Imagine that a manufacturer installs an AI-enabled camera system to identify risky lifting techniques. Everyone has a different perspective of the camera system. Executives see a tool that can prevent musculoskeletal disorders (MSD). The safety team sees faster hazard detection. Workers see something else: a camera recording their every move, an algorithm judging their performance and a potential new source of discipline.

The system may work exactly as designed by the manufacturer and still fail as a safety intervention.

This happens when leaders treat worker trust as a communications issue rather than an operational control. If employees fear how managers will use AI-generated information, they may avoid the system, work around it, withhold context or stop reporting mistakes. Technical accuracy cannot compensate for missing information and defensive behavior.

A recent survey found that 75% of surveyed safety leaders trust AI-generated insights. That finding shows growing confidence among decision-makers. But do the workers affected by those insights share that confidence?

Trust Must be Calibrated

Trusting AI does not mean accepting every recommendation. Blind confidence creates its own safety hazard. Workers and supervisors need enough confidence to use the technology but with enough skepticism to question an output that conflicts with conditions on the ground.

By calibrated trust, I mean confidence matched to the system's demonstrated reliability and to the consequences of being wrong. Workers should know when they can rely on an output, when they should verify it and when they must escalate it to a human decision-maker.

That balance belongs at the center of AI safety. The National Institute of Standards and Technology frames trustworthy AI around characteristics such as reliability, transparency, accountability, privacy protection and the management of harmful bias. These qualities require more than a technically capable model. They require an organizational process that lets people identify, communicate and correct problems.

The American Society of Safety Professionals takes a similar position on workplace AI. The organization recognizes AI’s potential to improve occupational safety and health while also warning that the technology can introduce hazards and privacy concerns.

For EHS leaders, the practical implication is clear. Human oversight only works when humans feel permitted to exercise it. Employees must be able to challenge an AI recommendation, report a false alarm, explain missing context and request a second review without being branded as resistant to technology.

A system technically keeps a person in the loop when it requires a supervisor to approve an automated recommendation. Meaningful human oversight goes further. The supervisor must understand the recommendation, possess enough authority to reject it and feel safe doing so.

Workers Evaluate the Motive Before the Tool

Leaders frequently explain what an AI system does while neglecting to explain why they are introducing it and how they will use its data.

Workers fill that vacuum with their own predictions. They may assume that a fatigue-detection system will generate disciplinary records. They may fear that computer vision will measure productivity under the label of safety. They may wonder whether a near-miss reporting assistant will identify employees who admit mistakes.

A 2026 International Labour Organization analysis found that intrusive AI monitoring can create psychosocial risks through surveillance, reduced autonomy, work intensification and uncertainty about data use. The American Psychological Association has likewise warned that poorly designed electronic monitoring can communicate distrust and increase employee anxiety.

Safety leaders cannot overcome these concerns with slogans about innovation. They need enforceable boundaries. Those boundaries also help safety leaders earn employee buy-in, a core part of change management.

In my work advising organizations on risk management, workplace decision-making, change management and AI adoption, I see employees judge a new tool by the rules and incentives around it as much as by the tool itself. Buy-in should not mean persuading workers to accept a decision leadership has already made. Rather, it means giving workers enough information and influence early enough that their concerns can change the design, data rules, training or deployment plan.

Before implementation, leaders should state what information the system collects, why it collects that information, who can access it, how long the organization retains it and which decisions the data may influence. They should also identify prohibited uses. For example, a company might use posture data to redesign workstations and target ergonomics training while prohibiting its use for individual productivity rankings. A clear boundary gives employees something concrete to evaluate. A vague assurance that “the data will only support safety” does not.

Give Workers Influence Before Deployment

Organizations often introduce safety systems in a predictable, top-down order. Executives select a product. Information technology configures it. EHS develops training. Supervisors announce it. Workers encounter the system after the important decisions have already been made. That sequence sacrifices useful knowledge and creates avoidable resistance.

A recent survey from the MSD Solutions Lab, part of the National Safety Council (NSC), highlights the importance of asking frontline workers about the technologies they use. Employees understand where the work varies from written procedures, where equipment produces false signals and where a theoretically useful intervention creates a practical obstacle. Their involvement should begin during problem definition. Ask workers which hazard the organization should address, what information would help them, what kinds of monitoring would feel intrusive, and what could cause employees to ignore or circumvent the system.

The NSC's 2026 survey of 405 nonmanagerial workers puts numbers behind that point. Eighty-five percent said they had received adequate instruction, 83% had opportunities to practice using the technology and provide feedback, and 83% felt comfortable suggesting improvements. Yet only just over half of respondents said they played a role in decisions about their organization's use of the technology. Workers in more participatory environments also reported stronger perceived safety benefits, lower technology-related mental stress and higher job satisfaction. Training workers after a decision is made does not substitute for giving them influence over the decision itself.

The NSC has observed that safety technology introduced without worker involvement, clear communication or adequate training often fails to deliver its expected results. Participation therefore serves two purposes. First, it improves the system. Second, it gives employees evidence that leadership values their judgment.

Separate Safety Learning from Automatic Punishment

Every safety professional understands the value of reporting hazards, close calls and weak signals before someone gets hurt. AI systems make that reporting culture even more important because the technology can produce incorrect, incomplete or misleading outputs.

Employees will expose those weaknesses only when the organization responds constructively. The reason is informational. Traditional safety programs depend on weak signals: near-misses, recurring workarounds, equipment quirks, and gaps between written procedures and real work. AI adds another layer of weak signals, including false positives, missing context, fabricated guidance, skewed predictions or alerts that workers learn to ignore.

Frontline employees are often the first to notice these mismatches. If reporting them creates personal risk, the organization loses the information it needs to correct the system before a small defect contributes to an injury. OSHA’s guidance on worker participation calls for protecting employees from retaliation when they report injuries, illnesses and hazards or participate in safety programs. Organizations should apply that same principle when workers report problems involving AI.

Suppose an employee discovers that a computer vision system repeatedly flags a safe task as hazardous. The first response should focus on understanding the error, documenting its frequency and correcting the system. The organization should avoid blaming the employee for failing to follow an automated alert that does not fit the actual conditions.

The same approach should apply when workers admit that they ignored, misunderstood or worked around the tool. Those actions may reveal poor training, workflow conflicts, excessive alerts or a lack of confidence in the data. Treating every workaround as misconduct can erase the information leaders need to improve the system.

This does not require abandoning accountability. Deliberate disregard for a well-understood critical control still warrants intervention. The goal of nonpunitive reporting is to distinguish useful disclosure from reckless conduct and to prevent fear from concealing emerging hazards.

Preserve Human Authority at Critical Moments

AI can recognize patterns across large amounts of safety data, identify unusual conditions and surface information quickly. It can also miss context that experienced employees recognize immediately.

An alert may overlook maintenance work occurring outside the normal schedule. A predictive model may rely on historical records that underrepresent near-misses. A generative AI assistant may produce an authoritative answer that omits a site-specific requirement.

Organizations need a written rule describing when an employee must stop, verify or escalate an AI-supported decision. Safety-critical recommendations should have a clearly identified human owner. Workers should know whom to contact and how to proceed when the system conflicts with training, procedures or direct observation.

The National Institute for Occupational Safety and Health’s 2026 guidance on managing AI hazards in the workplace emphasizes practical risk management, so that new technological risks do not outweigh AI’s benefits. That principle should shape deployment decisions. The more severe the possible consequence, the stronger the verification and escalation process should become. No employee should hear, “The system says it is safe,” as the final answer to a credible concern.

Measure Trust as a Leading Indicator

Organizations commonly evaluate AI programs by tracking accuracy, alerts, adoption rates, time saved and incidents prevented. Those measures provide useful information, but they cannot reveal whether employees feel safe challenging the system.

EHS teams should add questions that measure employee experience, such as: 

• Do workers understand why the technology exists?
• Do they know what data it collects?
• Do they believe managers will use that data consistently with the stated purpose?
• Can they report an error without negative consequences?
• Do they know who retains final decision authority?

Leaders should also track behavioral indicators. A sudden drop in reported system errors may signal improved performance. It may also mean employees have stopped reporting them. Low override rates can indicate accurate recommendations, or they can reveal that supervisors fear contradicting the system.

OSHA recommends using leading indicators, such as worker participation, safety suggestions, reported hazards and employee opinions about program effectiveness. EHS leaders can adapt those measures to AI by monitoring worker feedback, disputed alerts, corrected outputs, escalation frequency and management response time.

Treat Trust as Part of the Safety Architecture

Trust grows through repeated evidence. Employees watch whether leaders explain decisions, respect agreed boundaries, correct system failures and protect people who raise concerns. One town hall cannot establish that record.

The emerging model of responsible AI calls for managing trustworthiness throughout the technology’s lifecycle. For safety leaders, that means worker involvement cannot end after the pilot. Employees need an ongoing role in evaluation, correction and governance.

The broader psychology of AI adoption leads to the same conclusion. People judge a new system through the behavior of leaders, the reactions of supervisors and the experiences of trusted colleagues. Having a formal written policy is important, but daily evidence determines whether employees believe the policy.

AI may help an organization detect hazards sooner, analyze incidents faster and make better-informed decisions. Those benefits depend on people who supply accurate information, challenge questionable outputs and speak up when the technology misses something important.

Trust, therefore, belongs beside data quality, training, maintenance, verification and emergency response as a genuine risk control. When workers trust the process enough to participate honestly, AI can strengthen the safety system. When fear drives problems underground, even a powerful system operates with a dangerous blind spot.