AI Applications by Safety Category
- Safe for individual use: Documentation, patient education materials, research summaries
- Safe with institutional approval: Clinical decision support, validated diagnostic tools
- Requires regulatory clearance: Diagnostic AI, treatment recommendations
- Not appropriate: Using ChatGPT/Claude for clinical decision-making
- Always: Human clinician retains responsibility for all clinical decisions
The Biggest Win: Reducing Documentation Burden
The most impactful — and safest — AI application in healthcare is administrative documentation. Doctors spend an estimated 2-4 hours per day on clinical notes, referral letters, and administrative correspondence. AI scribing tools are transforming this.
- Listens ambient, no dictation required
- Integrates with major EHR systems
- HIPAA compliant with BAA available
- Produces structured clinical notes for physician review
FDA-Cleared Diagnostic AI — What's Actually Approved
Over 700 AI-enabled medical devices have received FDA clearance or approval. The most established areas include:
- Radiology: AI tools that flag potential pneumothorax, pulmonary embolism, and intracranial haemorrhage in imaging — giving radiologists priority reading queues.
- Ophthalmology: IDx-DR (now Digital Diagnostics) — the first FDA-cleared autonomous AI diagnostic system, for diabetic retinopathy screening.
- Cardiology: ECG analysis AI (AliveCor, Apple Watch) cleared for atrial fibrillation detection.
- Pathology: AI-assisted cancer detection in histopathology slides, improving accuracy in high-volume labs.
Safe AI Use for Clinical Professionals
For research and education: Using ChatGPT or Claude to understand medical literature, get plain-language explanations of research papers, or generate patient education materials is generally appropriate. Always verify clinical claims against primary literature.
For administrative tasks: Drafting referral letters, generating template forms, summarising non-patient-specific information, and creating educational content are all appropriate uses of general AI tools.
For clinical decisions: Do not use general-purpose AI tools for clinical decision-making. This is not what they are built or validated for, and the stakes — patient safety — are too high for unvalidated tools.