Safety Norms: Machine Learning and Artificial Intelligence Engineering

1Purpose, scope, and safety principles

This document provides safety norms for machine learning and artificial intelligence engineering. Its purpose is to prevent injury, property loss, and public harm arising from the facilities, hardware, automated systems, and decisions associated with building and running these systems. It is educational guidance and a starting point for locally approved procedures. It does not replace professional authorization, formal training, supervision, or emergency services.

The scope covers model development laboratories, accelerator clusters, edge devices, robotics and autonomous test areas, data collection sites, and any software pipeline that can trigger a physical action. In this document, MANDATORY means a rule that must be followed without exception unless the responsible safety lead approves a documented alternative. RECOMMENDED means good practice that teams should adopt unless a documented reason exists not to. Laws, standards, codes, and manufacturer instructions are jurisdiction dependent and must be verified with competent authorities or qualified professionals.

• Eliminate hazards at the design stage before relying on warnings or personal protective equipment.
• Treat any model that controls energy, motion, or a process as machinery, not as ordinary software.
• Stop work when a hazard is unclear, and escalate rather than guessing.

2Roles, competence, and accountability

The engineering manager is accountable for the safety case of each project, including risk assessments and approval of physical tests. The facilities or data center operations lead controls power, cooling, fire systems, and rack access. A lab safety coordinator oversees wet labs and bench equipment. The safety officer reviews high risk activities and can halt work. Machine learning engineers are responsible for the safe configuration of jobs and models they launch, and for reporting hazards they observe. Data stewards protect information, and the security lead handles compromise risks.

Competence determines authority. ML engineers may run software jobs and operate approved test equipment after site induction. Only qualified persons may perform electrical work, open power distribution units, service cooling loops, handle refrigerants, or lift heavy equipment into racks. Work on high voltage or energized equipment requires qualification under local law and site policy.

• MANDATORY: Assign a named owner to every job that controls hardware or a physical process.
• MANDATORY: Engineers must not perform electrical, refrigerant, or structural work they are not authorized to do.
• MANDATORY: Record training and authorization for each person before first access to restricted areas.
• RECOMMENDED: Name a deputy for every critical role so work does not stop during absences.

3Course specific hazards and risk assessment

Realistic hazards in this field include electrical shock and arc flash from high current racks and busbars, fire from overloaded circuits or failed battery and UPS systems, burns from hot components and exhaust air, leaks or sprays from liquid cooling, and crush or pinch injuries from robotic arms, mobile platforms, and test rigs. Other hazards include lifting injuries when moving servers, musculoskeletal strain from long workstation sessions, hearing damage from fans and test cells, laser or radio emissions from sensors, and chemical exposure in wet labs. Data hazards include exposure of personal information in training sets and unauthorized disclosure through model outputs.

Risk is rated by combining likelihood and severity into low, medium, high, or critical levels. Any high or critical activity requires a documented controls plan before work starts. A risk assessment must be completed before the first run of new hardware, a new robot policy, a changed power cap, or a new sensor configuration. Example: a reinforcement learning policy controlling a robotic arm in a shared lab has a crush hazard rated high. Controls include a fixed guard, speed and force limits enforced outside the model, a hardware emergency stop, and a trained observer present during tests. Regulated domains such as healthcare, finance, transport, and critical infrastructure require additional professional review and approval.

• MANDATORY: Complete and sign a written risk assessment before first physical or high load operation.
• MANDATORY: Reassess risk after hardware changes, model version changes that affect actions, or any incident.
• RECOMMENDED: Review hazards with the whole team at project kickoff and keep a visible hazard register.

4Preventive controls and safe operating practices for compute

Compute equipment is often the largest source of physical risk in ML work. Racks draw substantial electrical current, run hot, and may use liquid cooling. Engineers should assume that any energized component can cause serious injury. Only facilities staff may open power distribution units, busbars, or rack power shelves, and all such work follows the site lockout tagout procedure. Engineers must never bypass interlocks, jumper sensors, or disable fire suppression to finish a training run.

Long running training and inference jobs can run unattended for days. Operators must confirm that alarms, shutoffs, and monitoring are active before leaving a job unattended. Jobs that write to hardware management interfaces, change fan curves, or alter power limits require approval from facilities. Job schedulers should enforce power and thermal caps so that aging circuits are not held at sustained maximum load.

• MANDATORY: Work only within approved rack layouts and keep aisles, egress paths, and floor tiles clear of cables and equipment.
• MANDATORY: Verify rated power and cooling capacity with facilities before adding an accelerator node or raising a power cap.
• MANDATORY: Label circuits, coolant lines, and emergency shutoffs, and post a responsible contact at each row.
• MANDATORY: Lift servers with mechanical aids or by two trained people, and never reach into an energized rack.
• RECOMMENDED: Configure temperature and power alarms with named responders and test them on a schedule set by facilities.
• RECOMMENDED: Keep a spare hardware inventory so failed components are replaced by qualified staff rather than improvised repairs.

5Robotics, autonomy, and process safeguards

When a model commands a motor, valve, vehicle, drone, or industrial process, the software becomes a machine control system. Engineers must apply the same discipline used for conventional machinery, including risk assessment, guarding, and independent safety functions. Machinery and functional safety standards are jurisdiction dependent. Before any deployment that can injure people or damage property, consult competent authorities and qualified safety engineers.

Learned policies can behave unexpectedly when sensors fail, lighting changes, or unfamiliar objects appear. Safety functions must therefore sit outside the model. Controllers should enforce speed, force, and workspace limits regardless of the model output, and hardware emergency stops should remove actuator power without relying on software.

• MANDATORY: Run physical tests only in marked and barriered zones with a trained observer holding an accessible emergency stop.
• MANDATORY: Provide a hardware emergency stop that cuts actuator power independently of the software.
• MANDATORY: Enforce speed, force, and workspace limits in the controller, outside the learned model.
• MANDATORY: Test new policies first in simulation, then at reduced speed with a tethered or caged setup, before any unrestricted run.
• MANDATORY: Keep people out of the working envelope during autonomous operation, using fixed guards, light curtains, or laser scanners where the risk assessment requires them.
• RECOMMENDED: Log sensor inputs, model version, and commanded outputs for every physical run so incidents can be reconstructed.
• RECOMMENDED: Test behavior with sensor dropouts, glare, and unfamiliar objects before expanding the operating scope.

6Personal protective equipment, environmental controls, and information safeguards

Personal protective equipment is selected from the task risk assessment. Safety glasses are required near bench work, cutting, or cable stripping. Closed toe shoes are required in server rooms and test areas. Cut resistant gloves are used when handling sheet metal or sharp hardware. Hearing protection is required where a noise assessment or site rule identifies a hazard. Nitrile gloves and lab coats apply in wet labs. Ratings and suitability must be verified with manufacturers and competent authorities. Insulated tools and rated gloves are used only by qualified persons for work that is authorized.

Environmental controls include spill kits for coolant and battery electrolytes, secondary containment for liquids, and ventilation checks where equipment generates heat or fumes. Lithium batteries, electronic waste, and refrigerant must be handled by certified recyclers or technicians under local rules. Data centers should track energy and water use where required and should review cooling setpoints to avoid unnecessary consumption.

Information safeguards protect people as well as systems. Training sets often contain personal or sensitive data, so teams should apply data minimization, role based access, and encryption in transit and at rest according to policy. Restricted datasets must not be copied to personal devices. Model endpoints that trigger external actions require authentication, rate limits, and monitoring for prompt injection and tool misuse. Privacy and data protection laws are jurisdiction dependent and must be verified with qualified counsel.

• MANDATORY: Select and inspect PPE against the documented task assessment before each activity.
• MANDATORY: Dispose of batteries, electronic waste, and refrigerant only through approved channels.
• MANDATORY: Do not copy restricted datasets to personal devices or unapproved storage.
• MANDATORY: Require review and authorization before a model is connected to a system that can change physical or financial state.
• RECOMMENDED: Adjust workstation height, chair, and monitor position, and take scheduled breaks from screen work.

7Incident reporting, emergency response, and escalation

Report every injury, near miss, electrical fault, coolant leak, unexpected robot motion, unsafe model behavior affecting physical systems, and suspected security compromise within the same shift. Reporting should be simple and blameless. The person reporting does not need to prove cause. Immediate actions are to stop the job using the kill switch, move people away from the hazard, and call the emergency number for fire, electric shock, serious injury, chemical exposure, or a person struck by machinery. Untrained persons must not attempt rescue from energized equipment or enter a hazard area without the protection the situation requires.

After an incident, the responsible safety lead decides when equipment may be restored. Logs, checkpoints, and sensor data must be preserved without overwriting. If a deployed system has caused physical, financial, or privacy harm to the public, the product safety lead must start the withdrawal, correction, and notification process. Reporting obligations and timelines vary by jurisdiction and must be confirmed with competent authorities or qualified professionals.

• MANDATORY: Call local emergency services for fire, electric shock, serious injury, or chemical exposure.
• MANDATORY: Do not restore power or motion until the safety lead signs off in writing.
• MANDATORY: Preserve logs, model versions, and sensor data related to the incident.
• MANDATORY: Notify the security lead immediately for suspected compromise, following the response plan timeframe.
• RECOMMENDED: Hold a blameless post incident review within a period set by the organization, and track corrective actions to completion.

8Records, training, inspection, and review

Records provide evidence that controls were planned and followed. Required records include risk assessments, permits to work, maintenance logs, training and authorization records, test plans, model cards that state safety limits, and incident reports. Permits to work are required for high voltage work, work at height on racks or roofs, hot work, and entry into confined spaces such as cooling tanks. Each permit names the authorized person, the hazards, the isolation steps, and the check needed before work is released.

Training covers site hazards, lockout awareness, emergency stop operation, incident reporting, and the specific norms for each project. Refresher training follows a schedule set by the organization. Inspections include weekly walkthroughs of labs and rack rows, scheduled checks of fire detection and suppression by facilities, calibration of safety sensors, and review of alarm logs. This document and the project norms are reviewed after every incident, major hardware change, or material model change, and at least annually.

• MANDATORY: Issue a permit to work before high voltage, work at height, hot work, or confined space entry.
• MANDATORY: Keep maintenance and inspection logs for all safety critical equipment.
• MANDATORY: Complete training before access to restricted areas and record completion.
• RECOMMENDED: Track near miss counts, reporting times, and corrective action closure to guide improvement.

Implementation checklist

• Confirm a signed risk assessment exists before any new hardware, robot policy, or power change.
• Verify every physical test area has barriers, a trained observer, and a hardware emergency stop.
• Check that software limits and safety functions sit outside the learned model.
• Confirm only qualified staff are performing electrical, refrigerant, or lifting work under a permit.
• Inspect PPE against the task assessment and replace damaged items before use.
• Confirm batteries, electronic waste, and coolant are routed to approved disposal channels.
• Ensure restricted data is stored, accessed, and shared only under approved controls.
• Report every incident and near miss within the shift and preserve logs before restarting.
• Schedule the next review of these norms and the project safety case.