Generative AI Revolutionizes Safety Monitoring on Construction Sites

A Deadly Reality in Construction

Last winter, a tragic accident occurred during an affordable housing project on Martha’s Vineyard, Massachusetts, where a 32-year-old worker, Jose Luis Collaguazo Crespo, fell from a ladder to his death. This incident highlights a grim reality: over 1,000 construction workers die annually in the US due to slips, trips, and falls, making construction the most dangerous industry for such fatal accidents.

Balancing Safety and Productivity

Philip Lorenzo, an entrepreneur and executive at DroneDeploy, spoke at Construction Innovation Day 2025 about the challenge of prioritizing safety over productivity on job sites. Despite the common rhetoric that 'safety is the number-one priority,' shortcuts and risk-taking persist. To address this, DroneDeploy developed Safety AI, a tool that analyzes daily digital models of construction sites—created from videos and images via "reality capture"—to detect violations of OSHA safety regulations with 95% accuracy.

How Safety AI Works

Safety AI is unique in employing generative AI, specifically a type of visual language model (VLM), which combines language understanding with image analysis. Unlike traditional object detection methods that only recognize objects like ladders or hard hats, Safety AI can interpret scenes, reason about safety risks, and flag violations accordingly. It can, for example, assess whether a worker is using a ladder unsafely by analyzing multiple factors such as points of contact and positioning.

The system was trained on a "golden data set" of tens of thousands of images depicting OSHA violations, curated over years with explicit customer permission. Human safety experts guide the AI by asking strategic questions and adjusting inputs to improve the model's reasoning capabilities.

Advantages and Limitations

Visual language models excel in dynamic environments like construction sites, where conditions change daily. However, challenges remain, such as dealing with rare edge cases and the AI's occasional hallucinations or incorrect assessments. To mitigate this, Safety AI incorporates older machine-learning techniques like image segmentation and photogrammetry to build spatial models, enhancing situational awareness.

Despite imperfections, Safety AI offers an invaluable digital "extra set of eyes" for safety managers who often oversee multiple sites simultaneously. It delivers alerts remotely, saving time and potentially improving worker safety.

Industry Perspectives and Worker Concerns

While some in the industry, like Aaron Tan, see Safety AI as a useful tool to support overstretched safety managers, workers sometimes fear such technology could become "bossware," invading privacy and being used punitively. Acceptance will depend on demonstrating tangible safety benefits.

Other companies, such as Safeguard AI and Buildots, utilize more traditional machine learning methods with human intervention to monitor safety or progress on construction sites. These approaches are considered reliable, though often require extensive labeled data and training time.

The Future of AI in Construction Safety

Experts emphasize that AI tools should complement, not replace, human safety inspectors. Verified human oversight remains crucial to ensure that AI-generated alerts are accurate and actionable. As AI technologies evolve, ongoing refinement and independent audits will be essential to increase trust and efficacy in improving construction site safety.