Governance · 6 min read

Human review in AI-assisted clinical research

Where language models can help with protocol work, where they shouldn’t be used, and how to design approval so that review is real rather than a rubber stamp.

Narrow the task

Proposing structured criteria from protocol text is a bounded task with a checkable output: every proposal can be compared with the passage it cites. Deciding whether a person is eligible is not a task to hand to a model.

Verify before showing

A proposal that cites text that doesn’t exist in the document should never reach a reviewer. Verifying citations mechanically removes a whole class of errors.

Design for real review

Show the source beside each proposal, flag conflicts, avoid uncalibrated confidence scores, and require the approver to attest that they compared the criteria with the protocol. Two-person approval helps when the drafter and approver would otherwise be the same person.