AI Safety & Compliance Glossary
37 plain-English definitions of the terms that show up in AI security reviews, procurement checklists, and regulator questionnaires. Written by the BotDoors team, updated as the field moves.
Alignment
- AI AlignmentAlignment is the problem of making AI systems reliably pursue the goals humans actually intend.
- RLHF (Reinforcement Learning from Human Feedback)RLHF trains a model to prefer outputs humans rated as better.
- Constitutional AIConstitutional AI trains a model against a written set of principles instead of human ratings alone.
- SycophancySycophancy is an AI model's tendency to tell users what they want to hear instead of what is true.
- InterpretabilityInterpretability is the science of understanding what is actually happening inside an AI model.
- Mechanistic InterpretabilityMechanistic interpretability reverse-engineers neural networks into human-readable circuits and features.
Compliance
- EU AI ActThe EU AI Act is the European Union's risk-tiered law regulating AI systems placed on the EU market.
- NIST AI Risk Management FrameworkNIST AI RMF is a voluntary US framework for identifying, measuring, and managing AI risks.
- ISO/IEC 42001ISO/IEC 42001 is the international management-system standard for AI, analogous to ISO 27001 for security.
- High-Risk AI SystemUnder the EU AI Act, high-risk AI systems are those affecting safety, fundamental rights, or critical services.
- PII (Personally Identifiable Information)PII is any data that can identify a specific person, directly or in combination with other data.
- GDPRGDPR is the EU regulation governing personal-data processing, with global reach and fines up to 4% of turnover.
- Shadow AIShadow AI is the unsanctioned use of AI tools by employees, outside the organization's governance.
- AI GovernanceAI governance is the set of policies, roles, and controls that keep an organization's AI use safe, legal, and accountable.
Ethics
- Model CardA model card documents what an AI model is, how it was trained, where it works, and where it fails.
- System CardA system card documents the full deployed AI system, not just the underlying model.
- Training DataTraining data is the corpus a model learns from; its provenance, licensing, and quality bound everything the model can do.
- Content ProvenanceContent provenance is verifiable metadata proving where a piece of content came from and how it was edited.
- DeepfakeA deepfake is synthetic audio, image, or video generated to convincingly impersonate a real person or event.
- AI WatermarkingAI watermarking embeds an imperceptible signal in generated content so it can later be identified as AI-made.
Evaluation
- AI HallucinationAI hallucination is when a model confidently states something false or fabricated.
- Retrieval-Augmented Generation (RAG)RAG grounds an AI model's answers in retrieved documents instead of relying on its parametric memory.
- GroundingGrounding ties an AI answer to verifiable sources so claims can be checked.
- AI EvalsEvals are structured tests that measure whether an AI system meets quality, safety, or capability targets.
- Prompt EngineeringPrompt engineering is the practice of writing and structuring instructions so an AI model gives better, more reliable answers.
- Context WindowA context window is the maximum amount of text an AI model can read and remember at once in a single conversation.
- Tokens (AI)Tokens are the small chunks of text, roughly word pieces, that AI models read and write, and the unit used to measure usage and limits.
Safety
- AI GuardrailsGuardrails are runtime checks that constrain what an AI model can say, do, or call.
- AI Red-TeamingRed-teaming stress-tests an AI system by actively trying to break its safety, security, and factual claims.
- Agentic AIAgentic AI describes systems that plan multi-step actions and use tools autonomously to reach a goal.
- Tool UseTool use lets an LLM call external functions — search, code, APIs — to extend what it can do.
- AI Content ModerationAI content moderation uses classifiers to filter unsafe user input and model output at runtime.
Security
- Prompt InjectionPrompt injection is an attack where hidden or hostile text overrides an AI model's original instructions.
- Indirect Prompt InjectionIndirect prompt injection hides adversarial instructions inside content the AI later reads, not the user's message.
- JailbreakA jailbreak is a prompt that bypasses an AI model's built-in safety rules.
- Differential PrivacyDifferential privacy adds mathematically calibrated noise so no individual record can be reverse-engineered from a model.
- Data LeakageData leakage is when sensitive information escapes into training data, logs, or model outputs.