AI for legal research: what these tools do and do not do

Three different things wearing one name

1. A general model with no law behind it. ChatGPT, Claude or Gemini out of the box. These know law as a pattern, not as a source. Ask for authority and you are rolling dice.
2. A legal-specific wrapper. A general model with legal prompting, and sometimes a document set behind it. Better framing, much the same underlying risk on citation.
3. A retrieval tool over a real database. These search an actual corpus of case law and legislation and cite what they found. A different category of thing entirely.

Anyone comparing all three in a single ranked list is selling something.

What they do well

1. Orientation. An unfamiliar area, and thirty minutes to sound competent in it. This is genuinely transformative and it is the use nobody markets.
2. Narrowing. Turning a vague worry into three searchable questions.
3. Summarising the long thing you already found. Reliable, because the source is sitting in front of it.
4. Finding the thing you half remember. Very good, and a real relief.

What "grounded" actually means

This is the single distinction worth understanding, and it is not complicated.

An ungrounded model answers from what it absorbed in training. It is recalling a pattern of what legal authority looks like, which is why the fabrications are so convincing - the shape is perfect and only the content is invented.

A grounded tool searches a real corpus first, then writes its answer from what it retrieved, with links. It can still summarise a case badly. It cannot invent one, because it is working from documents actually in front of it.

So the question to ask a vendor is never "does it hallucinate". Everyone says no. The question is "what did it search, and can I click through to it". A tool answering that clearly is in a different category from one answering it smoothly.

Where they break

1. Citation. A grounded tool cites what it retrieved. An ungrounded one cites what sounds right. On the page, the two look identical, which is precisely the problem.
2. Currency. Is this still good law? A retrieval tool can tell you where the database flags it. A chatbot will not, and will not tell you it cannot.
3. Weight. Knowing a first-instance decision from a Court of Appeal judgment, and knowing which one the other side will quote at you.
4. The argument. Research tools find the point. They do not find the way to run it.

The cost question

This is where most of these decisions are actually settled, and the listicles are quiet about it.

The grounded tools sit on top of the big legal databases, and they are priced accordingly - per seat, annually, at a level making sense for a firm doing research every week and no sense at all for one doing it twice a quarter. A general model with a decent subscription costs roughly what one person spends on coffee.

So the sensible question is not which tool is best. It is how much genuine research your work involves. If the answer is "not much, but when it happens it matters", you are usually better off with a cheap general tool for orientation and paying for the authority properly when a real question lands.

Firms routinely get this backwards. They buy the expensive tool for the research they rarely do, and use the free one for the work they do daily.

The checking routine worth having

Whatever you use, the habit matters more than the tool. Four steps, and they take less time than they sound.

1. Open every authority cited. Every one, every time. If a citation will not open, it does not exist - and that is the most common way an invented case reaches a document.
2. Check the case says what the summary claims. Roughly half the errors are real cases described wrongly, which is harder to spot than a fabrication.
3. Check it is still good law, in the database rather than in the chat window.
4. Ask what the tool did not tell you. It answers the question you asked, and is entirely untroubled by the better question you did not.

What to check before you buy

1. What is it searching? If nobody at the vendor can name the corpus, there is not one.
2. Does it link to the primary source, every time, in one click? If not, walk away.
3. What happens to what you type? Ask specifically whether prompts are used for training, and get it in writing.
4. Can you get out? A twelve-month lock-in on a market moving this fast is a bad trade.
5. What does a general model already do for you at a fraction of the price? Sometimes the honest answer is most of it.

Frequently asked questions