AI risk terms can sound abstract until a workflow breaks. This glossary explains the words business teams need before they publish AI content, buy AI tools, or deploy AI into operations.
Quick Map
| Term | Plain-English meaning | What to do |
|---|---|---|
| Hallucination | AI presents wrong information confidently. | Verify facts against sources. |
| Bias | Output treats groups or cases unfairly. | Test examples across user types. |
| Privacy risk | Sensitive data is exposed or retained inappropriately. | Limit inputs and check vendor terms. |
| Security risk | AI or connected tools can be manipulated or abused. | Use access controls and monitoring. |
| Copyright risk | Generated content may resemble protected work or use unclear training/input rights. | Review usage rights and avoid copying protected material. |
| Human-in-the-loop | A person reviews or approves AI output. | Set review gates for high-impact work. |
Core AI Risk Terms
Hallucination
A hallucination is an output that sounds plausible but is wrong, unsupported, or fabricated. It is especially risky in legal, medical, financial, technical, and news content.
Bias
Bias means the system produces unfair or skewed results. Bias can come from training data, prompts, business rules, or how people interpret the output.
Privacy Risk
Privacy risk appears when personal, customer, employee, or confidential business data is entered into systems without clear controls. Teams should know what data is allowed, where it is stored, and whether it can be used for training.
Security Risk
Security risk increases when AI connects to tools, files, code, or business systems. Prompt injection, unsafe plugins, excessive permissions, and weak logging are common issues.
Copyright and IP Risk
Generated text, images, and code still need review. Teams should avoid copying protected works, uploading confidential materials without permission, or assuming generated output is automatically safe to commercialize.
Explainability
Explainability means users can understand why an output or recommendation happened. In business workflows, source visibility and audit logs often matter more than perfect technical explanations.
Model Drift
Model drift happens when performance changes over time because data, users, prompts, or model behavior changes. Periodic review is necessary for repeat workflows.
Human-in-the-Loop
Human-in-the-loop means people approve, edit, or reject AI output. It is essential for customer-facing, regulated, sensitive, or high-impact actions.
Risk Triage
| Risk level | Examples | Default control |
|---|---|---|
| Low | Brainstorming, formatting, internal outlines | Human skim |
| Medium | Customer emails, summaries, reports | Review before sharing |
| High | Finance, legal, HR, security, compliance, customer-impact actions | Approval workflow and logs |
Bottom Line
AI risk management starts with shared language. Once a team understands these terms, it can design better prompts, safer workflows, and clearer review rules.