THE CORE AIM OF VOCABULARY MASTERY · AI SAFETY VOCABULARY · HAZARD → RISK → EVALUATE → MITIGATE → MONITOR
AI safety vocabulary is the language used to describe risks, failure modes, safeguards and evaluation practices for artificial intelligence systems. Terms such as hazard, risk, alignment, robustness, guardrail, red teaming, evaluation, oversight, misuse and monitoring matter because responsible AI requires more than asking whether a model is capable.
The core aim of vocabulary mastery for AI safety vocabulary is risk-and-control clarity. Learners should be able to identify what could go wrong, who or what could be harmed, how likely and severe the outcome is, what mitigation reduces the risk and what evidence shows the safeguard works.
This page is the AI Safety Vocabulary owner inside the eduKateSG Vocabulary hub. For wider AI concepts, use Artificial Intelligence Vocabulary. For generative-system failure modes, use Generative AI Vocabulary.
Central proposition: AI-safety vocabulary is mastered when vague concern becomes a specific hazard, characterised risk, tested mitigation and monitored deployment condition.
The AI Safety Vocabulary Router
- Hazard: failure mode, misuse, harmful output.
- Risk: likelihood, severity, exposure, impact.
- Evaluation: benchmark, test, red team, robustness.
- Mitigation: guardrail, filter, access control, constraint.
- Oversight: human review, approval, escalation.
- Operations: monitoring, incident, rollback, remediation.
Hazard and Risk Are Different
A hazard is a source or scenario with potential for harm. Risk considers the likelihood and impact of that harm under actual conditions. Keeping them separate helps teams prioritise mitigations rationally.
Robustness and Red Teaming
Robustness is the ability of a system to continue behaving acceptably when inputs or conditions change. Red teaming is structured adversarial testing intended to uncover weaknesses and failure modes so defences can improve.
Alignment and Oversight
Alignment broadly concerns whether AI behaviour follows intended goals and constraints. Oversight concerns how people supervise, review or intervene in system behaviour.
Monitoring and Incident Response
Safety continues after launch. Monitoring looks for drift, misuse and unexpected failure. Incident vocabulary includes detection, containment, mitigation, escalation, rollback and remediation.
How to Learn AI Safety Vocabulary
- Choose a concrete AI use case.
- List plausible hazards.
- Characterise impact and exposure.
- Define measurable evaluations.
- Map mitigations to risks.
- Specify human approval points.
- Monitor after deployment and revise controls.
Common Mistakes
Using safety as a synonym for accuracy
Repair: include misuse, robustness, privacy, security and operational harm.
Confusing hazard and risk
Repair: separate potential harm from likelihood and impact.
Treating one benchmark as proof of safety
Repair: use layered evaluations and deployment monitoring.
Frequently Asked Questions
What is AI safety vocabulary?
It is the language used to describe AI hazards, risks, evaluations, safeguards, oversight and operational response.
What is red teaming?
It is adversarial testing used to discover weaknesses so defences can improve.
What is a guardrail?
It is a technical or procedural constraint designed to reduce unwanted system behaviour.
The AI Safety Vocabulary Standard
Mastery means stating the hazard, characterising the risk, naming the mitigation and identifying the evidence used to test it.
That is the standard: safety language precise enough to turn concern into engineering and governance action.