What is artificial intelligence applied to occupational safety and health?
In prevention, an AI system can classify images, detect anomalies, estimate probabilities, extract information from texts, or generate recommendations. It can be integrated with sensors, cameras, maintenance logs, assessments, or incident reports. The result is an inference: it depends on the data, the model, the chosen threshold, and the context of use.
The preventative purpose must be defined before selecting the technology. Alerting someone about entering a restricted area is not the same as scoring a person’s behavior or deciding their job assignment. The greater the impact on rights, working conditions, or safety, the greater the evidence, oversight, and safeguards must be.
Possible applications
Possible uses include:
- detect presence in dangerous areas or absence of protection;
- identify anomalous patterns in equipment to guide maintenance;
- prioritize documents or communications for technical review;
- support ergonomic or exposure analyses when the method is validated;
- identify trends in incidents and pending measures;
- adapt training materials or facilitate access to information;
- assist in planning without replacing responsible decision-making.
The usefulness of AI depends on whether the problem lends itself to this approach and whether the output leads to effective action. An alert that arrives late, produces too many false positives, or lacks accountability can create noise and false confidence. AI also fails to compensate for faulty sensors, biased data, or incomplete risk assessments.
Differences with automation and conventional analytics
Automation executes a predefined rule or sequence, such as sending a notification when a review is due. Conventional analytics summarizes or compares data using known rules and models. AI encompasses systems that infer how to generate predictions, content, recommendations, or decisions with varying levels of autonomy. In practice, these can be combined.
Not every advanced function requires AI. If a simple rule solves the problem reliably, understandably, and less intrusively, it is usually preferable. Nor does every system commercially labeled as AI perform the same function or present the same risk. The evaluation should consider the intended use, the data, the people affected, and the consequences of the error, not the product name.
Preventive benefits and risks
The ILO notes that automation and intelligent monitoring can reduce hazardous exposures and support improved working conditions, while EU-OSHA warns of potential effects on autonomy, workload, and psychosocial risks. The clearest benefits emerge when technology eliminates exposure, improves a technical barrier, or helps detect a situation that would otherwise go unnoticed.
Risks include false negatives, excessive alarms, bias, loss of context, dependency, secondary use of data, disproportionate surveillance, work intensification, and shifting of responsibility onto the system or the person being observed. A high average accuracy is not enough: critical errors must be analyzed by task and by group, including the possibility that the system performs worse under certain conditions or with certain individuals.
How to implement it successfully
A responsible project can follow these phases:
- Define the hazard and the expected preventive outcome.
- Check if there are more effective or less intrusive non-technological measures.
- Consult with workers, representatives, prevention, data protection and cybersecurity.
- Determine legal requirements, purpose, necessary data and limits of use.
- Validate with representative data and risk-related metrics, including false negatives.
- Conduct a controlled test without removing existing barriers.
- Establish human oversight, alert response, and appeals mechanism.
- Document versions, changes, incidents, and decisions.
- Monitor technical, organizational and psychosocial effects and withdraw the system if it does not provide guaranteed value.
The company maintains its duty of prevention even if it uses an external provider.
Participation, transparency and human oversight
Those affected need to know what system is being used, for what purpose, what data it processes, what its output means, and who makes the decisions. Human oversight is not simply about automatically accepting recommendations: those who review them need competence, time, authority, and access to sufficient information to question them. They must also be aware of automation bias, that is, the tendency to place excessive trust in computer output.
Prior consultation is especially important if AI modifies the organization, pace, allocation, or evaluation of work. Article 33 of Law 31/1995 includes consultation on the introduction of new technologies when they may affect safety and health. Participation helps to identify actual uses, unforeseen effects, and less intrusive alternatives.
Practical example
A plant is studying a system that analyzes vibrations to detect equipment degradation. Before implementing it, they confirm that the machine has adequate safeguards, is undergoing scheduled maintenance, and has a safe shutdown procedure. The alerts are validated with maintenance personnel, and it is established that no system output authorizes work on energised equipment.
During the pilot program, alerts are compared with independent inspections, false negatives are recorded, and the responsible party is identified. A model update requires further validation. If the system alerts, it triggers a review; if it doesn’t, it doesn’t eliminate scheduled checks. This design uses AI as an additional signal and maintains technical barriers, procedures, and human decision-making.
Regulatory framework in the European Union and Spain
Law 31/1995 remains the foundation: it mandates risk assessments, adapting prevention measures to changes, and consultation on the introduction of technologies with effects on safety and health. Regulation (EU) 2024/1689 on AI is applicable with a phased implementation schedule and prohibits certain uses. Its consolidated version of July 2026 postpones the application of the main obligations for systems classified as high-risk under Article 6.2 and Annex III, which includes certain applications related to employment and worker management, until December 2, 2027. These obligations include informing representatives and affected individuals before using a high-risk system in the workplace.
If personal data is processed, the GDPR applies: lawfulness, transparency, minimization, security, impact assessment where appropriate, and safeguards against decisions made solely by automated means with legal or similar effects. Specific classification requires case-by-case analysis; a preventative tool is not automatically high-risk simply because it is called AI.
