AI Hallucination Risk in Construction: What Happens When AI Gets It Wrong

The risk of AI hallucination in construction is that a tool will return confident but wrong output, such as an invented spec citation, a misread scope item or a quantity that is not in the drawings. Stanford found legal AI tools hallucinated 17% to 33% of the time, and no construction-specific incident data exists in the sources we found. Verification is the control.

Our take: construction has the same failure mode as law, with fewer published consequences so far. That absence of data is not evidence of safety; it reflects that nobody is counting. Treat the legal evidence as a measured analogue and build checking into the workflow before the first expensive error. For the larger market context, see AI in construction in 2026.

What is AI hallucination in a construction context?

A hallucination is output that reads as authoritative but is false or unsupported. Language models generate plausible text; they do not look facts up unless a tool is built to, and even then the match between source and answer can fail.

Hypothetical examples, labeled as such because we found no documented construction incidents:

• An assistant drafting an RFI response cites "Section 07 84 00, paragraph 3.2.C" for a firestopping requirement that does not exist in the project specification.
• A summarization tool states a submittal was approved when the log shows "approved as noted" with a resubmit requirement.
• A takeoff tool reports a quantity for an item that appears only in a superseded sheet.
• A generated toolbox talk quotes a regulatory requirement that was wrong or out of date.

The common feature is that the error looks like all the correct output around it. That is what makes it costly: it passes casual review.

How often do AI tools get it wrong?

No construction-specific error rate was found. The best-quality analogue is legal research. Stanford RegLab and HAI tested commercial tools designed for legal work, which retrieve real documents to ground their answers. They found hallucinations in 17% of benchmark queries for Lexis+ AI and 33% for Westlaw AI. The Westlaw figure was reported as 34% or more in the original study and 33% after revision, per LawNext's coverage of the redo.

Three caveats. The benchmark is legal, so rates in spec interpretation or RFI drafting could differ in either direction. The study is from 2024, and tools have changed. And the point is not the number but that purpose-built, retrieval-grounded tools still erred at rates a professional cannot ignore.

What happens when professionals skip verification?

Courts supply the clearest record of consequences. A district court in the Southern District of Florida imposed roughly $86,000 in sanctions in August 2025 (the ByoPlanet matter). A California appellate court imposed $10,000 in September 2025 in its first published opinion on AI-hallucinated case citations (Noland). A database maintained by Damien Charlotin counted 1,397 US cases involving hallucinated content as of May 6, 2026. An ABA GPSolo article from 2026 continues the coverage.

This is an analogy, not construction evidence. We found no construction-specific sanctions. But the principle courts apply, that the professional must personally verify what they submit, maps onto construction: an estimator who submits a bid, an engineer who stamps a drawing or a superintendent who signs a daily report owns the content regardless of how it was drafted.

The contract side is open. We searched for AIA contract-document language or guidance on AI use and found none, so do not assume your contract forms address AI-generated content.

How concerned are contractors?

In the Dodge/CMiC survey of 235 US GCs and trade contractors, 57% named reliability and accuracy of AI output as their chief concern, and 54% named data security and privacy. Yet 85% expected AI to reduce time on repetitive tasks. The survey was sponsored by a software vendor, CMiC.

Bluebeam's 2026 survey tells a slightly different story: only 2% of respondents cited lack of trust in outputs as a barrier. The wording differs, and the resolution is probably that contractors worry about reliability but adopt anyway, managing it through review. That makes the quality of the review process the actual risk variable.

What does a real verification protocol look like?

The best published example comes from a vendor and concerns takeoff. ConstructConnect's checklist for reviewing AI takeoffs runs: confirm the complete set and latest addenda; review complex pages first; verify scope per item; check materials and assemblies; compare revision clouds; check units, duplicates and outliers; and confirm risk, pricing and margin.

Its most useful feature is risk calibration. Spot-checking is acceptable for standard counts on clean plans; complex geometry, scope, specs and revisions need full review. Generalizing that logic gives a tiered approach:

Habits that reduce hallucination

• Supply the source text. Models do better summarizing a document you provide than recalling it from training. See prompt engineering for construction.
• Require quotes and locations. Ask for the exact passage and page or section so checking takes seconds.
• Check every citation. This is the same lesson as the court cases.
• Keep a miss log. Record errors by tool and task type; after a few weeks you have your own error rate.
• Control data. Do not paste confidential contract documents into consumer tools without a firm policy.

Why is hallucination harder to catch in construction documents?

Construction documents have properties that make quiet errors likely. Specifications and drawings are long, cross-referenced and revised repeatedly. A statement can be true of the bid set and false of the conformed set. Section numbers follow a standard format, so a made-up reference looks right. And the people who would catch the error are the same overloaded project engineers and estimators the tool was meant to help.

Time pressure compounds this. The reason a team reaches for AI on an RFI response or a scope letter is usually a deadline, which is also when the review step gets cut. The control therefore has to be cheap enough to survive a busy week. Requiring the tool to quote the passage it relied on, with a page or section number, turns a vague "does this sound right?" into a thirty-second lookup.

A worked example of tiered review

Illustrative example, with assumptions stated: a project engineer uses an AI assistant to draft an RFI about a conflict between a mechanical drawing and a specification section. The assistant produces a clear, professional draft that cites a spec paragraph and a detail number. Under a tiered rule, an RFI draft is medium stakes. The engineer opens the specification, confirms the paragraph says what the draft claims, confirms the detail number exists on the current sheet, and notes both in the RFI log. The check takes about five minutes.

If the paragraph turns out not to exist, the team has learned something about that tool on that task type, and the miss log records it. If the same tool is wrong three times in ten drafts, the firm has its own error rate and can decide whether the tool saves time net of checking. That is the arithmetic that vendor survey savings figures leave out.

Which construction tasks carry the most risk?

Match risk to consequence. Estimating tools can mis-scope work, and the cost of an error scales with bid size; see AI construction estimating. Safety content errors carry the highest human stakes, as discussed in AI for construction safety, where vision detectors also have documented limits in poor lighting and crowded scenes.

Agent systems that act with limited human prompting raise the stakes because an error can propagate. Gartner forecasts over 40% of agentic AI projects will be canceled by end of 2027 because of cost, unclear value and weak risk controls, which is a forecast and not a measurement.

What this means for GCs, owners and designers

GCs and subcontractors. Write a one-page AI use rule: allowed tasks, prohibited data, required checks by tier, and who signs off. Few firms appear to have this. RICS found 43% of construction respondents not aligned with its AI standard.

Owners and developers. Ask for disclosure of AI use in submittals, estimates and reports, and ask what verification occurred. Treat AI-drafted content as unreviewed until someone accountable says otherwise.

Designers. Responsibility for code compliance, coordination and sealed documents stays with the professional. An AI summary of a code section is not a code citation.

What to watch through 2027

• First documented construction incidents. An RFI, estimate or spec error traced to AI would move this from analogy to evidence.
• Contract language. Whether AIA or other forms address AI-generated content.
• Retrieval accuracy. Whether newer tools beat the 17% to 33% range in independent tests.
• Insurance. Whether carriers ask about AI use; we found no source yet.
• Standards. Uptake of professional-body AI standards such as RICS's, effective March 2026.