A wind sensor is not calibrated and reports wind speed values that are much higher than the actual wind speed.

One annotator labels an object as “mobile crane”, while another annotator labels the same type of object as “crane truck". 

A site layout image was collected six months ago, but the access routes and material storage areas have now changed.

A dataset is downloaded from an unknown online source, and there is no information about how the data were collected or labelled.

Some construction site images are missing because the camera stopped recording for two hours.

A model is trained using thousands of past concrete pour records, where each record includes concrete volume, crew size, pump capacity, site access condition, and the actual pour duration. The model learns to predict pour duration for future projects.

A path-planning AI for an autonomous site robot learns to move from the material storage area to the workface. It tries different routes in a simulated site, receives positive rewards for shorter safe paths, and penalties for collisions or entering exclusion zones.

A model is trained using a large number of unlabelled construction site photos. Parts of each image are hidden, and the model learns to predict or reconstruct the missing parts before being used for other visual tasks.

An excavator records fuel consumption and travel distance every 10 minutes. There are no manual labels, but the system identifies unusual operating patterns, such as high fuel use with very low movement.

A robotic arm learns how to pick and place construction components by trying different actions in a simulated environment. It receives rewards when the component is placed correctly and penalties when it collides or fails.

A model is trained using site images that have already been labelled as “crack”, “spalling”, “corrosion staining”, or “no visible defect”. The model learns to predict the correct defect category for new images.

A system identifies crane lifting cycles that take much longer than normal historical lifting operations.

A model predicts the expected duration of a concrete pour based on concrete volume, pump capacity, crew size and site access conditions.

A model predicts the final project delay in days based on current progress, labour availability and material delivery records.

A site image is analysed and labelled as containing a loader, mobile crane, concrete pump or concrete mixer.

An LLM drafts a toolbox talk for workers installing trench shoring near active excavation equipment.

A concrete surface image is classified as showing cracking, spalling, corrosion staining or no visible defect.

A project team uses an AI defect-detection tool designed to flag possible cracks in concrete images. Because the tool is usually accurate, they stop doing human review and automatically reject subcontractor work whenever the AI reports a defect.

A company originally installs AI site cameras for safety alerts, but later reuses the same camera data to track individual workers’ productivity and discipline workers, even though this was not the stated or validated purpose.

A company collects workers’ safety observation reports for hazard prevention, but later uses the same reports to create individual “risk scores” for workers and decide who should receive fewer shifts. This use was not disclosed to workers.

A site team uses an AI-generated toolbox talk without checking it. The LLM includes a safety instruction that does not match the actual excavation condition, but the team still follows it because “the AI said so.”

A site supervisor uses an AI hazard-detection camera to receive alerts about workers entering exclusion zones. The supervisor still checks the site condition before deciding whether to stop work.

A contractor uses an LLM to draft a toolbox talk based on project-specific safety notes. The site manager reviews the draft, checks whether it matches the actual site condition, and edits it before sharing it with workers.