How it works in a data room
Language models generate text by predicting likely words, not by looking up facts. When the relevant passage is missing, ambiguous or not retrieved, a model may fill the gap with something that reads correctly but is false. In a deal room this can appear as a summary stating a term the contract does not contain, an answer citing the wrong page or a number that blends two documents. Grounding answers in retrieved text with retrieval-augmented generation, showing citations and allowing the model to say “not found” all reduce the rate.
Why it matters in a deal
Diligence findings drive price, warranties and the decision to proceed. A hallucinated fact relied on without checking can lead to a wrong valuation, a missed liability or a misleading answer to a bidder. Teams should treat AI output as a lead to verify, record which findings were checked against source documents and prefer tools that show where every statement came from.
Example
A reviewer asks an assistant whether a supply agreement includes a price adjustment mechanism. The assistant says yes and quotes a clause. When she opens the cited page, the clause belongs to a different agreement with a similar name. She flags the error, and the team adds a rule that every AI-sourced finding in the diligence report must carry a verified page reference.