Hallucination

Why it matters in construction

In estimating, a hallucinated fact can become a margin problem. If a model reads a mechanical proposal and reports “includes seismic bracing” when the sub excluded it, the gap may surface after award. If an RFP summary invents a liquidated-damages figure, the go/no-go discussion starts with a number that does not exist.

The output can look polished. A hallucinated exclusion may be formatted exactly like a real one. A model can state “$12,000” with the same confidence whether the number came from the proposal or not.

How it works

Language models can hallucinate because they generate text instead of looking up verified facts.

1. Prediction, not lookup. The model produces the likely continuation of the text so far. If the true answer is missing, it may still produce a plausible-sounding answer.
2. Gaps in the prompt. When a tool asks “what is the bond rate?” and the proposal never mentions bonding, the model may supply a typical rate rather than “not stated.”
3. Long documents. Important details can be missed in long documents. A 40-page proposal with exclusions on page 31 needs careful retrieval and review.
4. Confidence is not proof. The model may use the same tone whether it found the fact or invented it.

Give the model source documents instead of relying on memory. Require it to quote the passage it used. Make “not found” a valid answer. Before using the tool on live bids, test it on proposals with known answers and measure its misses.

Example in practice

Consider a GC’s estimating team piloting an AI leveling tool on a $4.2M electrical package with six proposals. On the first pass, the tool reports that all six include temporary power. The estimator checks the sources. Four say it explicitly, one is silent, and one excludes it in a footnote on page three. Two of the six answers were hallucinated.

The team reconfigures the tool to require a quoted passage for every inclusion and to return “not stated” when nothing turns up. On the rerun, it flags the silent proposal for follow-up and catches the footnote exclusion, worth roughly $65,000 on that job. The team adopts a new rule: no leveling cell is accepted without a visible source.