Material change

The functional test this topic applies to amendment decisions: whether a filed algorithm's actual current operation, as it would be perceived by a member of the public or a regulator reading the original filing, still matches what the filing describes. A change is material if the filed description no longer accurately represents the algorithm's real behavior, regardless of whether the declared taxonomy category has technically stayed the same.

A change to a system that could plausibly move its failure modes, and therefore the trigger for a targeted literacy refresh. Distinguished from a cosmetic or back-end change that leaves behavior the same, which is logged rather than retrained.

The framework's middle risk tier, exemplified by model updates and autonomy adjustments, which require a full governance review involving the relevant named roles.

A modification to a cross-border transfer's underlying facts (a new overseas recipient, a shift in data volume across a threshold, an added data category) significant enough to require re-running the threshold decision tree, and in some cases (such as a changed recipient under an existing security assessment) to require a fresh mechanism application rather than a simple renewal.

A change to an already-filed AI system (in automation level, data categories, or use-case boundaries) significant enough that the system's original regulatory filing or disclosure no longer accurately describes its current behavior, triggering a duty to refile or amend rather than wait for a scheduled review.