Chile brought AI to the workplace safety table. Workers need the sensor veto

THE SIGNAL IN ONE SENTENCE

A camera can notice that a forklift crossed a painted line. A wearable can warn that heat exposure is climbing. A model can search incident reports for patterns a hurried manager missed. None of those systems gets to decide what a worker is worth. That is the useful tension inside a small international meeting held November 4 in Chinese Taipei. Government officials, specialists and worker organizations from Asia-Pacific economies met under the APEC umbrella to discuss artificial intelligence for occupational safety and health. Chile sent Macarena Pinto, head of advisers at the Subsecretariat of Social Welfare, to a panel with participants from Japan, South Korea, Thailand and Malaysia, among others. Chile's official account, dated November 5 and posted with a November 6 publication line, says the workshop examined how AI could identify hazardous situations earlier, monitor workplace conditions and analyze information that helps prevent accidents and occupational illness. That is a promising list. It is also a list of verbs that can slide from protection into control. Identify can become classify the worker. Monitor can become watch all day. Analyze can become score, schedule, discipline or dismiss. A system installed to catch a dangerous temperature can quietly become a system that measures bathroom time. A camera trained to spot missing protective equipment can become evidence in a performance review. A model that predicts injury risk can tempt an employer to exclude a worker instead of fixing the job. The workshop report does not say that any of this happened. It does not name a model, company, workplace, sensor package, accuracy result, false-alarm rate, privacy design or collective agreement. It describes a policy conversation, not a deployed system. That distinction matters because the next step should not be a procurement race. It should be a worker-governance test. Before a safety system turns on, the people being observed should be able to answer plain questions. What hazard is this tool meant to reduce? What data does it collect? Does it identify individuals? Who can see the raw feed, alert and historical record? How long is each retained? Can the information be used for discipline, productivity scoring, insurance or hiring? How does a worker challenge a false alert? What happens when the system misses danger? Who can pause it? Who can stop the work? Call the answer a sensor veto. The phrase is editorial, not a term from the APEC workshop or Chilean law. It describes a bundle of practical powers: workers can help choose the safety purpose, reject unnecessary collection, block unrelated secondary uses, inspect the evidence behind an alert, contest a mistake, pause unsafe automation and preserve their independent right to stop dangerous work. Chile already has a legal and institutional foundation for much of that idea. Decree 44, the country's preventive-risk regulation, places responsibility for workplace risk management on the employer. It calls for hazard identification, risk assessment, prevention programs, worker participation, consultation and dialogue. It requires participation by workers or their representatives in planning, implementing, evaluating and controlling workplace risks. It also says sensitive or confidential information in safety records must be protected. Article 184 bis of Chile's Labor Code adds a human stop control. The Labor Directorate says a worker may interrupt work and, if necessary, leave the workplace when the worker reasonably believes continued work creates a grave and imminent risk to life or health. A worker who lawfully exercises that right cannot be harmed for doing so. An AI alert should support that protection. It should never shrink it. If a sensor sees a danger the worker cannot see, the system should make the hazard legible and help people act sooner. If a worker sees a danger the model missed, the dashboard should not outrank the person standing beside the machine. No green score should cancel a reasonable stop-work decision. This is more than a Chilean question. The International Labour Organization says automation and smart monitoring can reduce hazardous exposure, prevent injuries and improve working conditions. It also warns that AI at work can bring intrusive surveillance, work intensification, reduced autonomy, privacy concerns and decisions that are hard to understand or contest. The same instrument can therefore produce two different workplaces. In one, the system looks for dangerous heat, toxic exposure, unstable equipment or risky traffic patterns. Workers and safety committees define the purpose. Alerts point to the hazard. The employer fixes the source. Data is minimized, retained briefly and kept away from discipline. False alarms are reviewed. Misses are recorded. Workers can challenge the machine and stop work. In the other, the system watches bodies continuously, scores behavior, hides its thresholds and treats every deviation as an individual failure. The employer learns more about the worker than about the broken process. Safety becomes the friendly label on a surveillance product. The difference is not whether the software contains AI. It is who has power over the sensor, data and consequence. This is where worker organizations belong in the design room, not in the apology meeting after launch. They know which shortcuts are forced by production targets, which alarms already fail, which protective equipment does not fit, which shifts run short-staffed and which official process exists mostly on paper. That operational knowledge is training data of the most valuable kind, but it should enter through participation, not extraction. A serious pilot would begin with one hazard and one measurable outcome. Reduce near misses at a loading bay. Detect dangerous heat early enough to trigger rest and cooling. Identify a machine condition before maintenance becomes an injury. Measure whether the hazard actually falls. Publish false positives and false negatives. Audit performance across shifts, job types, protective clothing, skin tones, body types, disabilities and environmental conditions. Then publish the boundary. Safety data is not productivity data. An alert is not proof of misconduct. A prediction is not a diagnosis. No individual sanction should rest solely on a system designed for prevention. Workers need notice, access to the relevant record, a human review path and the ability to bring a representative. The machine can help see the hazard. The people who do the work still need the final word on what the system sees, what it is allowed to become and when the job must stop.