AI and Workplace Safety: What the Evidence Shows So Far

Evidence that AI tools reduce workplace injuries is thin. A systematic review published in Systematic Reviews on July 11, 2025 screened 1,255 articles from 2018 to 2024 and found only two studies that measured worker injury or illness outcomes, neither of high quality, and its authors wrote that wide-spread use of AI for health and safety may be premature. Workplace AI also creates its own duties: in Ontario, employers with 25 or more employees must have a written policy on electronic monitoring.

What the best review found

The review, led by Arif Jetha and colleagues, searched for peer-reviewed quantitative studies of occupational health and safety tools that use AI and that measured an actual outcome in workers. Of 1,255 articles identified, 359 went to full-text review and two met the criteria. The authors rated none as high quality, one as medium, and one as low. Because there were so few studies and they differed so much, no pooled estimate could be calculated.

The review has limits that cut both ways. It excluded pilot and prototype studies, and about 10 percent of the full texts screened were feasibility studies, mostly laboratory-based, including robotic systems that may well prevent injuries. It also could not see proprietary tools whose vendors have not published peer-reviewed results. The gap it describes is therefore a gap in published, controlled evidence, not proof that the tools fail.

The single trial, and what it measured

The medium-quality study was a randomized trial in Japan, published in 2021, of a 12-week health promotion program delivered through an AI-assisted chatbot. It enrolled 94 white-collar workers, 48 in the intervention group and 46 in the control group, and measured neck and shoulder stiffness or pain and low back pain. In the intervention group, 75 percent reported subjective improvement in pain, against 7 percent of controls, and the share with severe symptoms fell from 77 percent to 33 percent, against a change from 76 percent to 67 percent in the control group.

That result is encouraging and narrow. It concerns symptoms in office workers, not injuries on a construction site, a warehouse floor, or a plant. Readers who see AI safety claims for those settings are looking at a different kind of evidence.

A laboratory example of promise and error

NIOSH described in a December 9, 2020 bulletin a computer vision method that estimates how far a worker bends at the trunk while lifting, using only the size of a box drawn around the person in video. It was trained on 105 computer-generated postures and validated on 180 lifting tasks by five people in a laboratory, with a research-grade motion capture system as the reference. The mean absolute difference from the reference was 15.85 degrees, with an R-squared of 0.80. The bulletin presents the work as a feasibility step toward measuring posture without a human observer.

An average error of nearly 16 degrees may be acceptable for sorting risky lifts from safe ones and unacceptable for a precise ergonomic score. The point of the example is that a headline such as "AI spots unsafe lifting" depends on error size, training data, and the setting in which it was tested.

Hazards the technology adds

NIOSH has been direct about this. In a September 9, 2024 bulletin, NIOSH Director John Howard and Paul Schulte wrote that AI adoption can result in a broad range of hazards and risks to workers. They named the black-box nature of many systems, which makes decisions hard to understand and erodes trust, discriminatory outcomes, and design flaws that may stay hidden until deployment. A May 24, 2021 NIOSH bulletin added that risk rises when data are incomplete, inappropriate, or insecure, and when systems run without human oversight, and it noted that changed schedules, hours, or lone work could create new exposures and mental health effects.

The 2024 bulletin summarizes five options for managing these risks: reskilling, joint evaluations by developers and safety practitioners, independent audits, certification, and a safety case that documents why a system is safe to operate.

Monitoring rules when the tool watches workers

Many safety products work by watching people: cameras that check protective equipment, wearables that track location or movement, and vehicle telematics. In Ontario, the Employment Standards Act, 2000 requires employers with 25 or more employees on January 1 of a year to have a written electronic monitoring policy in place before March 1 of that year. The policy must say whether the employer monitors electronically and, if so, how and in what circumstances, for what purposes the information may be used, and the dates of preparation and of any change. Employees get a copy within 30 calendar days.

The province's guide states that these requirements do not create new privacy rights for employees. The rule is about disclosure. Whether a given camera or wearable program is lawful under privacy, human rights, or collective agreement terms is a separate question. This section is general information, not legal advice, and employers and workers can confirm specifics with counsel or the provincial ministry.

Questions to put to any safety AI claim

• Which outcome was measured: injuries, near misses, symptoms, or only detections in video?
• Was there a control group or a before-and-after comparison, and for how long?
• Where was it tested, and does that site resemble yours?
• What is the error rate, including false alarms and missed hazards?
• Who sees the data, how long is it kept, and is there a written monitoring policy?
• Does a person review the output before action is taken against a worker?

FAQ

Does AI reduce workplace injuries?

The published, controlled evidence is too limited to say. The most recent systematic review found no high-quality study measuring injury or illness outcomes and cautioned against wide-spread recommendation.

Can an employer in Ontario use cameras or tracking powered by AI?

The monitoring policy rule requires disclosure for employers with 25 or more employees. Other laws and agreements may limit what monitoring is allowed, so the policy rule alone does not settle legality.

Does the Ontario rule apply to small employers?

It applies only to employers that had 25 or more employees in Ontario on January 1 of the year in question. Smaller employers have no policy requirement under this rule.

Who is responsible when an AI safety system misses a hazard?

This article does not decide that. Responsibility depends on the jurisdiction, the facts, and contracts, and NIOSH lists independent audits and documented safety cases among its risk management options.

Related reading

• Where Canadian AI law stands
• Arguments against the AI threat view
• Moral questions raised by AI
• AI in organizational security

More in AI Policy and Ethics.

Last verified: October 2026

Sources

• Do occupational health and safety tools that utilize artificial intelligence have a measurable impact on worker injury or illness? Findings from a systematic review, Jetha A et al., Systematic Reviews 14:146, July 11, 2025
• Exploring Approaches to Keep an AI-Enabled Workplace Safe for Workers, Howard J, Schulte PA, NIOSH Science Blog, September 9, 2024
• The Role of Artificial Intelligence in the Future of Work, NIOSH Science Blog, May 24, 2021
• Computer Vision Development for Estimating Trunk Angles, NIOSH Science Blog, December 9, 2020
• Your guide to the Employment Standards Act: written policy on electronic monitoring of employees