> Fact checking an AI is still massively easier than finding and reading all the precedent yourself.

Actually fact-checking an AI requires finding and reading all the precedent yourself to veryify that the AI has both cited accurately and not missed contradictory precedent that is more relevant (whether newer, from a higher court, or more specifically on-point.)

If it has got an established track record, just as with a human assistant, you can make an informed decision about what corners you can afford to cut on that, but then you aren't really fact-checking it.

OTOH, an AI properly trained on one of the existing human-curated and annotated databases linking case law to issues and tracking which cases apply, overrule, or modify holdings from others might be extremely impressive—but those are likely to be expensive products tied to existing offerings from Westlaw, LexisNexis, etc.

What do you mean "finding"? The AI would just return links or raw text of the cases. Reading the findings would be the same as reading any precedence. But the AI could weight the results, and you'd only have to read the high scoring results. If the AI got it wrong, you'd just refine the search and the AI would be trained.

To the cost. If it removed the need for one legal assistant or associate then anything less than the cost of employing said person would be profit. So if it cost < 50k a year you'd be saving. (cost of employing is more than just salary)

You can't validate that it is making the right citations by only checking the cases it is citing, and the rankings it provides of those and other cases. You have to validate the non-existence of other, particularly contrary, cases it should be citing either additionally or instead, which it may or may not have ranked as relevant.

This is a very important point that highlights the absence of anything we could call "understanding".