Search the Atlas

Search risks, controls, and glossary terms

MediumKnowledge StalenessTheoretical

Stale knowledge / training cutoff

Data & Input Integrity

Description

The model's knowledge is frozen at its training cutoff, producing confidently outdated answers about rates, regulations, or products.

Example scenario

The assistant quotes a superseded repo rate because its training data predates the latest RBI revision.

Real-world evidenceTheoretical

Accessibility audits and academic studies consistently show AI-driven interfaces (voice assistants, conversational agents) perform worse for older adults and users with disabilities, and advocacy groups have raised concerns. However, a documented production incident where an AI interface's inaccessibility directly caused quantifiable harm to a vulnerable individual or group has not been publicly confirmed at scale.

No public incident on record — evidence level: Theoretical

Primary mitigations

  • RAG grounding to live sources
  • cutoff disclosure
  • freshness checks
  • scheduled re-grounding
  • retrieval-first answering.

Detection signals

Date-stamped fact checks; freshness SLA monitoring; outdated-claim sampling.

Mitigating controls

5
Non-agentic controls

Related risks in Data & Input Integrity