{
  "id": "knowledge-cutoff",
  "code": "PTL-0112",
  "term": "Knowledge Cutoff",
  "aliases": [
    "training cutoff",
    "training data cutoff"
  ],
  "category": "foundations",
  "definition": "A model's knowledge cutoff is the date after which its training data contains little or no information, so it does not know about later events unless they are supplied in the prompt or retrieved.",
  "description": "Questions about recent events, prices, or current officeholders are answered from stale knowledge unless the system adds search or retrieval. Research has found that a model's effective cutoff can differ from its reported one, and can differ across topics, because web data from earlier years is overrepresented and deduplication is imperfect. Stating the current date in the system prompt helps the model reason about what it might not know.",
  "example": null,
  "broader": [],
  "narrower": [],
  "related": [
    "retrieval-augmented-generation",
    "web-search-tool",
    "hallucination"
  ],
  "introduced": null,
  "added": "2026-10-10",
  "sources": [
    {
      "title": "Dated Data: Tracing Knowledge Cutoffs in Large Language Models",
      "authors": "Cheng et al.",
      "year": 2024,
      "url": "https://arxiv.org/abs/2403.12958"
    }
  ],
  "url": "https://protologue.com/t/knowledge-cutoff/",
  "citation": "Protologue. (2026). Knowledge Cutoff. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0112). https://protologue.com/t/knowledge-cutoff/"
}