At EBIS-HSE, we’re committed to advancing global understanding in health and safety through the latest insights and evidence-based research. This series highlights groundbreaking studies and innovative thinking from around the world, helping professionals stay informed and inspired in shaping safer workplaces.
Research shows that AI can perform many safety-related tasks with impressive technical accuracy, but strong model performance does not automatically translate into fewer injuries or healthier workplaces. The most important message is that AI should support established safety systems, not replace competent supervision, worker involvement, engineering controls, or professional judgement.
Summary
Artificial intelligence is increasingly being tested as a tool for occupational health and safety. Applications range from predicting accident rates and analysing incident reports to monitoring PPE compliance, assessing posture, detecting falls, and identifying changes in heat, noise, air quality, and other environmental conditions.
This systematic review examined 43 experimental studies published between 2013 and 2024. Construction and industrial settings featured heavily in the evidence, reflecting the suitability of computer vision, sensors, and predictive analytics for workplaces involving machinery, physical hazards, and observable work activities.
Many studies reported strong technical performance, with several achieving F1 scores above 0.80. Computer vision systems were able to identify helmets and other PPE, while machine-learning models detected patterns in accident records, classified safety reports, and generated early warnings from sensor data.
However, the review makes an important distinction between technical accuracy and real-world safety effectiveness. A model that performs well on a controlled or historical dataset may behave differently when lighting, camera position, equipment, work methods, or workforce characteristics change. Poor-quality data, incomplete incident records, false alarms, and missed detections can also affect reliability.
The evidence therefore supports AI mainly as a decision-support tool. It can help safety professionals process large volumes of data and identify patterns more quickly, but it should be integrated into existing risk-management systems rather than treated as a substitute for prevention, worker consultation, or professional judgement.
Methodology
The review followed PRISMA 2020 and was registered with PROSPERO. Searches were conducted in PubMed, Scopus, and Google Scholar up to 19 July 2024.
The search identified 540 records, with 43 experimental studies remaining after duplicate removal, screening, and full-text assessment. Two reviewers independently screened the studies, with disagreements resolved through discussion and, where necessary, a third reviewer. Inter-rater agreement was high, with a Cohen’s kappa score of 0.87.
The authors also developed a 34-point appraisal tool, TROSH-IA, to assess workplace context, datasets, algorithms, hardware, performance, external validity, limitations, and practical implications. Because this tool was created specifically for the review and has not been widely validated, its quality assessments should be interpreted cautiously.
Key Findings
1. AI Can Support Accident Prediction
Several studies used historical accident data to predict incident frequency, injury rates, or levels of risk. Techniques included random forests, artificial neural networks, support vector machines, and long short-term memory models.
These systems may help safety teams identify patterns across large datasets and prioritise higher-risk activities, locations, or conditions. However, the quality of the prediction depends heavily on the quality of the underlying data. Underreporting, inconsistent classifications, and changes in reporting practice can all reduce reliability.
AI can identify associations, but it does not necessarily explain why incidents occur. Investigation, worker consultation, and professional judgement remain essential before decisions are made about controls.
2. Computer Vision Can Monitor Visible Safety Conditions
Computer vision was widely used to identify helmets, high-visibility clothing, posture, movement, proximity to hazards, and falls. In large or complex work areas, these tools may help identify conditions that would be difficult to monitor continuously through human observation alone.
Their performance, however, can deteriorate because of poor lighting, weather, camera angle, distance, occlusion, or changes in PPE design. Continuous monitoring also raises questions around privacy, trust, and interpretation.
A system may detect that a worker is not wearing PPE, for example, but it cannot automatically explain why the equipment is absent or whether a higher-level control should have removed the hazard altogether. This is why AI monitoring should sit within the hierarchy of controls rather than replacing it.
3. AI Can Provide Earlier Warnings
Some studies combined sensors and machine learning to monitor falls, heat, noise, air quality, and physical activity in real time. These systems may help safety teams identify changing conditions earlier and respond before exposure becomes more serious.
However, an alert only improves safety if someone receives it, understands what it means, and takes appropriate action. Too many false alarms can create alarm fatigue, while missed detections may create misplaced confidence in the system.
Employers therefore need clear procedures for who responds to automated warnings, what action is expected, and how the system’s performance is monitored over time.
4. Incident Reports Can Be Analysed at Scale
Natural language processing and classification tools were used to analyse large collections of accident narratives, near-miss reports, and safety observations. These tools can identify recurring themes, equipment, activities, or locations much faster than manual review alone.
Their usefulness still depends on the quality of the original reports. Short descriptions, missing context, and inconsistent terminology can all limit what the system is able to identify.
Automated analysis should therefore help prioritise and organise information rather than replace access to the original reports or detailed human review.
5. Accuracy Does Not Prove Injury Prevention
One of the review’s most important findings is that high model accuracy should not be treated as evidence that AI reduces workplace harm.
Many studies reported strong classification or prediction results, yet relatively few examined whether implementation led to measurable reductions in injuries, occupational illness, or hazardous exposure. External validation was also inconsistent, and much of the available evidence came from construction and industrial environments.
The evidence is therefore stronger for AI’s ability to perform specific technical tasks than for its ability to improve long-term safety outcomes.
6. Governance Is as Important as Technical Performance
The authors identify privacy, data quality, algorithmic bias, cost, technical capability, and lack of standardisation as major barriers to implementation.
Organisations should begin with a clearly defined safety problem rather than adopting AI simply because the technology is available. Systems should be tested in the workplace where they will actually be used, with agreed standards for acceptable error rates and regular review.
Workers should also understand what information is being collected, how it is used, and who can access it. Responsibility for validation, updates, incident response, and ongoing system performance must remain clearly assigned.
Takeaways for Practice
- Start with a clearly defined workplace hazard or safety decision rather than the technology itself.
- Test AI systems under real operating conditions before relying on their outputs.
- Keep competent human oversight and allow workers to question or challenge system errors.
- Monitor false alarms, missed detections, and changes in system performance over time.
- Use AI alongside the hierarchy of controls, not as a substitute for hazard elimination or engineering controls.
- Establish clear rules for privacy, data use, accountability, validation, and review.
- Retain access to original incident reports and supporting evidence when using automated analysis.
- Measure injury, illness, and exposure outcomes rather than judging success only through model accuracy.
- Involve workers and safety professionals in the design and implementation of AI systems.
Read the full research study here:
La Torre, G., Manai, M. V., Meucci, S., Lucente, A., Picerno, A., Ammirati, S., & De Sio, S. (2026). Artificial intelligence and occupational health and safety: a systematic review. Journal of Public Health. Published 25 March 2026. https://doi.org/10.1007/s10389-026-02738-8