This audio is automatically generated. Please let us know if you have any comments.
Max Mahdi Roozbahani is a senior professor teaching machine learning and natural language processing at Georgia Tech. The opinions are the author’s own.
Imagine a team taking a photo near a job site. In the background, a work truck on the freeway flashes a red light.
A human understands the context: it’s likely a utility vehicle. A generative artificial intelligence system could label the beacon as “police presence”, infer that there was an accident, and summarize the day as an incident. That incorrect tag can show up in a daily report, security log, or claim file months later, with a timestamp and a safe narrative that no one intended to create.
Construction companies are rapidly adopting generative AI copilots to search and summarize project documents, emails and schedules. The goal is speed. With tight margins and chronic labor shortagethe industry is desperate for leverage.
While the build is ready to benefit from AI, you shouldn’t treat AI outputs as gospel or rely on them for your final signature.
Fluency is not a test
I teach machine learning at Georgia Tech and my background spans computer science and civil engineering in both academia and industry. This point of view reveals a risk: teams combine well-written answers with ground truth. In construction, ground truth is what is physically installed and supported by field evidence, including photos, timestamps, and locations. But no single text
Max Mahdi Roozbahani
Permission granted by Max Mahdi Roozbahani
In finance or law, the text is often the reality. A contract is the deal. But in construction, the text is just a proxy. A daily report may be incorrect. A presentation may be out of date. An invoice may be premature. A neat summary can hide a qualifier that matters. A large language model can read this record and still be wrong, because the record itself can be incorrect, incomplete, or out of date.
The danger is greatest in work that becomes invisible once covered: foundations, reinforcing steel, post-tensioning, fireproofing, and critical mechanical, electrical, and plumbing routing. These are the places where mistakes become catastrophic.
Consider seismic risk. A geotechnical report warns of liquefaction and recommends a specific pile count. Later, additional borings update these assumptions and the design is revised to reflect an even larger stack count. If a generative AI assistant is asked, “Does the foundation plan deliver?” it could roll back a previous revision, omit a conditional statement, and reduce the engineering nuance to a definite “Yes.”
If this phrase influences a decision, the project drifts from engineering judgment to automated optimism. The first time these assumptions are tested may be after a disaster.
This is not a theoretical concern.
In my research, my collaborators and I created EIDSeg, a dataset designed to assess post-earthquake damage from social media photos. The work required pixel-level labeling, a multiphase protocol, and sustained attention to consistency, because even highly motivated human annotators disagree when the evidence is messy.
The central lesson is uncomfortable but essential: it takes time and rigor to make visual AI reliable even for a very defined task. A live construction project is much more complex, with more ambiguity, more lost context, and higher consequences.
Multimodal AI, meaning models that interpret photos and text together, is a promising direction, but it introduces new failure modes. The context is thin in a single frame.
Based on the previous example, the red beacon on a utility truck is interpreted as a police presence. The police may involve an accident. A temporary condition is confused with a permanent installation. Missed an exit path blocked by a shadow.
The most dangerous pattern is not the AI making mistakes. It’s that the AI makes mistakes with an authorized pitch and speed.
How is AI still useful?
A reasonable objection is clear: if humans have to verify all outputs, what’s the point of AI?
The point is leverage. Many tasks in construction are repetitive: drafting routine communications, summarizing meetings, mapping shipments, and flagging missing attachments. If AI reduces time spent on repetitive work, engineers can spend more time on field checks and verifications.
AI should speed up preparation so professionals can certify reality. This means that building requires training, not just software. Responsible AI is now an engineering skill. Teams must understand data provenance, audit trails, and failure modes. They also need to learn to ask better questions.
Matters of incitement. Vague questions invite vague answers. Safer prompts require citations, list assumptions, and force the system to say what it cannot confirm. These practices reduce hallucinations, but do not eliminate them. Precisely for this reason, human review remains essential.
Leaders should establish a core procurement standard for any AI used in compliance, security or claims. Each response must cite specific sources with revision dates and identifiers. Answers about the condition of the field should bring to light supporting photos and inspection records. When evidence is missing or contradictory, the system should refrain from drawing a conclusion and flag for review.
Generative AI can make building faster. It can help teams find information buried in folders and emails. But the built environment depends on proof, not prose. When chatbots wow, infrastructure pays.
