches to AI failures like "hallucination" are fragmented and lack standardization.
• This hinders cumulative knowledge building and reproducible management strategies for AI systems.
• A unified framework is needed to address AI anomalies systematically.

Purpose of the Study:

• To propose Medicine for Artificial Intelligence (MAI), a novel clinical framework for AI anomalies.
• To adapt medical nosology and clinical workflows to the AI lifecycle.
• To establish a reproducible foundation for safer, more resilient, and auditable AI systems.

Main Methods:

• Reconceptualized AI anomalies as diseases within a clinical framework.
• Formalized core medical constructs (disease, symptom, diagnosis, treatment) for AI.
• Developed DSA-1, a prototype taxonomy of 45 AI disorders across nine functional chapters.
• Mapped a clinical workflow (examination → diagnosis → intervention) onto the AI lifecycle.

Main Results:

• The MAI framework clarifies ambiguous AI failure modes, such as distinguishing hallucination subtypes.
• Linked AI diagnoses to actionable interventions and standardized evaluation metrics.
• Enabled lifecycle practices like triage and "AI health checks" for AI systems.
• Integrated epidemiological and risk-assessment constructs for AI governance.

Conclusions:

• MAI provides a reproducible methodology for diagnosing and managing AI anomalies, akin to clinical medicine.
• This framework facilitates collaboration between AI researchers, clinicians, and regulators.
• MAI supports the development of safer, more resilient, and auditable artificial intelligence systems.