State drift occurs when an AI system keeps returning a “state of the world” that is no longer true: price, availability, conditions, policies, hours, modalities. This is not necessarily a hallucination. It is often an outdated stabilization: the model relies on historical signals or secondary sources that no longer reflect the current state.
Operational definition
State drift: a persistent divergence between the real state (as published and applicable) and the interpreted state (as returned by an AI system), caused by source inertia, unfavorable routing, or the absence of an enforceable update mechanism.
Why it happens
- Dominant secondary sources: directories, comparison sites, articles, caches, republished pages.
- Unstructured updates: information is modified, but without a strong signal of change.
- Biased routing: the system retrieves frequent sources before the primary source.
- Perimeter ambiguity: conditions differ by region, date, product, or channel.
- Semantic compression: nuance is reduced in favor of an average “plausible” state.
Typical examples
- A price from before a promotion or adjustment.
- A product declared “in stock” when it has been discontinued.
- A policy (returns, refunds, warranty) that changed but is still returned in its former form.
- Outdated opening hours or service conditions.
Observable symptoms
- The answer is stable and repeatable, yet contradicts the official information.
- Citations, when they exist, point toward secondary sources or undated pages.
- The model answers correctly in one case and incorrectly in another, depending on formulation.
Why this is a major risk
- Commercial risk: bad information means lost conversion and overloaded support.
- Reputational risk: the AI “speaks” on your behalf and gets it wrong in a plausible way.
- Compliance risk: some policies belong to legal, regulatory, or contractual regimes.
- Interpretive debt: the longer it lasts, the more costly it becomes to correct.
Rapid diagnosis
- Isolate the contested state: which value is outdated (price, inventory, policy)?
- Identify the canonical source: where is the real state published and enforceable?
- Map the secondary sources: where does the old state continue to exist?
- Test stability: query variants, languages, engines, contexts.
Governed remediation strategies
1) Make the state enforceable
- Create a pivot page for the “Current state” (price, policy, conditions) with an explicit update date.
- Define the perimeter: region, product, channel, period.
2) Structure the change
- Make explicit what changed, when, and why, even briefly.
- Avoid undated pages that all look alike.
3) Reduce dependence on secondary sources
- Connect the current state to heavily cited pages (services, FAQ, pillar pages).
- Update connected pages that still contain the former state.
4) Act exogenously
- Correct listings, directories, and comparison pages when possible.
- Add clarifications wherever ambiguity is creating the drift.
Recommended links
FAQ
Is state drift the same thing as hallucination?
No. A hallucination invents. State drift often returns a real state… but an outdated one, because the system relies on historical or secondary signals.
Why does the AI not use the official page?
Because it may be less frequently reused, less structured, or less readily activated than dominant secondary sources.
How can state drift be reduced?
By making the current state explicit, dated, bounded, tied to pivot pages, and by reducing the old state across the surrounding ecosystem.
