An Australian administrative court has found an increase of around 40% in cases involving generative AI between two financial years, with ChatGPT accounting for more than 75% of assisted filings. From 20 October 2026, applicants will have to declare their use of AI and verify all facts, references and links cited. One figure from the study is worth pausing on: non-native English speakers use it twice as much as native speakers.
We have covered lawyers sanctioned for citing invented case law. Here is the institutional response, and it is smarter than a ban.
What the measure requires
Two distinct obligations, and the second is the more important.
Declare usage. Flag that AI has been used in the preparation of a filing.
Verify every reference. This is the core of the scheme. The applicant must ensure that the facts, the cited decisions and the hyperlinks actually exist and say what they claim.
This second obligation is in fact a restatement of a duty that already existed: one has never been allowed to cite a non-existent decision. What the new rule does is make it explicit at the precise moment when the tool makes the error easy and invisible.
The figure that changes the picture
Here is the most interesting piece of information, and it complicates the usual narrative: people whose first language is not English use AI twice as much as native speakers.
That says something important about what is really going on. These people are not trying to cheat: they are compensating for a disadvantage. Drafting a legal filing in a language one does not master well, facing an administration that expects a precise register, is an obstacle that has nothing to do with the merits of the case.
This is exactly the mechanism we described regarding job applications and the shame of not knowing: the tool lifts a formal barrier that penalised some more than others.
Those who benefit most from the tool are also those least able to verify its outputs. Someone who does not master the language and the law well has no way of spotting that a cited decision does not exist. AI gives them access to correct form while exposing them to content they cannot control. A verification obligation, in this context, weighs most heavily on those least equipped to meet it.
Why this approach is better than a ban
Three reasons.
It is enforceable. A ban would require detection, and we have shown that detection does not work, with false positives hitting precisely non-native speakers.
It targets the right problem. The harm does not come from using the tool; it comes from invented references. A rule that targets verification treats the cause rather than the symptom.
It preserves access. Banning AI would amount to reinstating the language barrier. The chosen measure allows continued use, provided one answers for what one files.
This is the shift we advocated in our article on transparency: we stop tracking the machine and return to the responsibility of the person using it.
What is still missing
An obligation to verify does not help those who do not know how. If the court genuinely wants to address the problem it has identified, the measure should come with a means of verification: simple access to case law databases, or a tool to check that a reference exists.
Without that, the rule protects the institution against defective filings without helping those who submit them in good faith. That would be progress for the court, but not necessarily for justice.
What to take away
This decision is a good example of proportionate regulation: it observes a measured phenomenon, targets the real harm, and does not ban what it could not enforce.
It also reminds us of something we repeat in various forms: the problem with these tools is almost never that they produce falsehoods. It is that they produce falsehoods that look exactly like the truth, and that verification falls on the person least able to carry it out.