AI Doesn't Know the Standard Changed
AI can cite real but outdated regulations. Reliable compliance workflows verify authority, version, effective date, and provenance at delivery.
- AI can cite a real regulation that is no longer current, producing a document that survives plausible internal review and fails at audit.
- A recent practitioner discussion about an outdated OSHA requirement is a useful failure signal, not evidence of how often the problem occurs.
- Existence checks are insufficient. A compliance workflow must verify the current version, effective date, and authority of each material citation.
- Current tools and retrieval can reduce stale-knowledge risk, but the defensible control is source-of-record verification at delivery.
- For compliance systems, provenance and freshness are part of system quality alongside model capability.
A discussion among safety practitioners describes a deceptive AI failure: a generated document cited a real OSHA standard and a valid section, but the requirement came from an older version. The current rule said something different.
One thread cannot establish prevalence. It does expose the mechanism. A fabricated citation invites suspicion. A superseded citation looks legitimate, resolves to a real authority, and can pass a quick review.
The Wrong Frame
Hallucination controls ask whether a source exists and supports the claim. Temporal controls ask another question: is this the authoritative source now? A system can pass the first test and fail the second.
A real citation is not necessarily a current citation. In regulated work, validity depends on authority, version, effective date, and jurisdiction—not existence alone.Base-model knowledge can be stale. Retrieval helps only when its corpus is current, correctly versioned, and scoped to the applicable authority. Browsing or tool use can improve freshness, but it does not prove that the system selected the controlling source or captured a recent amendment.
Three Temporal Failure Modes
Superseded standards. The cited document was once valid but has been replaced. Old copies remain searchable, so both people and models can retrieve them.
Amended regulations. The identifier still resolves, but the operative language has changed. A citation-level existence check will not detect a changed requirement inside the same part or section.
Withdrawn guidance. An agency has removed or disavowed guidance without publishing a direct replacement. In that case, the right answer may be that no current guidance supports the claim.
Why Retrieval Is Not Enough
A retrieval-augmented system inherits the update discipline of its corpus. If an old document remains indexed without an effective-date boundary, the model can cite it confidently. Retrieval has moved the freshness dependency; it has not removed it.
RAG does not solve temporal drift by itself. It makes corpus currency, document lineage, and version selection part of the reliability boundary.The same constraint applies to web-enabled agents. A search result is not a source-of-record check. Ranking, cached copies, jurisdictional ambiguity, and publication lag can still put the wrong document in context.
Verification as Architecture
A defensible workflow verifies every material citation against the current authoritative publication at delivery time. When an official API exists, use it. When it does not, use the regulator's current publication and retain enough provenance to reproduce the check.
- Generate or retrieve the candidate requirement and citation.
- Resolve it against the authoritative source, such as the current eCFR publication.
- Record the source URL, jurisdiction, version or effective date, and retrieval time.
- Flag mismatches and ambiguous authority for human review; do not silently substitute text.
- Repeat the check when the document is delivered or materially revised.
This is the Validator Asymmetry Principle applied to temporal risk: when an authoritative source is queryable, checking a proposed citation can be more reliable than asking a model to remember the current rule.
What the Practitioner Signal Changes
The safety discussion shows why a human workaround often survives after AI adoption: practitioners still open the regulator's site and compare the text. The architecture should preserve that control while making it traceable and efficient.
For regulated decisions, freshness is not a data-maintenance detail. It is a product capability with an auditable control path.Model quality matters, but it cannot certify temporal validity. Trust comes from combining capable generation with current authority, recorded provenance, and a review path for unresolved changes.
Architecting Intelligence: Join the early-reader list
Related Reading:
- The Validator Asymmetry Principle — why validation can buy reliability more cheaply than generation quality
- The End of Determinism — the boundary between deterministic software and probabilistic behavior
- The Context Window Fallacy — why more context does not replace information architecture