The Substantiation Gap

Updated 2026-07-13 · markdown version

The substantiation gap is the distance between generating a claim and proving it: generative AI produces fluent, plausible, regulation-shaped sentences, but a sentence's fluency carries no information about whether its content exists in a source. In ordinary marketing, that gap is an embarrassment risk. In pharmaceutical promotion — where every claim must be verifiable against evidence at the page/column/paragraph level — the gap is the entire regulatory exposure of the industry, concentrated in one place.

The gap has a precise shape. A model writing "reduced LDL-C by 42% (p<0.001)" performs three acts at once: composing language (which it does superbly), asserting a fact (which it cannot check), and implying provenance (which it does not have). Reviewers unwind those three acts claim by claim — that's what substantiation review is — so content that arrives with the gap unclosed simply transfers the closing cost to the most expensive people in the pipeline: committee reviewers and signatories at the MLR wall.

What makes the gap strategic rather than incidental: generation scales; the gap scales with it. Every one-button gain in drafting speed is a proportional increase in unverified claims arriving at review. The gap cannot be closed by better prose. It closes only from the evidence side — by linking claims to source passages at writing time, so provenance travels with the sentence. Tools built that way (sourced-only, evidence-linked — the co-creation shape) don't narrow the gap; they refuse to open it.

A working test for any AI medical-writing tool: ask where a given generated claim came from. If the answer is a source passage, the gap is closed. If the answer is a confidence score, it isn't.