Why your AI keeps citing overruled authority (and how to catch it)

By Michael Nadalin, Founder, Market Lead · 24 August 2026 · 8 min read

The dangerous output is not the invented case. Invented cases get caught, because the citation goes nowhere and the associate who checks it finds nothing. The output that reaches a filed submission is the one where the case is real, the citation is correct, the paragraph number lines up, and the proposition stopped being good law six years ago.

Every reported judgment in Australia sits online in exactly the form it was handed down. A first instance decision that was reversed by the Full Court reads the same today as it did the morning it was delivered. Nothing on the face of it says so. The correction lives somewhere else entirely, in the later judgment and in the citator, and most AI tools have never looked there.

This is not a bug that a smarter model fixes. It is a consequence of how these systems are built and what they were trained on. Once you see the mechanism, the checks that catch it become obvious, and they take about five minutes per authority.

The model is not searching, it is predicting

A general purpose language model does not look up your case. It generates the most probable sentence about that case, based on how the case was discussed across everything it was trained on. If a decision was cited approvingly in hundreds of judgments, textbooks, seminar papers and firm updates over two decades, that weight of approval is what the model has absorbed. The later judgment that departed from it is one document against many.

So the model is not lying to you and it is not confused. It is accurately reporting the centre of gravity of a body of text, and that centre of gravity lags the law by years. Confidence in the output tracks how often the proposition was written down, not whether it is currently correct.

The same mechanism explains a quieter version of the problem: a case that was never overruled but now construes a section that has since been amended. No citator flag exists for that. The reasoning is intact and useless.

Five ways an authority stops being good law

Overruling is only one of them, and it is the one people check for. The others are harder to see and just as fatal to an argument.

Worth noting where the real exposure sits. A wholly overruled leading case is usually well known in its field and someone in the room will catch it. Partial disapproval and appellate reversal are what get through, because the case still appears in current writing and still holds up on a quick read.

Why the judgment on your screen carries no warning

Australian judgments are published as delivered and are not amended when later courts depart from them. The record of subsequent treatment lives in a separate layer: LawCite on AustLII, CaseBase on LexisNexis, FirstPoint on Westlaw AU, the citator view on JADE. That layer is built by editors reading later judgments and classifying how each one treated the earlier decision.

Very few AI research tools index that layer. Retrieval systems are usually built over the text of judgments and legislation, because that text is available in bulk and the citator data generally is not. So a tool can retrieve the correct paragraph of a real judgment, quote it accurately, and still be quoting a proposition that a superior court has since rejected. Retrieval improves the accuracy of the quote. It does nothing about the currency of the law.

Ask any research tool you are evaluating one question: does it check subsequent judicial treatment, and where does that treatment data come from. A vague answer is an answer.

The five minute check that catches it

This is not new practice. It is noting up, which every graduate learns and which quietly stopped happening in some firms once answers started arriving pre written. The only change is that you now run it against output you did not produce and cannot interrogate.

Run it in order. Steps one and two kill most bad citations before you spend money on a database search.

Prompting habits that lower the failure rate

None of this removes the need to verify. It does reduce how much bad material you have to wade through, and it makes the weak answers easier to spot, because a model given room to say it is unsure will sometimes take it.

The single biggest improvement is refusing to ask leading questions. Ask a research tool to support a position and it will support the position. Ask it what the strongest argument against you is and you get a far more useful map of the field, including the authorities your opponent is going to run.

Make it a system, not a matter of individual diligence

Personal discipline decays under deadline, which is exactly when these tools get used most heavily. The firms that avoid trouble here have written the check into the workflow rather than relying on everyone remembering it at nine on a Sunday night. That means a short AI use policy, a verification step in the document review process, and clear supervision responsibility for junior practitioners who may not yet have the instinct that a fluent answer can be entirely superseded.

The professional exposure is not exotic. Your duties to the court and your client under the Australian Solicitors' Conduct Rules do not soften because a machine drafted the passage, and several Australian courts have now issued practice notes and guidelines on the use of generative AI in litigation, some requiring disclosure or verification of AI assisted material. Read the ones that apply to the courts you appear in, because they differ, and check them again periodically since they are being updated.

Tooling helps if you choose it on the right criterion. Ask whether the system is grounded in current Australian primary materials, whether it surfaces subsequent judicial treatment rather than just the text of the judgment, and whether it links you back to the source so you can read the paragraph yourself. That is the standard we built Legal Brain to meet, and it is the standard worth applying to any tool you assess, including ours. A research tool that cannot show you its sources is asking you to stake your practising certificate on its confidence.

Frequently asked questions

Can ChatGPT tell me whether a case has been overruled in Australia?

Not reliably. A general purpose model has no live access to citator data and answers from patterns in its training text, where decades of approving citations outweigh a single later judgment that departed from the case. It will often assert a case is good law with complete confidence. Check subsequent treatment in LawCite, CaseBase, FirstPoint or JADE before relying on it.

What does noting up a case mean and is it still necessary if I use AI research?

Noting up means checking how later courts have treated a decision: applied, distinguished, doubted, not followed, overruled or reversed. It is more necessary with AI, not less, because AI output gives you a real citation and a plausible proposition without any indication of currency. Treat it as the mandatory last step on every authority the tool hands you.

Why do AI tools cite first instance decisions that were reversed on appeal?

Because the first instance judgment remains published in full and reads as authoritative, with nothing on its face recording the appeal. The reversal is a separate document. Unless the tool checks the appellate history, it has no way to know, so always trace a case upward before you cite it.

Do Australian courts allow solicitors to use generative AI for legal research?

Generally yes, subject to conditions, but the rules vary by court and jurisdiction and several have issued practice notes or guidelines covering generative AI, including requirements around verification and disclosure in some circumstances. Your duties to the court are unchanged regardless. Check the current practice note for each court you appear in, and check it again periodically because these are being revised.