Claude Wrote the Literature Review. Scite Caught the Two Claims That Would Have Sunk It.

Claude drafted my literature review in an afternoon, and the draft was genuinely good — balanced, readable, structured like something a reviewer would nod along to. The citations looked plausible. That's precisely the danger, and Scite exists for the morning after.

I ran every reference through Scite's citation context search. Most checked out. Two did not: one study Claude cited for a moderate claim was actually contradicted by a larger follow-up, and another was real but cited for nearly the opposite of what it found. Both errors were invisible in the prose. The prose was that good.

That's the failure mode worth understanding: strong models cite plausibly, not verifiably. Claude had no way to know what later studies said about the papers it referenced, and I had no way to know without a tool that tracks citation sentiment across the literature.

The workflow is now non-negotiable for anything I publish: Claude drafts, Scite vets every citation, Claude revises. Total overhead is about an hour on a multi-day project. The alternative — trusting fluent prose because it reads like expertise — is how a confident paragraph becomes a retraction.