Safety Norms: Natural Language Processing

1Purpose, Scope, and Safety Principles

This document sets safety norms for natural language processing work, including coursework, research projects, and applied development. Natural language processing, or NLP, is the practice of building software that reads, classifies, translates, summarizes, or generates human language. Its practical hazards are mostly informational, electrical, ergonomic, and organizational rather than mechanical.

The scope covers people who collect, label, or process text and speech data, train or evaluate models, operate shared servers and GPU workstations, and deploy language tools to users. Where this document says must, the norm is mandatory under the host institution or employer policy. Where it says should, the practice is recommended and may be adapted through local approval. References to law, standards, or manufacturer instructions are jurisdiction-dependent and must be checked against current versions.

• Minimize personal data, limit access to least privilege, keep humans accountable for consequential decisions, and stop work when a required control is missing.
• Every project member follows the mandatory norms regardless of seniority or project size.
• Anyone may pause work that appears unsafe, and no one may be penalized for doing so in good faith.

2Roles, Competence, and Accountability

The project lead owns the safety case for the project. The lead must maintain a data inventory, approve data sources, confirm that a written risk assessment exists before work begins, and ensure that every member knows how to report an incident. The data steward controls storage locations, retention settings, and access grants. The systems administrator maintains servers, network access, backups, and patching. Each member must follow these norms, report concerns, and never bypass a control for convenience.

Competence must be demonstrated, not assumed. Members must complete the institution's data protection and information security training before receiving access to restricted corpora. Work with clinical, legal, financial, employment, or child-related text, or with any dataset that may identify individuals, requires authorization from the data owner and, where applicable, review by an ethics or institutional review body. This is educational guidance and does not replace professional authorization or formal training.

• Supervisors approve access to electrical rooms, server racks, and high-power workstations.
• Students must not perform hardware installation, power distribution work, or cooling maintenance without trained supervision.
• Assign a named deputy for each role so that work is never blocked when one person is unavailable.

3Course-Specific Hazards and Risk Assessment

Hazards in NLP fall into six groups: data exposure, unsafe model behavior, electrical and thermal hazards, ergonomic strain, software supply chain risks, and environmental impact. Rate each hazard by likelihood and severity using low, medium, or high. Any high rated risk must have documented controls and sign-off from the project lead before work continues.

The following examples are typical for NLP projects and must be addressed in every assessment:

• Personal data exposure: names, contact details, health or financial details inside transcripts, chat logs, or scraped web text. Likelihood high, severity high.
• Re-identification: combinations of location, age, and rare job titles can identify people even after names are removed. Likelihood medium, severity high.
• Harmful or misleading outputs: classifiers or generators may produce defamatory, discriminatory, or dangerous advice. Likelihood medium, severity high in advisory settings.
• Model misuse: released models or tools may be repurposed to generate phishing messages or harassment. Likelihood medium, severity medium to high.
• Electrical hazards: GPU workstations, power supplies, power strips, and rack distribution units can overload circuits or cause shock. Likelihood low to medium, severity high.
• Thermal and fire hazards: overheated power supplies, blocked vents, and damaged lithium battery laptops. Likelihood low to medium, severity high.
• Musculoskeletal strain: long sessions at fixed screens and repetitive keyboard work. Likelihood high, severity medium.
• Noise: cooling fans and server rooms can be loud during extended exposure. Likelihood medium, severity low to medium.
• Supply chain: unverified packages, pretrained models, or datasets with malicious code or unclear licenses. Likelihood medium, severity high.

4Preventive Controls and Safe Working Practices

Apply controls in this order: remove the hazard, substitute a safer method, use engineering controls, use administrative controls, and rely on personal safeguards last. In NLP, removing unnecessary personal data is the strongest control and must be the first step in every project. The following norms are mandatory unless the data owner and institution have approved a documented exception.

Electrical and thermal work follows strict boundaries. Workstations and servers must be installed by trained staff or approved vendors. Nobody may open power supplies, replace capacitors, or bypass thermal protections. Keep the clearance specified by the manufacturer around vents and do not stack equipment. Use power strips rated for the load and never daisy chain them. Report any burning smell, crackling, discoloration, or repeated shutdown to the systems administrator immediately.

• Data minimization: collect only the fields needed for the stated purpose and remove direct identifiers before analysis.
• Redaction: apply an approved tool to names, phone numbers, email addresses, account numbers, and addresses. Test it on a sample and record misses in the project file.
• Access: grant read or write rights to named people for named projects. Shared accounts are prohibited for restricted data.
• Storage and transfer: no personal data in public repositories, public notebooks, chat tools, or external model services unless the data owner and institution approve that transfer.
• Secrets: never commit credentials, tokens, or keys to code. Rotate any exposed secret at once.
• Dependencies: pin package versions, install from trusted sources, and review licenses before use.
• Human review: outputs that affect a person's access, pay, housing, health, or legal status must be reviewed by a qualified person before action. Fully automated decisions of this kind require prior written approval.
• Evaluation: test models on representative data, including dialects and sensitive groups where relevant, and document known failure modes before release.
• Deployment guardrails: restrict outputs for advisory uses, apply approved content filters, and give users a visible way to report problems.
• Annotation welfare: annotators receive clear instructions, may skip disturbing content, know a support contact, and may stop a task at any time. Exposure to graphic or abusive text must be limited and rotated.
• Permits: work on high-power equipment, electrical rooms, or rack maintenance requires a written permit signed by the facilities owner before it starts. The permit lists the task, participants, isolation steps, and completion check. Lone working in server rooms during high-power tasks is prohibited.

5Equipment, Environment, Information, and Protective Safeguards

Personal protective equipment is used where a task requires it. Safety glasses are required when handling damaged rack components under supervision. Only qualified electricians may work inside power equipment, using insulated tools and approved isolation. Hearing protection is recommended in server rooms where noise is uncomfortable over extended periods. Exposure limits are jurisdiction-dependent and must be verified locally.

Environmental controls include adequate cooling, dust control, and fire protection. Keep server rooms clean, prohibit food and drink near equipment, and keep flammable storage out of computing areas. Charge lithium battery devices on hard, open surfaces, remove damaged batteries from service, and report them. Dispose of hardware through the approved e-waste process and sanitize or destroy storage media according to the data owner's instructions. Training large models uses substantial electricity, so teams should record compute use, prefer efficient hardware, and schedule long jobs with facilities staff where power limits apply.

• Keep emergency exits, fire extinguishers, and electrical isolation switches unobstructed at all times.
• Store credentials only in approved password managers or secret stores.
• Use encryption at rest and in transit for all restricted text data.
• Check ergonomic setup: monitors at roughly eye level, keyboards at elbow height, and short breaks at intervals agreed with the team.

6Incident Reporting, Emergency Response, and Escalation

Report incidents within the timeframe set by institutional policy, and report at once when anyone is at risk. Incidents include data exposure, leaked credentials, model outputs that cause harm or distress, electrical or fire events, injuries, and near misses where a control failed but no harm occurred. Near misses are reported without blame because they reveal weak controls before someone is hurt.

Respond in this order. Stop the activity and make the area safe if it is safe to do so. Call local emergency services for fire, electric shock, injury, or threats to life. Do not repair power equipment, use water on electrical fires, or touch a person still in contact with a live source. Isolate affected accounts, revoke exposed keys, and preserve logs without deleting evidence. Notify the project lead, data steward, and systems administrator. Where personal data may have been disclosed, the data protection officer or equivalent decides on any legal notification, which varies by jurisdiction. Serious injuries, large data exposures, harmful outputs published to the public, and regulator contact must be escalated to the safety office and legal counsel. Investigations identify causes and corrective actions, not people to blame.

• Each lab and server room posts emergency contacts, exit routes, extinguisher locations, and power isolation points.
• Teams rehearse the emergency plan at intervals set by the facilities owner.

7Records, Training, Inspection, and Review

Keep a project safety file containing the risk assessment, data inventory, access list, redaction and evaluation results, permits, inspection findings, and the incident log. Retention periods are set by institutional policy and law, which vary by jurisdiction, so confirm them with the records office.

Continuous improvement means treating each review as a chance to remove a hazard rather than add paperwork. Share lessons learned in a short written summary for the whole group.

• Training: complete data protection, information security, electrical awareness, and incident reporting training before first access.
• Inspection: check cable condition, cooling, clearances, fire equipment, and storage at intervals set by the facilities owner, and record results.
• Review: update the norms after any incident, new dataset, new deployment, hardware change, or significant change in users.
• Metrics: track near misses reported, redaction test results, access reviews completed, and corrective actions closed. Use these to improve systems, not to punish reporters.

Implementation checklist

• Confirm a written risk assessment exists before starting any project.
• Remove direct identifiers and test redaction on a sample.
• Grant access by named person and least privilege only.
• Keep secrets out of code, notebooks, and public repositories.
• Have a qualified person review every consequential model output.
• Report incidents and near misses immediately through the approved channel.
• Never open power equipment or bypass electrical or thermal protections.
• Keep exits, extinguishers, and isolation switches clear.
• Use a written permit and never work alone on high-power tasks in server rooms.
• Dispose of hardware and storage media through the approved process.
• Record training completion in the project safety file.
• Verify laws and standards with competent authorities for your jurisdiction.