The practice of connecting an AI model's responses to real, verifiable information — such as a live database, a set of documents, or a search index — rather than relying solely on what it learned during training. A "grounded" AI checks its answers against actual sources before responding, which makes its outputs more reliable and easier to fact-check.

For data reporters, grounding is the difference between an AI assistant that makes things up and one that cites its work. Imagine uploading a year's worth of city council meeting minutes and asking an AI to summarize every vote on housing policy: without grounding, the model might hallucinate votes that never happened. With grounding, it is constrained to the documents you provided, and every claim it makes can be traced back to a specific page. The same principle applies to building a chatbot that answers questions about your publication's archives, or a tool that cross-references campaign finance disclosures against a database of federal contractors.

Retrieval-augmented generation, commonly called RAG, is the most widely used technical approach to grounding. Instead of asking a model to answer from memory, a RAG system first searches a database for relevant documents, then feeds those documents into the model's context window along with the question. The technique is closely related to embeddings, which are used to find the most relevant chunks of text to retrieve. Grounding is also a key strategy for reducing hallucinations: when a model is anchored to a specific set of sources, it has less opportunity to invent facts.