LM
Larissa Meredith-Flister
/

Catching Plausible but Wrong AI Inferences

0 views

We all talk about the risk of AI hallucinating cases that don’t exist. There is a quieter risk that gets less attention: the inferences AI draws that do not immediately raise red flags. A date that was never in the document. A rule reconstructed from training data rather than the statute. A paragraph reference that looks right. Plausible but wrong, and by the time you notice, it is in your draft.

I’ve created a skill to address this, now available on Lawve AI.

It forces a different mode. Claude answers only from materials you provide or online sources it has actually accessed. Every claim is anchored to a pinpoint reference and labelled in one of five categories: expressly stated in materials, verified online, supported but not expressly stated, not found, or flagged inference. Gaps stay as gaps. “Not found” is treated as a correct answer, not a failure.

Built for document review, chronologies, citation checks, evidence summaries, and any drafting task where fidelity to the record matters more than a complete-sounding answer.