AI at work: what leaders need to know about safety
Artificial intelligence is no longer a side project. It's embedded in controls, scheduling, monitoring, and decision-support across operations. As adoption grows, so do questions about worker safety and health.
New guidance from NIOSH offers a practical way to manage AI-related risks using the same occupational safety science you already rely on. It treats AI as a software factor that changes how existing hazards show up-not as an entirely new hazard category.
What the guidance changes
The big shift: stop treating "AI" as a single, vague thing. Focus on how a trained algorithm affects a specific system and workflow. That clarity makes it easier to plug AI into your current hazard identification, exposure assessment, and control processes.
Algorithms don't create new physical, chemical, or biological hazards out of thin air. But they can adjust equipment behavior, process timing, and human interactions in ways that increase or decrease exposure. That's where the risk lives.
NIOSH also calls out psychosocial risks. Changes to autonomy, monitoring, and job expectations can raise cognitive load, stress, and uncertainty. Treat these like any other occupational health risk-assess, control, and follow up.
Use precise language: "trained algorithm," not "AI"
"AI" is a catch-all. A trained algorithm is concrete: software using data to influence outputs, prioritize tasks, tune parameters, or trigger actions. Naming it this way keeps your team focused on how the software actually functions in your workplace.
The algorithmic hygiene framework
NIOSH introduces an "algorithmic hygiene" lens that adapts industrial hygiene principles to software-driven systems. It links software characteristics to known hazard categories and control strategies.
System characteristics that influence risk
- Data and methodology design: data quality, bias, drift, model updates, and validation routines.
- Worker-algorithm trust: clarity of decision logic, explainability, and predictable behavior under edge cases.
- Worker-management trust: transparency on monitoring, metrics, and how data is used for evaluation.
- Job reskilling demands: new competencies, training access, and the pace of change.
- Cybersecurity and hardware integration: secure interfaces, fail-safes, and physical interlocks.
These factors influence both tangible and psychosocial exposures, changing how risks present and how controls should be applied.
NIOSH positions this framework as a starting point. The goal is a scientific basis for actionable guidance that employers, developers, and policymakers can put to work.
Prevention and control strategies
Work design controls (employer responsibility)
- Redefine roles to clarify what the algorithm decides and what humans decide.
- Add human oversight for high-consequence decisions and edge cases.
- Update SOPs, permits-to-work, LOTO, and change management to reflect algorithmic behavior and update cycles.
- Integrate AI into routine safety reviews, JHAs, and exposure assessments-don't create a separate process.
- Monitor psychosocial risks: workload, autonomy, fairness perceptions, and monitoring practices.
- Plan reskilling early. Provide training before changes go live and refresh regularly.
Software design controls (developer/provider responsibility)
- Build transparency: log decisions, expose key performance limits, and identify operating boundaries.
- Run alignment evaluations to confirm behavior matches intended parameters across realistic scenarios.
- Design for safety from the start: fail-safe defaults, safe states on data loss, and graceful degradation.
- Enable independent audits with accessible documentation, test suites, and versioned change logs.
Employers and developers need a shared plan. Collaboration across the lifecycle-procurement, configuration, updates, decommissioning-keeps risks visible and controllable.
Manage AI risks over time
- Schedule independent audits and algorithmic transparency assessments.
- Use safety system and safety case approaches to document hazards, controls, and evidence over time.
- Track model changes, data updates, and retraining events like you track equipment modifications.
- Consider voluntary certification programs that reward predictable, safe system behavior.
- Close the loop: measure outcomes, investigate deviations, and feed lessons back into design and training.
Why this matters to managers
This guidance lets you extend the safety program you already have. You don't need to reinvent your process-just apply it to software-driven systems with the same discipline you use for physical equipment.
Framing AI as a factor that modifies known hazards keeps teams grounded. The algorithmic hygiene view gives you a structured way to assess risks, prioritize controls, and make clear accountability calls.
Quick-start checklist for leaders
- Map where trained algorithms influence operations (decisions, timing, setpoints, alerts).
- Identify affected hazards (physical and psychosocial) and run exposure assessments.
- Assign ownership: who validates outputs, who approves changes, who handles incidents.
- Add human-in-the-loop for high-impact decisions; define escalation criteria.
- Update SOPs, training, and emergency procedures to reflect algorithm behavior and failure modes.
- Require developer documentation: model scope, data sources, limits, and update policy.
- Set up audit schedules, logs, and a change control process for models and data.
- Measure impact: error rates, near misses, stress indicators, rework, and downtime.
Where to start
Review the latest NIOSH resources and align your safety management system with their direction. Pair that with a risk framework your team already understands.
If you're building team capability, consider focused training that connects AI concepts to daily operations and safety goals.
Bottom line
AI changes how risks show up; it doesn't replace them. Treat trained algorithms as part of your system, hold them to the same safety standards, and keep improving with evidence. That's how you protect people and keep operations steady while you adopt new tools.
Your membership also unlocks:
