State, Commonwealth or California: where legal AI gets it wrong

By Michael Nadalin, Founder, Market Lead · 7 September 2026 · 7 min read

Ask a general purpose chatbot for the limitation period on a personal injury claim and you will get a confident number in about four seconds. Ask it which Act, which state, and whether the clock runs from the date of injury or from discoverability, and the answer starts moving. That movement is the whole problem. The first answer was not knowledge. It was an average.

Australia has nine bodies of law that can govern one client's problem: the Commonwealth, six states and two territories. Most general AI tools treat that as a single corpus called "Australian law", then fill the gaps with the largest body of legal text ever published on the open web, which is American. The result reads fluently, uses the right register, and cites nothing you can open.

This is not an argument that AI is useless for legal research. It is a description of two specific failure modes, where each one comes from, and what a two line change to your prompt and your checking habit will catch.

One question, nine possible answers

The federal structure is the first thing a model flattens. Some areas sit squarely with the Commonwealth, some sit squarely with the states, and a large middle band is concurrent, with section 109 of the Constitution deciding what happens when the two collide. A tool that does not hold that structure will happily blend a Commonwealth Act with state case law and hand you a paragraph that is internally consistent and legally nowhere.

It is also not a clean binary. The Australian Consumer Law is Schedule 2 to the Competition and Consumer Act 2010 (Cth), but it is applied as a law of each state and territory and enforced by state fair trading bodies as well as the ACCC. Defamation runs the other way: the model provisions are drafted to be uniform, then enacted separately in each jurisdiction, which is why the 2021 serious harm amendments commenced on different dates and Western Australia and the Northern Territory came later than the eastern states. "Uniform" and "identical" are not the same word, and a model trained on general text does not feel that difference.

So a question about a client in Hobart has one right answer and eight plausible neighbours. A fluent model will pick a neighbour without telling you it made a choice.

Where the boundary actually sits

The areas below are the ones where a misplaced boundary does real damage, because the wrong answer still sounds correct to a reader who practises in a different state. Note how often the honest answer is "it depends on the entity, the employer type or the date", which is exactly the kind of qualification a model strips out when it is optimising for a clean reply.

Criminal work is the clearest example. Queensland, Western Australia, Tasmania and the Northern Territory are code jurisdictions. New South Wales, Victoria and South Australia are common law jurisdictions with statutory overlay. Commonwealth offences sit in the Criminal Code Act 1995 (Cth) on top of all of it. An answer that discusses the elements of an offence without naming the source is not an answer.

The American accent in the answer

The second failure mode is simpler and more common. The published volume of American legal material, case reports, law school outlines, firm commentary, forum answers, dwarfs anything Australian on the open web. Australian primary law is public but comparatively thin, less repeated, and often sitting in structures a general crawler handles poorly. When a model is asked a legal question with no jurisdiction attached, the statistical centre of "law" is Washington, not Canberra.

It rarely announces itself with a US case name. It leaks through vocabulary and procedure first, which is why it survives a quick read. If you are skimming for the answer rather than the framing, you will not notice that the tool has quietly changed countries.

How a wrong jurisdiction travels through a file

Nobody sends a chatbot transcript to a client. The risk is not the bad answer, it is the laundering. Once a proposition is rewritten in your own words in a file note, it loses its source and gains your firm's authority. Two steps later it is in advice, and there is no longer anything on the file that says where it came from.

Several Australian courts have now issued practice notes and guidelines on the use of generative AI in proceedings, and the direction is consistent: the responsibility for what is filed stays with the practitioner who signs it. A tool cannot hold a practising certificate.

A checking routine that fits inside a minute

You do not need a policy document. You need a habit that makes the jurisdiction explicit before the answer is generated, and one that makes the primary source cheap to open afterwards. Most of the value is in the first step: models that are given a jurisdiction, a court and a date behave very differently from models that are given a bare question.

The version in force is the step people skip. A consolidated Act as at today is the wrong text if the facts arose two years ago and the section has since been amended. General tools almost never volunteer which version they are describing, and often cannot tell you.

What to demand from any tool before you rely on it

Judge legal AI on four things, none of which are about how good the writing is. Does it state the jurisdiction on the face of every answer, without being asked. Does it give a pinpoint citation you can open in one click, to the actual text rather than a paraphrase. Does it tell you which version of the provision it is reading and as at what date. And does it say it does not have something, rather than producing a plausible substitute.

The fourth one is the hardest and the most important. A tool built to always produce an answer will always produce an answer, and the failure will look identical to a success. A tool built on an Australian corpus with the jurisdiction pinned can decline. Legal Brain was built for that constraint: Australian primary sources, the jurisdiction stated up front, and a link to the text so verification takes seconds instead of an afternoon.

Whatever you use, the allocation of risk has not changed. The tool does not hold the file, does not sign the advice and does not answer to the client or to the court. You do. Make the jurisdiction explicit, open the primary source, and the American default never reaches your file note.

Frequently asked questions

Does ChatGPT know Australian law?

It has seen a lot of Australian legal text, but it holds no reliable model of which jurisdiction applies to your question. Asked a legal question with no state or court named, it tends to answer from the largest body of legal material it was trained on, which is American. It can be useful for framing an issue or drafting, but any proposition of law needs to be checked against the primary source before it goes on a file.

Why do AI tools cite American cases in Australian matters?

Published American legal material vastly outweighs Australian material on the open web, so it dominates the training data. When nothing in the prompt anchors the answer to Australia, the model produces the statistically ordinary legal answer, which is a US one. Naming the state, the court and the relevant date in the prompt removes most of that drift.

Is Australian law state or federal?

Both, and often at the same time. Some areas sit with the Commonwealth, some sit with the states and territories, and a large middle band is concurrent, with section 109 of the Constitution resolving inconsistency. Some Commonwealth statutes, such as the Australian Consumer Law, are also applied as laws of each state, so the question is never simply one or the other.

How do I stop AI giving me the wrong state's law?

State the jurisdiction, the court and the date the facts arose in the prompt itself, and require a pinpoint citation to the Act and section rather than a summary. Then open the primary text on the relevant legislation register or AustLII and read the provision. If the tool will not give you a citation you can open, treat the answer as a starting point and nothing more.