citation
In plain wordsA pointer from something an AI says to the source that supports it — a link, document or quoted passage — so people can check the claim for themselves.
In AI systems, a reference attached to generated text that identifies the source supporting a specific claim — a document, passage, web page or quoted span — typically produced by retrieval-augmented generationretrieval-augmented generation, n.: Giving an AI an open-book exam: before answering, it looks up relevant passages in a set of documents and bases its reply on them, often citing where each point came from. or agentic RAGagentic RAG, n.: Retrieval in which the AI runs its own research: it decides what to search for, looks at the results, searches again if needed, and only then answers. systems that instruct or train the model to cite what it retrieved. Citations make output verifiable and are evaluated by citation recall (is every claim supported by some cited source?) and citation precision (does each cited source support its claim?), often judged with entailment models or an LLM-as-a-judgeLLM-as-a-judge, n.: Using one AI language model to grade or compare the answers of another, as a cheaper and faster stand-in for human reviewers.. Known failures include references to papers or court cases that do not exist — a form of hallucinationhallucination, n.: When an AI states something false or made up — a fake fact, quotation or source — fluently and confidently, as if it were true. — real sources that do not say what is claimed, and citations of low-quality pages. Some model APIs return structured citations pointing to exact spans of supplied documents, which reduces fabrication.
Every sentence in the summary carries a citation to the contract clause it came from, so legal can check it quickly.
Origin
OpenAI's WebGPT (December 2021) and DeepMind's GopherCite (March 2022) trained language models to back answers with references or quotations, and Hannah Rashkin and colleagues proposed an 'attributable to identified sources' standard for evaluation (2021). Citations became a standard feature of AI search products from 2023, the year fabricated case citations led to sanctions against lawyers in the US case Mata v. Avianca.
Usage note
A citation shows where a claim purports to come from, not that it is true; checking that the source says what the model claims remains essential.
Related
- retrieval-augmented generationretrieval-augmented generation, n.: Giving an AI an open-book exam: before answering, it looks up relevant passages in a set of documents and bases its reply on them, often citing where each point came from.
- groundinggrounding, n.: Connecting an AI's words to something real — trusted documents, a live search, images or the physical world — so that its answers rest on evidence rather than only on what it absorbed in training.
- hallucinationhallucination, n.: When an AI states something false or made up — a fake fact, quotation or source — fluently and confidently, as if it were true.
- factualityfactuality, n.: How far what an AI says is actually true about the world. A factual model gets facts, numbers, names and dates right and does not make things up.
- agentic RAGagentic RAG, n.: Retrieval in which the AI runs its own research: it decides what to search for, looks at the results, searches again if needed, and only then answers.
- deep researchdeep research, n.: An AI feature that spends minutes rather than seconds on a question — planning, running dozens of web searches, reading many sources — and then writes a long report with citations.
Sources
- Nakano, R. et al. (2021). WebGPT: Browser-assisted question-answering with human feedback. arXiv preprint.
- Menick, J. et al. (2022). Teaching language models to support answers with verified quotes. arXiv preprint.
- Gao, T., Yen, H., Yu, J. & Chen, D. (2023). Enabling Large Language Models to Generate Text with Citations. EMNLP 2023.
Mentioned in
deep researchdeep research, n.: An AI feature that spends minutes rather than seconds on a question — planning, running dozens of web searches, reading many sources — and then writes a long report with citations., retrieval-augmented generationretrieval-augmented generation, n.: Giving an AI an open-book exam: before answering, it looks up relevant passages in a set of documents and bases its reply on them, often citing where each point came from., agentic RAGagentic RAG, n.: Retrieval in which the AI runs its own research: it decides what to search for, looks at the results, searches again if needed, and only then answers., generative engine optimizationgenerative engine optimization, n.: Shaping web content so that AI chatbots and AI search features are more likely to find it, quote it and recommend it — the AI-era version of search engine optimization., groundinggrounding, n.: Connecting an AI's words to something real — trusted documents, a live search, images or the physical world — so that its answers rest on evidence rather than only on what it absorbed in training., WebGPTWebGPT, prop. n.: A 2021 OpenAI research system that answered questions by searching and browsing the web in a simple text browser, collecting sources and citing them — an early ancestor of today's research agents.