Can AI Make Construction Safer?

The problem, in numbers

In 2024, roughly one thousand construction workers in the United States died on the job — the highest absolute toll of any industry, and close to one in five of all workplace deaths.^{Source 1} The rate has stayed near double the all-industry average for more than a decade. The problem is not a lack of rules. It is that a rule no one is watching does not prevent a fall. The pattern is global: construction ranks among the most dangerous industries in Australia, across Asia's fast-building economies and in the United States alike — the figures below are US federal data, the best-documented series, but the dynamic travels.

What improves, and why

Computer-vision systems now watch live site feeds and flag what a human supervisor cannot catch continuously: a worker without a hard hat, a body too close to a swinging load, an unprotected edge, a fall in progress. The meaningful change is not detection — it is timing. Historically a violation surfaced in an end-of-day report, if at all. Now the alert arrives while intervention is still possible.

There is a second, quieter effect. Continuous monitoring changes behaviour on its own — workers self-correct when they know the site is watched. That is genuine value, and also the first place governance questions appear.

What the evidence actually says

Here the honest picture matters more than the exciting one. Vendors report large drops in recordable incidents — one national contractor cited a 78% reduction after deployment.^{Source 5} These numbers are directional, not conclusive: they are self-reported, rarely controlled, and selected for publication.

The most rigorous reviews are more sober. A 2025 peer-reviewed survey found that most systems deliver stand-alone detection and that real safety impact depends on complete alerting and supervisory workflows — not detection accuracy alone.^{Source 4} The clearest proof that the curve can bend comes from a non-AI case: focused OSHA enforcement cut trench-collapse deaths nearly 70% since 2022.^{Source 3} The lesson is uncomfortable for technology vendors — attention and follow-through save lives, with or without a model.

Where it falls short

• Coverage, not omniscience. A camera protects only what it frames. Blind spots, dust, low light and occlusion are exactly where fatal incidents cluster.

• Alert fatigue. Detection that fires constantly and routes nowhere trains crews to ignore it. The failure is organisational, not technical.

• No causal proof yet. We have compelling response-time data and weak harm-reduction data. Claiming the latter from the former is the field's central overreach.

Where governance decides the outcome

This is the dimension the other three cannot buy. Continuous worker surveillance raises consent, data-retention and disciplinary-use questions that must be settled before deployment, not after. Who sees the footage? Is it used to coach, or to punish? What happens to a false positive that stops work? Safety improves only when detection is wired into a supervisory workflow with clear accountability — the same finding the evidence keeps returning.

How to think about adoption

Start where the risk profile justifies it — large, high-turnover sites with real fall or struck-by exposure. Treat the camera as the cheapest part of the system: budget for integration, training and the safety headcount that acts on alerts.^{Source 6} Define governance first. Measure response time from day one, and treat any incident-rate claim as a hypothesis to be tested on your own site, not a result to be assumed.