The Core Problem Grounding Solves
Large language models are extraordinarily good at producing fluent, confident text. They are less reliable at knowing whether the specific facts and claims in that text are accurate. Left unchecked, an AI will sometimes generate plausible-sounding information that has no basis in any real source — a phenomenon researchers call hallucination. Grounding is the set of techniques designed to prevent this by anchoring every AI output to verifiable input.
In a grounded AI system, the model is not generating from memory or statistical pattern-matching alone. It is working from a specific body of source material — retrieved documents, database records, or provided context — and its outputs are constrained to what can be supported by that material. Grounding does not eliminate all errors, but it dramatically reduces the rate of fabricated or unsupported claims.
Retrieval-Augmented Generation: Grounding in Practice
The most common grounding architecture in production AI systems is retrieval-augmented generation, or RAG. In a RAG system, when a user submits a query, the system first retrieves relevant documents from a verified database — in legal research, this means actual court opinions. The retrieved documents are then provided to the AI model as context, and the model generates its analysis based on that context rather than from its training data alone.
This architecture has two important properties. First, the AI is working from real documents that the system has verified exist. Second, the analysis the AI generates can be checked against those same documents — if the AI claims a court held something, you can look for that holding in the retrieved text. CaseMatch AI uses this architecture: when you run a search, the AI analyzes real opinions retrieved from our legal database, not text generated from a language model's memory.
The Grounding Score: Measuring Accuracy Automatically
Knowing that a system uses RAG architecture does not tell you how well any specific output is grounded. Even with real documents as context, an AI can over-interpret, extrapolate beyond what the text supports, or make claims that are adjacent to the source material rather than directly traceable to it. A grounding score measures this at the claim level.
CaseMatch AI's grounding check works by taking every claim the AI generates — winning factors, legal authority citations, case impact summaries — and verifying whether the key concepts in each claim appear in the source case text. A claim that the court suppressed evidence is grounded if the words "suppress" or "suppression" appear in the opinion's summary, holding, or full text. A claim about a specific constitutional basis is grounded if that constitutional provision is mentioned in the source. The resulting score tells you what percentage of the AI's claims passed this verification test.
What a Grounding Score Does and Does Not Tell You
A high grounding score — say, 90% or above — is a strong signal that the AI's analysis is closely tied to the actual case text. It does not guarantee that every claim is perfectly accurate in a legal sense, because text overlap is a proxy for grounding, not a perfect measure of it. A case might mention "probable cause" in a context that does not support the AI's specific claim about it.
A low grounding score, on the other hand, is a clear warning sign. If 40% or less of the AI's claims can be traced back to the source text, that analysis warrants careful scrutiny before you rely on it. The score gives you a fast, consistent signal about where to focus your verification effort.
Grounding as a Legal Ethics Requirement
Bar associations have increasingly emphasized that attorneys using AI tools retain full responsibility for the accuracy of their legal research. The duty of competence requires understanding the tools you use, and the duty of candor to the tribunal requires that citations you submit be real and accurate. Grounding technology does not transfer that responsibility — but it gives attorneys a concrete, systematic way to meet it. When a research tool shows you a grounding score alongside every AI analysis, verification becomes a built-in workflow step rather than an afterthought.