Generative AI has captured everyone’s imagination. ChatGPT, Claude, Gemini—these systems can produce remarkably human-like text. Naturally, healthcare leaders ask: can we use this for clinical documentation?

The answer is complicated. Let me share what I’ve learned.

The Documentation Burden Is Real

Let’s start with the problem. Clinicians spend far too much time on documentation. Studies suggest doctors spend two hours on documentation for every hour of patient contact. Nurses face similar burdens.

This isn’t bureaucratic waste—it’s patient safety, communication, and legal protection. Documentation matters. But the volume has become unsustainable.

If generative AI could reduce documentation burden, it would be genuinely valuable. The question is whether it can do so safely.

What Generative AI Can Do Today

Ambient clinical documentation. Systems that listen to clinical encounters and generate draft notes are the most promising application. Products like Nuance DAX (now Microsoft) and various startups offer this capability.

How it works: The system records the clinical consultation (with patient consent), uses speech recognition to transcribe, then uses generative AI to structure the transcription into a clinical note format.

Early implementations show promise. Clinicians review and edit AI-generated drafts rather than writing from scratch. Time savings of 30-50% are reported.

Letter and referral drafting. Generating draft referral letters, discharge summaries, and patient correspondence based on clinical data. The AI structures and phrases; the clinician reviews and approves.

Template completion. Filling in standardised documentation templates based on clinical context, reducing repetitive data entry.

Summarisation. Condensing lengthy patient histories into concise summaries for handover or referral purposes.

What Generative AI Can’t Do Reliably

Clinical decision support. Generative AI isn’t designed to provide clinical recommendations. It generates plausible text, not accurate clinical guidance. Using ChatGPT to suggest diagnoses or treatments is dangerous.

Medical coding. Clinical coding for MBS, PBS, and casemix requires precision. Generative AI makes errors that are hard to detect and have financial and compliance implications.

Structured data extraction. Converting narrative documentation to structured data fields is unreliable. AI extracts plausible information, not necessarily accurate information.

Handling rare conditions or edge cases. Generative AI performs best on common patterns. Unusual presentations or rare conditions produce less reliable output.

The Hallucination Problem

Here’s the fundamental challenge: generative AI sometimes generates text that sounds correct but is factually wrong. In other contexts, this is a minor annoyance. In healthcare, it’s a patient safety issue.

An AI-generated clinical note might:

• Include findings that weren’t observed
• Omit findings that were observed
• Misrepresent patient-reported information
• Confuse current and historical information
• Fabricate medical history details

These errors are often subtle. The text reads fluently. Without careful review, they could become part of the medical record.

This is why generative AI for clinical documentation requires human review. Always. Without exception.

Current Regulatory Position

The TGA hasn’t issued specific guidance on generative AI for clinical documentation. It’s an evolving space.

Key considerations:

• Documentation aids (helping clinicians write faster) face lighter regulation than clinical decision support (recommending actions)
• Products that remain clearly advisory with mandatory human review have clearer pathways
• Products that blur the line between documentation and decision support face harder regulatory questions

AHPRA’s guidance on medical practitioners using AI emphasises that clinical responsibility remains with the practitioner. AI-assisted documentation doesn’t change this—the clinician who signs the note is responsible for its content.

Practical Implementation Guidance

If you’re considering generative AI for clinical documentation:

Start with lower-risk applications. Administrative correspondence, template completion, and summarisation carry lower risk than primary clinical documentation. Build experience before tackling higher-stakes applications.

Mandate human review. No AI-generated clinical documentation should enter the medical record without clinician review and approval. Build workflows that make review unavoidable, not optional.

Monitor error rates. Track how often AI-generated drafts require significant correction. If error rates are high, the time savings might be illusory (clinicians spend as long correcting as they would writing).

Informed patient consent. Patients should know when AI is involved in documenting their care. A simple disclosure in your consent processes is appropriate.

Audit and quality assurance. Periodically audit AI-generated documentation for accuracy. Don’t assume it’s working correctly—verify.

What I’m Watching

Several developments will shape this space:

Specialised clinical models. Generative AI trained specifically on clinical documentation, with healthcare-specific guardrails, may perform better than general-purpose models.

Integration with clinical context. Systems that can access EMR data while generating documentation could be more accurate (because they have more context) or more dangerous (because they might hallucinate about actual patient data).

Regulatory clarity. TGA and international regulators are developing frameworks. Clearer guidance will help organisations understand compliance requirements.

Clinical evidence. We need studies of AI documentation accuracy in Australian clinical settings. Vendor claims aren’t enough.

My Current Recommendation

Generative AI for clinical documentation shows promise, but caution is warranted.

For administrative documentation—letters, referrals, non-clinical correspondence—pilot carefully with appropriate review processes.

For primary clinical documentation—clinical notes, assessments, care plans—proceed very cautiously. The risks of AI errors entering medical records are significant.

For any application, ensure clinicians understand that AI generates drafts, not finished documents. Review isn’t optional.

The technology will improve. Regulatory clarity will emerge. Evidence will accumulate. In two years, my recommendation might be different. But today, enthusiasm should be tempered with appropriate caution.

If you’re working with external partners on AI documentation projects, AI consultants Brisbane and similar firms can help navigate the implementation complexities, though the clinical governance decisions remain yours to make.

Documentation burden is a real problem worth solving. Generative AI might be part of the solution. But not all solutions are ready for deployment, and not all deployments are ready for primetime.

---

Dr. Rebecca Liu is a health informatics specialist and former Chief Clinical Information Officer. She advises healthcare organisations on clinical AI strategy and implementation.