Generative AI Engineering

AI Safety and Privacy Terms

Learn the reliability, attack, privacy, moderation, fairness, guardrail, and red-team terms needed for responsible AI operation.

12 matching terms

Reliability

Hallucination (confabulation)

Meaning

A plausible-sounding output that is unsupported by the available evidence or is factually incorrect.

When to use it

Treat it as a risk to manage through grounding, verification, evaluation, and user disclosure.

Practical example

Verify names, dates, numbers, and citations against authoritative sources before publishing.

Attacks

Prompt injection

Meaning

An attempt to manipulate a model with instructions that conflict with the application's intended rules.

When to use it

Separate untrusted content from instructions and enforce permissions outside the model.

Practical example

Treat webpage text as data, not as authorization to call tools.

Attacks

Indirect prompt injection

Meaning

Malicious instructions hidden in external content that an AI system reads or retrieves.

When to use it

Defend document, email, webpage, and tool-result pipelines that feed agents.

Practical example

Ignore instructions embedded in retrieved documents and show them as untrusted content.

Attacks

Jailbreak

Meaning

A prompt strategy intended to bypass a model or application's safety restrictions.

When to use it

Use layered safeguards, monitoring, red-team tests, and least-privilege tools.

Practical example

Block dangerous actions at the tool authorization layer even if the model is manipulated.

Privacy

Personally identifiable information (PII)

Meaning

Information that directly or indirectly identifies a person.

When to use it

Detect, minimize, redact, or obtain proper authority before processing it with AI.

Practical example

Replace customer names and account numbers with placeholders in the evaluation dataset.

Privacy

Data retention

Meaning

Rules governing how long inputs, outputs, logs, and derived data are stored.

When to use it

Define it before production use and align it with purpose, law, contracts, and deletion needs.

Practical example

Delete raw prompts after 30 days while retaining anonymized aggregate metrics.

Privacy

Data residency

Meaning

The geographic location where data is stored or processed.

When to use it

Check it when laws, contracts, or organizational policy restrict data location.

Practical example

Route regional customer data only to approved in-region services.

Controls

Guardrail

Meaning

A technical or procedural control that constrains, checks, or redirects AI behavior.

When to use it

Combine input checks, output validation, permissions, monitoring, and human review.

Practical example

Validate generated SQL against an allowlist before execution.

Controls

Content moderation

Meaning

Detection and handling of content that violates safety or platform policies.

When to use it

Apply it to inputs and outputs with actions suited to the risk level.

Practical example

Flag high-risk content for review and provide a safe alternative when appropriate.

Bias and fairness

Bias

Meaning

A systematic tendency that can produce skewed or unfair outcomes.

When to use it

Measure it across relevant groups, languages, contexts, and failure costs.

Practical example

Compare false rejection rates across supported languages.

Bias and fairness

Fairness

Meaning

The goal and criteria used to assess whether benefits, errors, and treatment are appropriately distributed.

When to use it

Define it for the specific use case because different fairness measures can conflict.

Practical example

Document which fairness metric applies to the hiring-assistance workflow and why.

Assurance

AI red teaming

Meaning

Structured adversarial testing that searches for harmful failures, misuse paths, and weak controls.

When to use it

Run it before launch and after meaningful changes, using realistic threat scenarios.

Practical example

Test prompt injection, data leakage, unsafe tool actions, and multilingual bypass attempts.