Model Risk Basics: Bias, Drift, Hallucination, and Explainability

A compliance professional new to AI governance often hears these four terms used casually in meetings — bias, drift, hallucination, explainability — without a precise working definition of any of them, which makes it nearly impossible to ask a sharp follow-up question or correctly document a finding. This lesson gives you working definitions precise enough to use in a real review, building directly on the pattern-learning model from Lesson 1.

Four Terms, Four Distinct Risks

Bias. A systematic pattern where an AI system's outputs unfairly disadvantage a particular group, traceable to a corresponding pattern in the historical data it was trained on. A recruitment-screening tool trained on years of hiring decisions that favored one group over another will tend to reproduce that pattern — not through any explicit rule, but because it learned the pattern statistically. Bias is a distinct risk from accuracy: a model can be highly accurate against its own training data's historical outcomes and still be systematically biased, because the training data's outcomes were themselves the product of a biased process.

Model drift. The tendency for an AI system's real-world performance to degrade over time as the environment it operates in changes and no longer resembles the data it was trained on. A fraud-detection model trained on last year's fraud patterns gradually becomes less effective as fraud tactics evolve, not because the model changed, but because the world did and the model's underlying assumptions did not keep pace. Drift is why an AI system's compliance status is not a one-time determination — a system compliant at launch can drift into non-compliance without any code change at all.

Hallucination. The tendency of a generative AI system — most notably a large language model — to produce a fabricated fact, citation, or detail with the same fluent confidence as a correct one. Hallucination is a structural property of how these systems generate text, not an occasional bug, and it is a particularly acute risk anywhere an AI-generated output reaches a customer, a regulator, or a legal document without independent verification.

Explainability. The degree to which a human can understand why an AI system produced a specific output. Some AI systems can provide a reasonably clear account of the factors behind a decision; others are closer to a "black box," where even the people who built the system cannot fully reconstruct why it produced a specific result. Explainability is not an abstract technical nicety — it is frequently a direct regulatory requirement for consequential, individual-level decisions, covered specifically in Lesson 4.

Why an AI System Can Be Both "Accurate" and Non-Compliant

A recurring compliance mistake is treating a vendor's accuracy figure as evidence that a system is safe to deploy without further scrutiny. Accuracy measures whether a model's predictions match its training data's labeled outcomes on average — it says nothing about whether that average conceals a bias pattern against a specific group, whether the model's performance is drifting as conditions change, whether it occasionally hallucinates a fabricated detail, or whether anyone can explain a specific output well enough to satisfy a regulator or a court. A system can score extremely well on accuracy and still fail on any of the other three dimensions.