The gradual change in the meaning of a term or data definition across systems or over time.
Semantic drift refers to the gradual, often unnoticed change in the meaning or interpretation of a term, metric, or data definition across different systems, teams, or time periods. It occurs when the same word or concept is used in slightly different ways across an organization — 'active user' might mean 'logged in within 30 days' in one system and 'made a purchase within 90 days' in another. Over time, these small differences compound into significant inconsistencies.
Semantic drift has become one of the most expensive problems in enterprise AI. When AI agents query multiple data sources with drifted definitions, they produce contradictory outputs that erode trust. Organizations that have invested in semantic layers and ontologies to combat drift are seeing dramatically better AI reliability. Detecting and measuring semantic drift programmatically — using tools like semantic similarity checkers — is now a standard practice in AI governance.
Semantic drift can be detected by comparing the current definition of a term against a canonical reference definition using semantic similarity scoring. A similarity score below a threshold (typically 0.85) indicates meaningful drift. Automated drift monitoring systems track definitions across all data catalogs, documentation, and code, alerting data teams when definitions diverge beyond acceptable thresholds.
A SaaS company tracks the definition of 'Monthly Active User' (MAU) across their product analytics system, their billing system, and their investor reports. Over 18 months, the product team quietly changed the definition to include free trial users, while the billing system still counted only paid users. An AI agent querying both systems for MAU reports wildly different numbers, causing confusion in a board meeting.