Can AI See What Workers Miss? What 14 Studies Reveal About Computer Vision for Construction Safety

A Camera Caught What the Inspector Missed

In a 2025 experiment, researchers pointed an AI system at real construction site photos and asked it to identify every safety violation. The AI flagged 96 out of every 100 actual hazards --- a detection rate no human inspector could realistically sustain across an eight-hour shift. But it also flagged dozens of things that were not actually hazards: a shadow that looked like an unguarded edge, a reflective vest folded on a railing, a compliant guardrail seen from an odd angle (Wang et al.).

That result captures the central tension of AI-powered safety monitoring in 2026. The technology is remarkably good at catching hazards. It is also remarkably bad at knowing when something is fine. And that gap --- between catching everything and crying wolf --- is the question every safety manager needs to understand before deciding whether this technology belongs on their jobsite.

To get a clear picture of where the science actually stands, we reviewed 14 studies published between 2015 and 2026, drawn from peer-reviewed engineering journals and leading research repositories. What follows is not a sales pitch for AI and not a dismissal of it. It is a practical assessment of what works, what does not, and what it means for the people responsible for keeping workers alive.

What These Systems Actually Do

If you have used a security camera system with motion alerts, you already understand the basic concept. AI-powered safety monitoring uses cameras --- sometimes fixed on structures, sometimes mounted on drones or robots --- to continuously watch a construction site. Software analyzes the video feed and flags potential hazards.

But the technology has evolved well beyond simple motion detection. Today's systems fall into three broad categories, and understanding the differences matters because each has distinct strengths and limitations.

Object detection systems are the most straightforward. These use algorithms trained on thousands of labeled photographs to recognize specific items: hard hats, safety vests, guardrails, ladders, heavy machinery. The most common technology is called YOLO (You Only Look Once), which can scan a video frame and identify every recognizable object in it within milliseconds (Adil et al.; Choi and Greer). Think of it as a very fast, very focused visual checklist. The system knows what a hard hat looks like, scans the frame, and reports whether each worker is wearing one.

Vision-language models represent a significant leap forward. Instead of just identifying objects, these AI systems can describe what they see in plain English and reason about whether a situation is dangerous. Point one at a construction photo and it might report: "A worker on the second-floor scaffolding is not wearing a fall harness, and the guardrail on the east side appears to be missing a midrail." These models --- including commercial systems from OpenAI, Google, and Anthropic --- understand context, not just objects (Chaudhary et al.; Chen and Zou).

Integrated pipelines combine multiple AI approaches into a single system. One study mounted an AI pipeline on a four-legged robot that autonomously walked a jobsite, used one AI model to describe each scene, used another to look up the relevant OSHA regulation, and used a third to generate a written safety inspection report --- all without human intervention (Naderi et al.). Another system combined camera feeds with on-site audio recordings and cross-referenced everything against a database of safety regulations to produce inspection reports (Wang et al.).

The Numbers: How Well Does It Actually Work?

This is where the research gets interesting --- and where safety managers need to pay close attention, because the headline numbers can be misleading without context.

The Best Detection Rates Are Impressive

The highest-performing system reviewed, a framework called SiteShield that combines visual and audio analysis with regulatory cross-referencing, caught 96% of actual hazards in real-world construction site data. It achieved what researchers call an F1-score of 0.82 --- a combined measure of both catching real hazards and avoiding false alarms, where 1.0 would be perfect (Wang et al.).

A robot-mounted inspection system running entirely on free, open-source AI models caught 92.2% of hazards in its best scenario. That system was tested across 442 video frames containing 20 different safety violations spanning four OSHA categories: fall hazards, housekeeping and trip hazards, electrical safety, and PPE compliance (Naderi et al.).

For the specific task of hard hat detection --- the most studied application --- object detection models achieved 93.3% accuracy at simply locating workers and equipment in a frame (Adil et al.). A separate study showed that even general-purpose AI models with no construction-specific training could identify hard hats with roughly 65% accuracy across more than 5,200 images (Choi and Greer).

But the False Alarm Problem Is Real

Here is the number that matters most for practical deployment: when researchers directly compared AI performance to human inspectors on the same set of construction safety violations, the results were revealing. Human inspectors correctly identified hazards with 95.6% precision --- meaning that when a human said "that is a violation," they were almost always right. The best AI model achieved only 18.2% precision on the same task, meaning that roughly four out of every five flags it raised were false alarms (Chen and Zou).

To put that in operational terms: if an AI system generates 100 safety alerts in a shift, roughly 80 of them may not be actual hazards. That volume of false alarms creates a real risk of alert fatigue --- the same phenomenon that causes nurses to ignore beeping monitors and drivers to tune out parking sensors. If your safety team learns to dismiss AI alerts because most of them are wrong, the system becomes worse than useless because it creates a false sense of coverage.

However, the AI's recall --- its ability to catch hazards that actually exist --- was 89.4%, compared to the human inspector's 66.6%. The AI missed fewer real hazards. The human inspector missed more real hazards but almost never flagged something that was not actually a problem (Chen and Zou).

This is not a story about AI being better or worse than human inspectors. It is a story about AI and human inspectors having complementary strengths. The AI catches more things. The human knows which things actually matter. Together, they cover more ground than either one alone.

How You Ask the AI Matters More Than Which AI You Use

One of the most practically important findings across these studies is that the way you configure and prompt an AI system has a bigger impact on its accuracy than which AI model you choose.

Chaudhary et al. tested five of the most powerful commercial AI systems currently available --- Claude-3 Opus, GPT-4.5, GPT-4o, GPT-o3, and Gemini 2.0 Pro --- on the same set of 16 real construction site photographs. When given a simple instruction ("identify hazards in this image"), the average accuracy across all models was poor, with F1-scores around 0.31 --- meaning the systems missed most hazards and flagged many non-hazards. When given step-by-step reasoning instructions (a technique called chain-of-thought prompting), accuracy doubled to an average F1 of 0.64. The improvement was statistically significant (p < 0.001), and under the improved prompting strategy, the differences between the five AI models were no longer statistically significant (Chaudhary et al.).

That finding has direct cost implications. If how you ask the question matters more than which product you buy, then a safety manager using a less expensive AI system with well-designed prompts could outperform a competitor using a premium product with generic configuration.

Sammour et al. confirmed this pattern in a different context. Testing AI on 385 questions from professional safety certification exams, they found that prompt design could swing accuracy by as much as 13.5 percentage points. No single configuration worked best across all topic areas --- the optimal setup for hazard identification questions was different from the optimal setup for emergency response questions (Sammour et al.).

Small, Cheap Models Are Catching Up Fast

For safety managers concerned about cost, one of the most striking findings in the recent research is that small, free, open-source AI models are rapidly closing the gap with expensive commercial systems.

Sahraoui tested a model called Qwen2 VL with only 2 billion parameters --- small enough to run on a laptop --- on construction safety violation detection. Using a technique called prompt ensembling, where the system checks the same image with multiple different prompts and aggregates the results, this tiny model achieved 72.6% accuracy and caught 98% of actual hazards. For comparison, OpenAI's commercial GPT model in a standard configuration achieved only 32.2% accuracy on the same task (Sahraoui).

Adil et al. demonstrated that pairing a fast object detection model with a small vision-language model could achieve useful hazard detection performance while adding only 2.5 milliseconds of processing time per image --- fast enough for real-time video monitoring. Even the smallest model they tested, with only 1 billion parameters, showed a 15-percentage-point improvement when paired with the detection system. The entire setup could run on an edge computing device at the jobsite, with no cloud connection required (Adil et al.).

Naderi et al. built a complete autonomous inspection robot running entirely on open-source models and found it consistently outperformed OpenAI's GPT-4o across all tested scenarios --- at roughly one-tenth the per-image cost. Their system used four different open-source AI models working in sequence, each handling a different part of the inspection pipeline, and every intermediate step was visible for human audit (Naderi et al.).

The trajectory here is clear. In 2015, automated hard hat detection required a specialized research lab and custom-built datasets (Shrestha et al.). By 2020, it required machine learning expertise but could use smaller datasets through active learning techniques (Kim et al.). By 2026, a system running on open-source models on a commodity device can detect hazards, explain its reasoning, cross-reference OSHA regulations, and generate a written report --- all for less than the cost of a commercial software subscription.

What AI Still Cannot Do

The research is equally clear about the limitations, and safety managers should understand these before making purchasing decisions.

Fine-Grained Rule Interpretation Remains Weak

AI systems perform reasonably well at coarse hazard recognition --- "is that worker near an unguarded edge?" --- but struggle with the kind of detailed regulatory interpretation that experienced safety professionals do intuitively. When Chen and Zou tested AI models on specific OSHA rule compliance, accuracy on fine-grained rules dropped below 20%. The models could identify that something looked unsafe but could not reliably determine which specific regulation was being violated or whether an exception applied (Chen and Zou).

Similarly, Sammour et al. found that while AI passed professional safety certification exams with scores of 73-85%, it performed poorly on questions involving mathematical calculations, emergency response procedures, and fire prevention specifics. The error analysis was instructive: 38% of mistakes came from gaps in knowledge, 31% from flawed reasoning, 24% from context and memory limitations, and 7% from calculation errors (Sammour et al.).

The Lab-to-Jobsite Gap Is Real

Most of the studies reviewed were conducted in controlled environments or on curated image datasets. Real construction sites present challenges that laboratory conditions do not: heavy dust, dramatic lighting changes throughout the day, workers obscured behind equipment or materials, constantly changing site layouts, and the sheer visual chaos of an active jobsite.

Chharia et al. directly addressed one of these challenges --- visual occlusion --- by using four camera angles instead of one. Their system's accuracy jumped from 81.7% with a single camera to 92.0% with four cameras. Hard hat detection, the scenario most affected by occlusion (workers bending down, turning away, standing behind equipment), improved by 10.3 percentage points with multiple viewpoints. They also developed a synthetic scene generator that creates realistic construction environments for training AI systems, reducing the need for expensive real-world data collection (Chharia et al.).

But multiple cameras, edge computing devices, and the infrastructure to connect them all represent real costs and real complexity. The gap between "this worked in our experiment" and "this works on your jobsite every day" is where many promising technologies stall.

Privacy, Liability, and Worker Acceptance

The research papers focus almost entirely on technical performance. They do not address the practical questions that will determine whether this technology is actually adopted: Do workers accept constant camera surveillance? What happens when an AI system misses a hazard and someone is injured --- who is liable? How do you handle the data that these systems collect, especially if it captures workers making mistakes? What are the union implications?

These are not hypothetical concerns. Any safety manager deploying an AI monitoring system will need to navigate worker consent, data retention policies, and the legal question of whether AI-generated hazard alerts create a documented "knowledge" that increases liability exposure if not acted upon. The technology is ahead of the policy framework, and the research literature has not caught up to these operational realities.

The Real Opportunity: AI as Your Second Set of Eyes

If there is a single takeaway from these 14 studies, it is this: AI is not coming for the safety inspector's job. It is coming for the safety inspector's blind spots.

The data consistently shows that AI and human inspectors have almost perfectly complementary capabilities. Humans are precise --- when they identify a hazard, they are almost always right, and they understand the regulatory context intuitively. AI is thorough --- it watches everything, never gets tired, and catches hazards that humans walk past. The combination of high human precision and high AI recall is more powerful than either alone.

Liu et al. found that more than half of construction site hazards go unrecognized due to gaps in inspector experience and knowledge, particularly among less experienced personnel. Their work on using augmented reality to transfer expertise from experienced inspectors to newer workers points toward a broader theme: the real value of AI in safety is not replacing expertise but distributing it (Liu et al.).

For the practical safety manager, the research suggests the following approach:

Start with PPE compliance monitoring. Hard hat and safety vest detection is the most mature, most validated application. It works, it produces measurable results, and it is the easiest to validate against your own observations. If a vendor cannot demonstrate reliable PPE detection on your specific site conditions, their more advanced features are not ready either.

Invest in configuration, not just procurement. The research consistently shows that how you configure an AI system determines its value more than which system you buy. If you are evaluating AI safety tools, ask vendors about their prompting strategies, their false-positive rates under real conditions, and whether they allow you to customize the system's analysis approach for your specific site hazards. A well-configured mid-tier system will likely outperform a poorly configured premium one.

Design for human review, not automation. Build your workflow so that AI flags go to a competent person for evaluation, not directly into enforcement actions. The technology's strength is catching things humans miss, not replacing human judgment. Track your false-positive rate over time --- if it is above 50%, the system is creating more noise than signal and needs reconfiguration.

Watch the cost curve. The research shows that open-source models running on edge devices are approaching commercial performance for specific safety tasks. If budget is a barrier today, it is worth revisiting in twelve months. The cost of effective AI safety monitoring is dropping faster than almost any other safety technology in recent memory (Oliveira et al.; Pour Rahimian et al.; Adil et al.).

Consider multiple cameras. If you do deploy a vision-based system, the research from Chharia et al. makes a strong case for multiple camera angles. A four-camera setup detected hazards at 92% accuracy versus 82% for a single camera. The improvement is especially pronounced for PPE violations, where a single camera angle often cannot see whether a worker is compliant (Chharia et al.).

Where This Is Heading

The pace of improvement in this field is extraordinary. In 2015, the state of the art was detecting whether a blob of color on a construction site was a hard hat. In 2026, AI systems can watch a jobsite, describe what each worker is doing, identify which OSHA regulations apply, and generate a written inspection report --- using free software running on a device that fits in your pocket.

The research does not yet show a system ready for unsupervised deployment. Every study reviewed acknowledges the need for human oversight. But the trajectory is unmistakable, and the question for safety managers is not whether AI will become part of construction safety monitoring, but when and how you integrate it into your program.

The studies reviewed here suggest that "when" may be closer than most safety professionals realize.

Limitations of This Review

This article reviewed 14 studies available through academic journals and preprint repositories as of April 2026. It does not include proprietary industry research, unpublished pilot programs, or commercial product performance data. Most studies were published between 2024 and 2026, and earlier work may have been superseded. Many studies rely on curated datasets that may not represent the full diversity of construction environments, worker populations, or regional regulatory frameworks. Field validation under authentic, long-term jobsite conditions remains limited across the literature. This is a narrative research review, not a formal statistical meta-analysis; the studies reviewed use different metrics, datasets, and evaluation methods that preclude direct statistical comparison.

Works Cited

Adil, Muhammad, et al. "Integration of Object Detection and Small VLMs for Construction Safety Hazard Identification." arXiv preprint, arXiv:2604.05210, 6 Apr. 2026, doi.org/10.48550/arXiv.2604.05210.

Chaudhary, Nishi, et al. "Prompt to Protection: A Comparative Study of Multimodal LLMs in Construction Hazard Recognition." arXiv preprint, arXiv:2506.07436, 9 Jun. 2025, doi.org/10.48550/arXiv.2506.07436.

Chen, Xuezheng, and Zhengbo Zou. "Are Large Pre-trained Vision Language Models Effective Construction Safety Inspectors?" arXiv preprint, arXiv:2508.11011, 14 Aug. 2025, doi.org/10.48550/arXiv.2508.11011.

Chharia, Aviral, et al. "Safe-Construct: Redefining Construction Safety Violation Recognition as 3D Multi-View Engagement Task." arXiv preprint, arXiv:2504.10880, 15 Apr. 2025, doi.org/10.48550/arXiv.2504.10880.

Choi, Lucas, and Ross Greer. "Evaluating Cascaded Methods of Vision-Language Models for Zero-Shot Detection and Association of Hardhats for Increased Construction Safety." arXiv preprint, arXiv:2410.12225, 16 Oct. 2024, doi.org/10.48550/arXiv.2410.12225.

Kim, Jinwoo, et al. "Towards Database-Free Vision-Based Monitoring on Construction Sites: A Deep Active Learning Approach." Automation in Construction, vol. 118, Oct. 2020, article 103376. doi.org/10.1016/j.autcon.2020.103376.

Liu, Pengkun, et al. "Sharing Construction Safety Inspection Experiences and Site-Specific Knowledge through XR-Augmented Visual Assistance." arXiv preprint, arXiv:2205.15833, 31 May 2022, doi.org/10.48550/arXiv.2205.15833.

Naderi, Hossein, et al. "Autonomous Construction-Site Safety Inspection Using Mobile Robots: A Multilayer VLM-LLM Pipeline." arXiv preprint, arXiv:2512.13974, 16 Dec. 2025, doi.org/10.48550/arXiv.2512.13974.

Oliveira, Bruno, et al. "Automated Monitoring of Construction Sites of Electric Power Substations Using Deep Learning." IEEE Access, vol. 9, 2021, pp. 27865-80. doi.org/10.1109/ACCESS.2021.3054468.

Pour Rahimian, Farzad, et al. "On-Demand Monitoring of Construction Projects Through a Game-Like Hybrid Application of BIM and Machine Learning." Automation in Construction, vol. 110, Feb. 2020, article 103012. doi.org/10.1016/j.autcon.2019.103012.

Sahraoui, Islem. "Automated Hazard Detection in Construction Sites Using Large Language and Vision-Language Models." arXiv preprint, arXiv:2511.15720, 13 Nov. 2025, doi.org/10.48550/arXiv.2511.15720.

Sammour, Farouq, et al. "Responsible AI in Construction Safety: Systematic Evaluation of Large Language Models and Prompt Engineering." arXiv preprint, arXiv:2411.08320, 13 Nov. 2024, doi.org/10.48550/arXiv.2411.08320.

Shrestha, Kishor, et al. "Hard-Hat Detection for Construction Safety Visualization." Journal of Construction Engineering, vol. 2015, 2015, article 721380. doi.org/10.1155/2015/721380.

Wang, Chenxin, et al. "Automating Construction Safety Inspections Using a Multi-Modal Vision-Language RAG Framework." arXiv preprint, arXiv:2510.04145, 5 Oct. 2025, doi.org/10.48550/arXiv.2510.04145.