Why Legal AI in India Needs to Be Source-Grounded, Not Just Fluent

General-purpose language models got very good, very fast, at producing text that reads like a competent human wrote it. For most writing tasks, that's the whole job. For legal work, it's the easy part.

The hard part is whether a specific factual or legal claim in that fluent text is actually true - whether a cited case exists, whether it says what the draft claims it says, and whether it's still good law. Fluency doesn't tell you any of that.

Source-grounding means an AI system's output is tied to an identifiable, checkable source - a specific judgment, a specific statutory provision, a specific document you uploaded - rather than generated from the model's general training data with no attached reference.

In practice, this looks like: a citation that links to the actual judgment text, a retrieval date showing when that source was checked, and a verification status showing whether the citation has actually been cross-checked against the underlying record - as opposed to a citation that simply appears in the output because it was statistically plausible to generate.

A wrong citation in a casual research summary is an inconvenience. A wrong or misattributed citation in a filed pleading is a professional liability - and once it's in front of a judge or opposing counsel, the cost of catching it has gone up substantially.

This is also why grounding, on its own, isn't sufficient. Grounding reduces the chance of a fabricated or misattributed source; it doesn't replace the advocate's own review of whether that source actually supports the argument being made. Both matter, and neither substitutes for the other.

If you're evaluating any legal AI tool, a fair set of questions is: Where does a specific citation actually come from? Can you see the underlying source document? Is there a visible indication of whether that citation was checked, or does it just appear because the model generated it? A vendor that can't answer these clearly is asking you to trust fluency instead of verification.