Grounding

A chat program like ChatGPT can form sentences that sound convincing but are wrong. While learning from vast amounts of text, it picked up patterns and often merely reproduces what seems plausible. Grounding is the attempt to prevent this. Before answering, the program is presented with matching documents: a webpage, a manual, a spreadsheet. It is then supposed to derive its answer from this material rather than invent it from its own memory. The English term literally means something like giving the answer a firm footing: it gets solid ground beneath it.

Why made-up answers become so costly

Language models are probability machines. They calculate which word fits best next. Whether a statement is true is not something they check. When a model freely invents a year, a law, or a price, this is called a hallucination. This is exactly where grounding comes in.

For companies, this is the decisive point. A bank must not tell its customers made-up fees. A clinic must not issue an invented dosage. Without a tie to sources, a language model is simply unusable for such tasks. That is why grounding shows up in almost every major provider’s product announcement.

A second reason is currency of information. A model only knows what was on the internet up until the end of its training. This point in time is called the knowledge cutoff. Everything after that is missing. Grounding supplies the missing new information afterward, without the model needing to be retrained.

From search hit to substantiated answer

The usual process consists of three steps. First, a search query is built from the user’s question. Then the system searches a data source and retrieves the most fitting pieces of text. Only after that is the language model presented with the question and the text pieces together, with the instruction to answer only from them. This procedure is called Retrieval Augmented Generation, or RAG for short.

The search itself usually does not rely on exact keywords. Instead, texts are converted into long sequences of numbers that capture their meaning. Two texts with similar meaning produce similar number sequences. This way, the system finds the right passage even when the document says “notice period” and the user wrote “quit”.

A common misconception: grounding does not make a model honest, only better informed. If the search delivers the wrong passage, the model dutifully builds a wrong answer from it. That is why grounding quality is measured in two parts. Does the search find the right thing? And does the answer afterward actually stick to what was found? In addition, many systems attach source citations so people can spot-check them.

Footnotes in the chat window

Grounding is most visible in search assistants. When Google shows an AI summary above the search results, or Perplexity places small numbers behind sentences, these are grounding citations. Web search in ChatGPT and Copilot works the same way: search first, then formulate.

In everyday corporate use, grounding is mainly embedded in internal assistants. A support chatbot accesses the product database, a law firm’s assistant accesses collections of rulings. Such systems are often chosen based on how reliably they connect to sources, not on how well they write poems.

In news coverage, you’ll usually encounter the term in connection with reliability. When a provider writes that a model is now “grounded in real-time data,” it means: it queries a current data source before answering. That is a considerably more modest claim than it sounds. It guarantees no truth, only a traceable origin of the statement.