Glossary
Updated
What counts as AI risk?
AI risk is any way an AI system could cause harm: to the people it makes decisions about, to the people who use it, or to the organization that runs it. Like other risks, it has two parts: how likely the harm is, and how bad it would be. AI adds some twists. The same system can give different answers to similar inputs, it can change when its model is updated, and people can use it for things nobody planned.
What are the main types?
Most AI risks fall into a few groups. Performance risk is outputs that are wrong, including confident fabrications, often called hallucinations. Fairness risk is outcomes that disadvantage people because of a protected characteristic, also known as AI bias. Privacy risk is personal or confidential data exposed through training data, prompts or outputs. Security risk covers attacks such as prompt injection, data poisoning and jailbreaks.
Third-party risk is a vendor changing a model, a term or a feature without telling you. Legal risk is a use that breaks a law, such as a practice the EU AI Act bans. With AI agents there is also action risk: an agent doing something it was never meant to do, with the access it was given. The OWASP Top 10 for Agentic Applications lists the most common ways that happens, and NIST's Generative AI Profile sets out 12 risk categories for generative AI, from confabulation and data privacy to information security and intellectual property.
How is AI risk measured?
Most organizations rate each risk on a likelihood scale and an impact scale, often five points each, and combine them into one rating. They rate it twice: once before any controls, the inherent risk, and once after, the residual risk. The gap shows what the controls are worth. Some risks can be measured directly, such as error rates or differences in approval rates between groups, and those numbers can be tracked over time as key risk indicators.
Is AI risk the same as a risk tier?
No. Laws such as the EU AI Act sort AI uses into tiers: prohibited, high-risk, transparency duties only, and minimal. A tier is a legal classification of the use, and it decides which duties apply. AI risk is the chance of harm from a particular system in a particular setting. A minimal-risk system under the Act can still carry serious business risk, such as an internal tool that leaks confidential data, and a high-risk system can be well controlled. You need both views. Our EU AI Act risk classification guide covers the tiers.
To manage these risks as a program, see AI risk management. In Enzai, each risk is attached to the AI system it belongs to and scored before and after treatment, in one register shared across teams. See AI risk and compliance.
Example
A retailer uses a third-party chatbot for customer service, and its review rates three risks. The chatbot could invent a refund policy the company doesn't offer: likely, with moderate impact. A customer could trick it into revealing another customer's order: unlikely, with high impact. The vendor could switch to a new model without notice: likely, with moderate impact. The retailer limits the chatbot to its published policy pages, lets it see only the signed-in customer's own orders, and adds a contract clause requiring notice of model changes. It then re-rates the residual risks and reviews them every quarter.