What large language models can and can't draft for a trial

Large language models drafting clinical trial documents has moved quickly from a speculative idea to an active area of applied research. Work exploring retrieval-augmented generation, feeding a model relevant source documents alongside the writing task rather than relying purely on what it learned during training, for clinical trial document drafting gives a genuinely more specific picture of what these tools can do well, and where they still need close supervision.

Why retrieval-augmented generation matters here specifically

A general-purpose language model, asked to draft a protocol section from scratch, risks producing plausible-sounding text that isn't actually grounded in the specific study's real data, prior documents, or the applicable regulatory guidance. Retrieval-augmented approaches change the task: the model is given the actual relevant source material, previous protocol versions, study data, specific guidance documents, and asked to draft from that material rather than from memory alone.

This matters because it shifts the main risk. A model working from retrieved source material is less likely to fabricate a plausible-sounding but incorrect fact, and more likely to fail in a different, more manageable way: missing something the retrieval step didn't surface, or synthesising the retrieved material imperfectly.

What this kind of system was found to do well

Document drafting tasks that are substantially about synthesising and structuring existing information, rather than generating novel scientific claims, are where these approaches perform most reliably:

• Producing a structured first draft of a document section from source material that already exists, saving genuine time on the mechanical work of assembling and formatting information that's already been decided.
• Maintaining consistency of terminology and structure across a document, something that's tedious for a human writer to enforce manually across a long protocol.
• Surfacing the specific source passages a drafted section was based on, which supports the kind of verification a human reviewer actually needs to do their job.

Where it still needs close human oversight

The same work is equally clear about where these tools fall short of being trustworthy without review:

• Judgement calls that require genuine scientific or regulatory expertise. A model can draft plausible text; it can't make the substantive decision about whether a particular endpoint or eligibility criterion is the right one.
• Catching what the retrieval step missed. If a relevant piece of source material wasn't retrieved for a given section, the model has no way of knowing that, and the resulting draft can look complete while missing something important.
• Anything genuinely novel to this specific study, rather than synthesised from existing material, still needs to originate from human expertise, not the model.

What this means for a study team considering these tools

1. Use retrieval-augmented approaches over general-purpose drafting, specifically because grounding the model in real source material meaningfully reduces the fabrication risk that general drafting carries.
2. Treat any AI-drafted section as a structured first draft, not a finished one, requiring the same named, accountable review any other first draft would need.
3. Verify that retrieval actually captured the relevant source material, rather than assuming a complete-looking draft means nothing was missed.
4. Reserve genuinely novel scientific and regulatory judgement calls for human authors, using AI assistance for the surrounding structural and synthesis work instead.

The honest state of the technology

The research picture here is more encouraging than a blanket "don't trust AI with regulatory documents" position, and more cautious than "AI can draft your protocol." Retrieval-grounded drafting genuinely reduces one of the biggest risks, fabricated claims presented confidently, that made earlier general-purpose drafting approaches risky for this kind of document. It doesn't remove the need for expert human judgement on the substance, and it introduces its own specific failure mode, incomplete retrieval, that a reviewer needs to actively check for rather than assume away.