The Verification TaxSome AI answers are sourced. Some are just confident guesses.
Grounding is not a feature name, it is a difference in how the answer got made. Here is what that difference actually is.
The symptom, as lawyers describe it
What it looks like
An AI-drafted citation and a real one look identical on the page.
Same format, same confident tone, sometimes a real case name attached to the wrong proposition.
You cannot tell them apart by reading. You can only tell by checking.
One attorney's account of that gap, from a tool with a retrieval feature attached:
What lawyers complain about
Quotes are verbatim public posts by practitioners, reproduced without attribution to the individual.
The fix, in plain English
What changes
A language model writes by predicting the next likely word from patterns in its training.
Left alone it writes citations the same way, assembling a case name, volume and page because that pattern is probable, not because it looked anything up.
That is generation. It can be fluent and still be wrong.
Retrieval is different. The system first finds an actual document, matched against available legal sources.
It then writes from what it found, so the citation points at something that exists.
That is grounding: the answer is tied to a source you can open.
Neither approach is instant proof of correctness.
A retrieved source can still be misread, and a well-grounded system can still be cited for the wrong point.
Grounding narrows what you have to check. It does not remove the checking.
Keep going
Pick the next fix
the babysitting problem, what it costs you when this distinction gets ignored.
See Irys on your own matter.
Tell us about your practice and we will set up a walkthrough on the work you actually do.
No credit card. Transparent pricing. 7-day free trial.