The claim
Data quality practices were designed largely for deterministic systems. Generative components make it possible to create new values that look correct without being grounded in a source.
What changes
Traditional checks still matter, but they operate after a more fundamental question: should this value exist at all? Provenance, retrieval quality, attribution, and semantic consistency become part of data validation.
A better response
AI output should cross a trust boundary before entering a consequential workflow. That boundary may combine source evidence, deterministic constraints, contextual validation, confidence, and human review.
The question to carry forward
Which AI-generated values can be safely reversed, and which can silently become part of organizational memory?