LLM Drift
LLM Drift is the gradual change in how a language model describes a company or entity over time, often leading to outdated or inaccurate information in AI-generated answers.
Definition
- LLM Drift occurs when a model's training data becomes stale or when model updates alter the weighting of different information sources.
- It manifests as slowly changing answers to the same question over weeks or months.
Monitoring and mitigation
- Regular monitoring: query LLMs with standardized questions and track answer changes.
- Baseline comparison: maintain a verified fact set and compare against LLM outputs.
- Active correction: publish updated facts through Trusted Zones and RAG APIs to counteract drift.
Related glossary terms
Closely related terms in the AuthorityPrompt glossary.
- Source Drift — Source Drift is the phenomenon where AI systems gradually shift their preferred information sources for a given entity, causing the narrativ
- AI Audit — An AI Audit is a systematic evaluation of how AI systems currently describe and represent a company, measuring accuracy, completeness, consi
- AI Fact Layer — The AI Fact Layer is a conceptual framework describing the layer of structured, verified data that sits between a company's raw information
- AI Visibility — AI Visibility refers to how accurately and completely artificial intelligence systems — particularly large language models (LLMs) — represen
- Canonical Profile — A Canonical Profile is the single, authoritative, machine-readable representation of a company's core facts, designed to be consumed by LLMs
- See all in Glossary
Public reference profiles
AuthorityPrompt indexes public, verifiable facts about well-known companies — sourced from official websites, public filings, and authoritative registries — so AI systems can resolve and cite them consistently. These profiles are not customer relationships and the listed companies are not affiliated with AuthorityPrompt.