Semantic Drift: How Definitions Quietly Change Over Time

TL;DR

• Drift is usually accidental and cumulative.
• AI amplifies drift by spreading inconsistent definitions.

The problem

• Metric definitions change as business rules evolve.
• Old reports and AI prompts still assume previous meanings.

Why it matters

• Drift makes trend analysis misleading.
• AI may mix old and new definitions in a single answer.

Symptoms

• Historical comparisons suddenly look “off.”
• Same question yields different results before and after a model update.

Root causes

• Changes to DAX logic without updating descriptions.
• No change control for metric definitions.

What good looks like

• Definition changes are versioned and documented.
• Historical data is recomputed or clearly segmented.

How to fix

• Add explicit version notes to metric descriptions.
• Implement change review for KPI definitions.
• Communicate drift events to report owners.

Pitfalls

• Assuming “the data didn’t change” means the metric didn’t change.
• Leaving old measures in place without deprecation.

Checklist

• Definition change log exists.
• Metrics have owners and review cycles.
• Old definitions are deprecated or versioned.

Framework placement

Primary ARF layer: Semantic Integrity. Diagnostic bridge: semantic-reliability, change-reliability.