Most writing about AI and construction risk is vendor material in which the tool never fails, or reviews in which nobody ran a real project through a model.

A January 2026 study in Buildings, selected as an Editor's Choice paper, is in a more useful place.¹ Researchers at the University of Zagreb's Faculty of Civil Engineering gave the full 606-page technical specification for a real judicial complex to three general-purpose AI models, with no fine-tuning and no preprocessing, and asked each to identify risks phase by phase, then put the outputs in front of three construction risk practitioners in Croatia, Germany, and the United States.

One caveat throughout: the models tested have since been superseded, so anything you open today will do this better. Read the ranking as a snapshot. The methodology, the expert reaction, and the failure modes are the durable part.

The setup matters more than the results, because the setup is the part you can copy.

Three off-the-shelf models, no domain training, no custom tooling. The test was what a construction professional could do on a Tuesday with a subscription and a PDF.

The work followed the European Commission's PM² methodology and its four phases: Initiating, Planning, Executing, Closing. Each model got a fresh session, a context-setting prompt, then one prompt per phase asking for a seven-column table: Code, Risk Source, Cause/Driver, Risk, Affected Area, Risk Category, Risk Bearer.

All four phases ran in one continuous session, so the model carried forward what it learned earlier, the way an analyst who has read the whole package gets sharper as they go.

The seven columns map onto a PM² Risk Log; the cause-risk-effect phrasing matches how PM² tells you to write a risk statement. Nothing needed reformatting afterward.

A second round then asked the model to re-cut the same risks by category and by stakeholder, without generating anything new. That is closer to how risk information actually gets consumed.

The GC does not want the register. The GC wants their rows of it.

The 26-versus-240 gap is the number everyone quotes. It measures how each model read an open instruction about granularity, not how much each one found. One consolidated "unforeseen ground conditions" into a single register line. Another split the same ground into four entries with individual write-ups. Both are legitimate readings.

Each model ran once, on one project, with one phrasing, and no ground-truth register existed to score against. So specify granularity yourself, or you get defaults that will change next quarter.

Two rated AI "very useful" for initial risk identification, one "extremely useful", and all three endorsed a hybrid approach without hedging. On trust, all three said only "moderately trustful”. Not one moved above that.

That pairing is the finding. It is useful the way a first-year engineer's memo is useful. You are glad to have it, but you read every line before it goes anywhere.

Preference split cleanly: