Why 'summarise this case' is the worst prompt in legal research

By Michael Nadalin, Founder, Market Lead · 2 September 2026 · 6 min read

You paste a sixty page NSW Court of Appeal judgment into a chat window, type 'summarise this case', and get back a tidy set of paragraphs: the parties, the procedural history, what the primary judge found, what the Court held. All of it correct. None of it useful. You already knew the parties. What you needed to know was whether the Court treated the absence of a documented inspection system as decisive on breach, or as one factor among several, because that is the only question your matter turns on.

Hand the same judgment to the least impressive paralegal you have ever supervised and you get a worse written product and a better answer. Not because they are smarter than the model. Because before they wrote anything, they asked you which matter it was for.

That gap is the whole problem, and it is not a limitation of the technology. It is a briefing failure, and it sits on your side of the desk.

Three words, and every fact that mattered is gone

Read 'summarise this case' as a set of instructions and see what it actually asks for: produce the average summary of an unnamed document, for an unnamed reader, for an unnamed purpose. There is no issue named, no jurisdiction named, no side named, and no statement of what the answer will be used for.

A language model does not leave those gaps open. It fills them with the most statistically ordinary reading, which is a neutral headnote written for a first year law student. That is a reasonable guess at what an average person wants from an average judgment. It is not what you want, and nothing in your prompt told it otherwise.

The result feels competent, which is the dangerous part. A visibly wrong answer gets discarded. A generically right answer gets skimmed, half trusted, and quietly relied on.

What a bad paralegal does that your prompt did not

A weak junior still sits inside the matter. They know there is a hearing in the District Court next month. They know you act for the defendant. They have the pleadings on the desk beside them. Before writing a word they narrow the task, usually by asking you two irritating questions.

They also fail out loud. 'I could not find anything on that point, do you want me to keep looking' is a genuinely valuable sentence, and it is the sentence a model given three words will almost never produce. Its default is to generate something rather than to stop, because stopping was not one of the options you offered.

So the comparison is not model against junior. It is a fully briefed junior against an unbriefed one. You would not walk into a graduate's office, drop a judgment on the desk, say 'summarise this' and walk out. That is exactly what the three word prompt does.

The five things a legal prompt has to carry

None of this is prompt engineering in the sense of magic phrasing. It is the same information you would give a person, written down. Miss any one of the five and the model substitutes an average, and its averages are drawn from a global corpus that includes a great deal of American law.

Rewriting 'summarise this case' so a junior could not misread it

Here is the same request, briefed properly: 'The attached text is a NSW Court of Appeal judgment in a slip and fall claim against a shopping centre occupier. I act for the occupier. The point I care about is breach: did the Court treat the absence of a documented inspection system as decisive, or as one factor weighed with everything else? Set out the Court's reasoning on breach only, in no more than eight dot points, with the paragraph number against each point. Quote the sentence where the Court states the test it applied. If the judgment does not deal with that point, say so and stop. Use only the attached text.'

That is about a hundred words and takes well under a minute to type, most of it spent on the question rather than on instructions to the machine. What comes back is either eight paragraph referenced propositions you can verify against the judgment in the time it takes to scroll, or a sentence telling you the point is not there. Both outcomes are useful. The generic summary was neither.

Constraints, ranked by how much they actually protect you

Most of the anti hallucination advice circulating in practitioner groups is close to worthless, because it asks the model to want to be accurate rather than telling it what to do. The instructions that work are the ones that either close off a source of invention or give the model a permitted way to fail.

Where the effort actually goes

The trade is one minute of briefing for an answer you can verify in one minute, instead of ten minutes of reading followed by the uneasy decision to redo it yourself. On a research task of any size that maths is not close.

It also does not shift the professional obligation. Australian courts and regulators have issued guidance on the use of generative AI in litigation, and the practitioner who puts a citation before a court owns that citation regardless of what produced it. Paragraph numbers and quoted passages exist so that verification takes seconds rather than being skipped.

Tooling takes some of this off your hands. A purpose built Australian research tool such as Legal Brain fixes the source material to Australian judgments and legislation and shows you the passage it relied on, which covers two of the five items above before you type anything. It does not cover the other three, and nothing will. Naming the issue, naming your side and naming what you will do with the answer are lawyer's work.

The prompt is the brief. Write it the way you would brief a junior you do not entirely trust, and you will get the answer that junior would eventually have produced, in about a minute.

Frequently asked questions

What is a good prompt for summarising a case?

One that names the source document, the court and jurisdiction, the single issue you care about, whose side you are on, and the shape of the answer you want. Ask for paragraph numbers against every proposition so you can verify it quickly. Add an instruction to say so and stop if the judgment does not deal with your point.

Why does AI invent case citations?

Because a vague prompt leaves gaps and the model fills gaps with whatever is statistically plausible, and a citation that looks like a real Australian citation is highly plausible text. Two instructions do most of the work against it: restrict the answer to the attached or retrieved source, and give the model an express option to report that the answer is not there.

Can I upload a client judgment or brief into a general AI chatbot?

Check your firm's policy and your professional obligations on confidentiality and client consent before anything leaves your system, and check whether the court you are appearing in has issued a practice note on generative AI. Where the material is sensitive, de-identify it or use a tool with terms that suit legal work rather than a consumer chat product.

Do I still need to check every authority the AI gives me?

Yes. The solicitor or barrister signing the document is responsible for every citation and every proposition in it, and no output is a substitute for reading the passage. This is precisely why you should demand paragraph numbers and quoted sentences in the prompt: verification you can do in seconds is verification that actually gets done.