Language Models in the Tumour Board

Language models are arriving in multidisciplinary meetings through the side door — summarising notes, drafting documentation, preparing case lists. Some of that is genuinely useful and some of it is a liability, and the line between them is fairly clear.

Where summarisation helps

• Assembling a case list from scattered documents ahead of the meeting.
• Drafting the administrative record of a decision already made and stated.
• Surfacing documents a clinician then reads — retrieval, not conclusion.
• Translating a recorded decision into patient-facing language for later review.

Where a human must stay in the loop

• Any clinical fact that will be acted on. Stage, biomarker status, prior treatment and performance status must be verified against the source, never accepted from a summary.
• Anything the model asserts without a citable span. Fluent output carries no signal about whether it is correct.
• Negation and uncertainty. “No evidence of” and “cannot exclude” are routinely flattened in summarisation.
• The recommendation itself. A model does not carry clinical responsibility, and the clinician who acts on it does.

Two practical rules

1. Every clinical claim in a generated summary must be traceable to the span of text it came from.
2. No patient-identifiable information into any system without a data-processing agreement covering it.