What Is a Grounding Score?
A grounding score is a quantified measure of how closely an AI's generated analysis tracks the source material it was given to work from. In legal research, this means: for a given AI-analyzed case, what percentage of the AI's claims about that case — its winning factors, legal authority, and case impact — can be traced back to actual language in the court opinion?
A score of 90% means nine out of ten claims the AI made are directly supported by text in the opinion. A score of 50% means roughly half the claims have clear textual grounding and half are interpretations, inferences, or extrapolations that go beyond what the source document explicitly says. The score is a fast, automated proxy for answer quality — a signal that would otherwise require you to read the entire opinion to evaluate.
Why Grounding Scores Matter More Than Overall AI Accuracy
You may have seen AI tools advertise overall accuracy rates — claims like "95% accurate on legal tasks." These aggregate metrics, while useful for comparing tools, do not tell you anything about the specific claim in front of you. A tool that is 95% accurate on average will produce inaccurate output on approximately 1 in 20 claims, and you have no way of knowing which 1 in 20 it is without checking.
A grounding score addresses this at the individual result level. Instead of telling you how the tool performs on average, it tells you how this particular analysis performed against this particular case text. That per-result, per-analysis signal is far more actionable for an attorney making a reliance decision about a specific piece of research.
How to Read a Grounding Score in Practice
When CaseMatch AI shows you a "Verified · 88%" badge on a case result, here is how to read it in practice. The AI analyzed this specific case and made a set of claims about why it was won — winning factors, legal authority cited, and case impact. Of those claims, 88% were verified to contain language traceable to the case's actual text in our database. The remaining 12% are claims where the AI's interpretation extended beyond what could be directly matched to source language.
For a case with an 88% grounding score, you can read the AI's analysis with reasonable confidence and spend your verification time confirming nuance rather than checking basic accuracy. For a case with a 45% grounding score, you should read the actual opinion before building any argument on the AI's characterization of it.
Grounding Scores and Legal Ethics
Bar associations have consistently held that attorneys retain full professional responsibility for AI-assisted work product. The duty of competence and the duty of candor do not come with an AI exception. Grounding scores are relevant to both of these obligations.
On competence: understanding what a grounding score means and using it appropriately is part of using an AI tool competently. An attorney who ignores a low grounding score on a case they plan to rely on has arguably not exercised the verification diligence that competent use of AI requires. On candor: citing a case based on AI analysis with a low grounding score, without independent verification of the holding, creates a risk of inadvertently misrepresenting the law — precisely the type of conduct bar authorities have flagged in their AI guidance.
Integrating Grounding Scores Into Your Research Workflow
The most practical way to use grounding scores is as a triage tool that tells you where to read and where to skim. Start by identifying the cases on your shortlist with the highest grounding scores — these are the ones where the AI's analysis is most reliable as a starting point. For these cases, read the holding to confirm the AI's characterization is accurate and look for any language the AI may have underweighted.
For cases with lower grounding scores that you still want to use, read the full opinion section rather than relying on the AI summary. The lower score is a signal that the AI's interpretation diverges from the text, which may mean the case is less strong for your position than it appeared — or it may mean the AI missed something important that actually helps you. Either way, the opinion itself will tell you.
Grounding scores do not replace legal judgment. They focus it. That is exactly what a good research tool should do.