AI for Construction Safety: What's Working in the Field

AI construction safety tools mainly use cameras and computer vision to flag missing PPE and unsafe conditions, and research systems also detect fall hazards and proximity risks. Published detector accuracy ranges widely by dataset, and studies show degradation in poor lighting and crowded scenes, so as of 2026 AI supplements site safety supervision and does not replace it.

Our take: presence detection is not proof of correct use, and no evidence we found ties these tools to fewer injuries. That does not make them useless. It means buyers should treat them as an additional layer of observation, and judge them on incident data from comparable sites. For the wider picture, see AI in construction in 2026.

Why does construction safety matter for the AI conversation?

The Bureau of Labor Statistics reported 5,070 fatal work injuries across US occupations in 2024, down 4.0% from 5,283. Construction and extraction occupations accounted for 1,032 of them. That is an occupational-group count, not an industry total, so it should not be quoted as 'construction industry deaths'.

Even so, it is the scale that makes safety the use case with the highest moral stakes. It is also the one practitioners rank low. In RICS's Q1 2026 monitors (3,148 respondents, 1,883 in construction), safety and sustainability ranked lowest for AI significance. The likeliest reading is that safety is regulated, hands-on and already culturally owned by people, so AI feels like an add-on. That is our interpretation, not a survey finding.

How is AI used for construction safety today?

The brief for this topic also includes wearables and near-miss analysis. We gathered no source on wearable accuracy or on near-miss identification tools, so we make no claims about them. Where this article is specific, it is about camera-based detection, the best-studied category.

How accurate are AI PPE detectors?

Researchers report mean average precision (mAP) of roughly 0.70 to 0.95 on curated datasets. Examples include 0.841 mAP@0.5 under variable lighting and 86.55% accuracy for YOLOv5x on a helmet dataset. The studies use different image sets, camera positions and labeling rules, which is why the numbers cannot be ranked against each other.

What the figures omit matters more than what they include:

• Lighting. Variable light, shadows and glare lower scores.
• Occlusion. Workers partly hidden by materials or equipment are missed.
• Clustering. Groups of workers confuse detectors.
• Scope. Most detect helmets and vests only; correct wear is not verified.

Illustrative example: a camera alert says "worker without hard hat, 3:12 pm, level 4." The supervisor walks up and finds the worker was holding the helmet while stepping into a break trailer. That false alarm costs a minute. A missed alert for a worker whose harness is clipped to nothing costs far more, and no common detector in these studies checks that.

Does AI actually reduce injuries?

We found no source that demonstrates it. Accuracy on a dataset is a laboratory measure; injury frequency is an outcome measure, and a long chain separates them: the alert must be right, reach a person quickly, change what happens, and the change must prevent a harm that would otherwise have occurred.

Vendor materials sometimes cite returns such as "300% ROI." Such claims are vendor-reported, and ROI of that kind depends on how avoided incidents are valued and counted. Ask any vendor for before-and-after incident rates from projects of similar type and size, over a meaningful period, and ask whether other safety changes happened at the same time.

The honest framing for a safety director: these tools may improve observation coverage, particularly for repeated checks on large sites, but the effect on outcomes is unproven in the evidence available to us. The same standard applies to the claim for any other safety intervention; it is not special pleading against AI.

How does AI fit into a site safety program?

Safety programs rest on layers: planning before the work (job hazard analysis, pre-task meetings), controls during the work (guardrails, training, supervision), and learning after (incident and near-miss review). Camera-based AI touches the middle layer, as extra eyes, and potentially the last layer, as a source of data. It does not touch the first.

That placement suggests where evaluation effort should go. The question is not "is the model accurate?" in the abstract. It is "does this alert reach the right person, quickly enough, for a hazard they would not otherwise have seen?" A detector that flags a missing helmet in a zone the superintendent walks every hour adds little. The same detector watching a laydown yard that is rarely visited, on a night shift, might add more. Value depends on the gap in existing coverage.

An illustrative pilot design

Illustrative example, with assumptions stated: a GC pilots camera-based PPE detection on two comparable projects for 90 days, one with the system and one without. Before the pilot it records observed PPE violations from supervisor walks, near-miss reports and first-aid cases for four weeks on both sites. During the pilot it tracks alerts raised, alerts confirmed by a supervisor, time from alert to response, and the same incident measures.

At 90 days the GC can compute a confirmation rate (confirmed alerts divided by alerts raised) and compare near-miss and violation trends between sites. Two projects over 90 days will not produce statistically strong injury results, since serious injuries are rare events, so the pilot is best read as a test of detection usefulness and workflow. It should not be read as proof of injury prevention. That limit is precisely why outcome evidence is missing from the field.

What are the risks of AI safety monitoring?

False confidence. A dashboard showing 98% PPE compliance can hide that the detector only looks for helmets. Teams may relax other observation habits.

Alert fatigue. If false positives are frequent, supervisors learn to ignore the feed.

Worker trust. Camera monitoring raises privacy and surveillance questions. We gathered no source on worker attitudes or legal requirements, so we do not characterize them, but any rollout should include clear communication about what is recorded, who sees it and how it is used. Data governance matters more as concerns rise: RICS found privacy and security worries among construction respondents grew from 22% to 30%.

Overreliance on text tools for safety content. Generative AI drafting job hazard analyses or toolbox talks can invent requirements or cite the wrong standards. That is the failure mode discussed in AI hallucination risk in construction, and every safety document an AI drafts should be checked by a qualified person.

What should a contractor ask before buying?

1. What exactly does the system detect, and what is out of scope?
2. What accuracy was measured, on which dataset, and how does it compare to my site conditions?
3. What are typical false positive and false negative rates in the field?
4. Who receives alerts, how fast, and what is the response procedure?
5. What happens to imagery: storage, access, retention and deletion?
6. Is there before-and-after incident data from comparable projects?
7. How will workers be told about the system?

Run any pilot with a baseline. The framework in measuring AI ROI in construction applies: record current observation practices and incident or near-miss reporting for a few weeks, then re-measure after 30, 60 and 90 days.

What this means for GCs, owners and designers

GCs and subcontractors. Use AI detection to extend, not replace, supervision. Keep human walkthroughs, pre-task planning and stop-work authority as primary controls.

Owners and developers. Ask your contractors about their monitoring practices and data handling. Don't treat a vendor's compliance dashboard as an indicator of safety performance without incident data behind it.

Designers. Design-side safety, such as prevention through design, is not covered by these tools. We found no evidence of AI-driven design safety review in the sources reviewed.

What to watch through 2027

• Outcome studies. Peer-reviewed or third-party evaluations that measure injury or near-miss rates, not detection accuracy.
• Correct-use detection. Systems that verify chin straps, harness connection and fall-protection setup.
• Low-light and crowded-scene results. Whether reported accuracy holds up outside curated datasets.
• Policy. Employer and owner requirements on camera monitoring and worker data.
• Adoption. Whether safety rises above its lowest ranking in future RICS surveys.