What Khan v Aldi tells employment lawyers about AI legal research

Two things came up repeatedly when I was talking to colleagues at the FNQLA Core CPD Seminar in Cairns in April this year.

The first was self-represented parties on the opposing side, running matters on AI generated material that goes nowhere, at the cost of everyone's time and their client's money. The second was clients turning up with the answer already, supposedly, formed, usually the wrong one, ready to tell their lawyer how to do their job.

Sadnan Khan v Aldi Pty Ltd [2026] FWC 3144 speaks to both. It is also more instructive than the coverage suggests for anyone working out what AI for unfair dismissal cases should actually look like. The interesting failure discussed in this decision was not a fabricated citation. It was relevance.

What the Fair Work Commission actually decided

Mr Khan commenced employment with Aldi in September 2025. He was notified of his dismissal at approximately 4:00pm on 26 March 2026 and filed an unfair dismissal application at 10:02pm the same day. He was three days short of six months.

Section 383(a)(i) of the Fair Work Act 2009 (Cth) (Act) fixes the minimum employment period as ending at the earlier of the time the person is given notice of the dismissal or immediately before the dismissal. On the uncontroversial evidence, Mr Khan's period ended on 26 or 27 March 2026. He was never eligible to bring the claim.

From the moment that problem was raised, his AI generated materials focused on a different and irrelevant date, being the date the dismissal took effect. Deputy President Easton wrote to him twice on 31 July 2026, in plain English, setting out the operation of s.383 of the Act and the Full Bench authority in Corner v SkyCity Adelaide Pty Ltd [2011] FWAFB 955. He was invited to discontinue and warned about costs. Aldi indicated it would apply for costs. Mr Khan filed two further AI generated submissions quoting the same wrong date, and conceded the point only at the hearing on 5 August 2026, where he discontinued.

Deputy President Easton ordered him to pay $1230 towards Aldi's costs under s.400A of the Act, describing it as a very rare step.

Two details about that figure matter. It is capped by Schedule 3.1 of the Fair Work Regulations 2009 (Cth), at two hours for the digital hearing book and four hours of preparation and attendance, and Deputy President Easton said he was quite certain Aldi incurred significantly more. And the order was expressly made for deterrence, to deter Mr Khan "and hopefully other applicants in the same position."

What the decision does not say

This is not a decision that says using AI is unreasonable, and practitioners citing it should be careful on that point.

Aldi sensibly confined its claim to costs incurred after 31 July (the date on which he was notified of the defects in his submissions), and Deputy President Easton therefore did not need to decide whether relying on AI generated submissions before that date was unreasonable. He was prepared to accept that until then Mr Khan did not know any better.

He went further than merely declining to decide it. He said, “Using AI to assist in drafting correspondence and even submissions is not, of itself, unreasonable. Artificial intelligence, in 2026 at least, is a drafting tool that can assist litigants when used properly.”

What he criticised was a change of role. It was "quite possible that Mr Khan adopted AI as his quasi-legal advisor and then blindly submitted AI generated outputs as his own material." The unreasonableness lay in continuing after the Commission had told him, twice and unambiguously, why the claim could not succeed. Parties, Deputy President Easton observed, must read and review all correspondence that comes from the Commission.

The distinction between a drafting tool and an advisor is central, and it was drawn by the bench rather than by a vendor.

The failure was not hallucination, it was relevance

The submission filed on 2 August was headed "Why the authority relied upon by the Deputy President does not properly apply to the present matter.” Deputy President Easton's assessment of it warrant particular attention:

“The submission was not complete AI slop, as that term has come to be used, insofar as some of the legal arguments presented were more or less correct, though irrelevant, and the High Court cases cited were real cases rather than hallucinations.”

Nothing was invented. The authorities existed. The reasoning was competent. It was a distinguishing argument, which is a recognisable piece of legal work, aimed squarely at the wrong question.

In recent years we’ve taught each other to audit for fabricated citations, because fabrication is visible. It fails a check we know how to run. Correct law aimed at the wrong issue fails no check at all. It doesn’t look like an error. It looks like a case.

And it is not only self-represented litigants who are exposed to it. In this decision, Deputy President Easton notes that Aldi’s first written submissions addressed the same wrong question. Aldi was represented by an employment specialist firm. The wrong question survived contact with competent lawyers until the member deciding the matter redirected everyone to s.383 of the Act. Once a matter has been framed around the wrong issue, the framing is contagious, and the volume of correct material generated within that frame makes it harder, rather than easier, to see.

What this means for AI tribunal case preparation

The practical exposure for employment practitioners is not that a chatbot will invent a case and embarrass them. Most people now check for that. It is the material arriving from the other side, and from clients, that is internally coherent, properly sourced, and directed at something other than the question in issue.

That material is expensive in a specific way. It generates work that looks like legal work. It draws submissions in reply. In Khan, it drew Aldi onto the same wrong date. Deputy President Easton's closing observation is the one worth taking note of.  He said, “The losses for both parties were utterly preventable. If Mr Khan had properly read for himself the first email he received from the Commission two days after his application was lodged, he would have immediately discontinued.” At that point he would have been better off than when he started, because the application fee was refundable and no costs exposure had crystallised.

The skill this calls for is not caution about AI. It is fluency: knowing what these systems reliably do, where they fail quietly, and which failures your existing checks will not catch.

Where AI legal research platforms now have to earn their place

Surfacing correct law stopped being the measure of success a while ago. Legal AI tools are being now designed to deliver reliable authorities.

The work that remains is the work Khan exposes: surfacing all relevant law, then helping the lawyer see what actually bears on the matter in front of them. i.e. read this authority rather than that one because the facts sit closer to your file. Skip the five decisions on the same point of law that will not move your matter an inch. Take note that the question you framed, is different to the test the legislation applies.

The tool is not making that judgment. A lawyer makes it, and that does not change. What a tool can do is shorten the run up to it, and it can only do that if it knows something about the matter. A system answering in the abstract will give you correct law about employment termination. A system carrying the matter into the answer is the one positioned to tell you that the date you are arguing about is not the date the section turns on.

That is the difference between AI legal research platforms that are pulling ahead in employment law and those falling well short of what the market now expects.

Professional obligations and compliance

For practitioners, the professional dimension of this is straightforward and worth stating plainly.

Where a lawyer has a practising colleague, an internal review step, or a platform's verification layer between the output and the advice, the obligation to the client and to the court remains entirely with the lawyer running the matter. Deputy President Easton's reasoning is consistent with that: Mr Khan was not excused by his reliance on AI, nor by the fact that he "is not a legal expert," because it remained his responsibility to assess for himself whether he was eligible and to ensure that what he filed addressed the concerns raised.

The practitioner’s final judgment call, grounded in knowledge of the matter, the client, and the full correspondence, is what catches the relevance failure that no citation audit will find. That is the case for building a second step into the process, and it is a case about accuracy rather than efficiency.

It is also why the Commission's own free material is part of this picture. The benchbook includes a page titled "How do you calculate the minimum period of employment?", which addresses the date the employee is notified of the dismissal. The correct answer was public, findable and free. It was routed around rather than unavailable.

Frequently asked questions

Did the Fair Work Commission decide that using AI is unreasonable?
No. Deputy President Easton said that using AI to assist in drafting correspondence and even submissions is not, of itself, unreasonable, and described AI in 2026 as a drafting tool that can assist litigants when used properly. The costs order was made because Mr Khan continued his claim after being clearly told, twice, that on his own version of events he could not succeed.

Can AI be used to prepare unfair dismissal submissions?
On the reasoning in Khan, yes, as a drafting tool. The difficulty arises when AI is treated as a source of advice and its outputs are filed without the party assessing them. Deputy President Easton described that as adopting AI as a quasi-legal advisor and blindly submitting the outputs as one's own material.

Why was the costs order only $1230?
The amount reflects the maximum rates claimable under Schedule 3.1 of the Fair Work Regulations 2009 (Cth). Deputy President Easton said he was quite certain Aldi incurred significantly more than $1230 in the relevant period.

Are costs orders common in unfair dismissal matters?
No. Parties are ordinarily protected from paying an opponent's costs. That protection can be lost under s.400A of the Act where an unreasonable act or omission in the conduct or continuation of the matter causes the other party to incur costs. Deputy President Easton described the order as a very rare step.

What is the risk for represented parties?
Khan is a reminder that the wrong question is not only a self-represented problem. Aldi's first written submissions addressed the same irrelevant date until the Commission redirected the parties to the correct test.

Read the full decision here.