Canonical concept · WikiSure
Semantic Drift
The silent divergence of a term's meaning across teams, systems, or time — the failure mode that semantic governance exists to prevent.
Definition
Semantic Drift is the silent, often unnoticed divergence of a term's meaning across teams, systems, geographies, or time. Drift is not a data quality problem — the values are correct; it is a meaning problem — the values answer different questions. Left unmanaged, drift produces inconsistent AI decisions, broken integrations, and reporting that disagrees with itself.
Business context
Drift compounds. Every new team, integration, or AI agent that consumes a term without resolving against a canonical source introduces a new variant. By the time drift is visible — usually as a regulatory finding or a reconciliation failure — the cost of correction is an order of magnitude higher than continuous governance would have been.
Insurance example
An insurer's 'active policy' meant 'premium paid in the current period' in the billing system, 'coverage in force at report date' in actuarial, and 'no cancellation event recorded' in claims. Three departments produced three different active-policy counts every quarter. WikiSure surfaced the drift, made the differences explicit, and approved one canonical definition with owner accountability.
Governance example
A risk model trained on 2022's definition of 'high-risk counterparty' silently underperforms in 2026 because the regulatory definition tightened twice in the intervening period. Semantic drift detection flags the model's reference definition as stale and triggers a governance review before the model's next deployment cycle.
Related concepts
Frequently asked
- How does WikiSure detect drift?
- WikiSure continuously compares candidate definitions (from documents, systems, and agent prompts) against the approved canonical version using an alignment score. Definitions that fall below the alignment threshold are flagged as drifted.
- Is semantic drift the same as data drift?
- No. Data drift describes a change in the distribution of values; semantic drift describes a change in what the values mean. Data drift can be fixed with retraining; semantic drift can only be fixed with governance.
- What is an acceptable level of drift?
- For regulated decisions, zero — the canonical definition is the only acceptable reference. For exploratory analysis, drift is tolerable as long as the analysis is not promoted to a governed surface without re-alignment.