In two single-judge costs rulings, the Federal Court of Australia ordered costs against plaintiffs who disclosed the scale of an AI-assisted document exercise late, and against a litigant who the court found had instituted proceedings without reasonable cause

What did the court decide in Jahani v Qiu?

Cheeseman J ordered the plaintiffs, liquidators of companies in the Ralan Group, to pay the defendants' costs thrown away when a final hearing listed for 28 September 2026 was vacated. This is a single-judge costs ruling at first instance, binding on the parties only. The substantive case, a corporate insolvency claim over purchaser deposits that the plaintiffs allege exceeded $288 million unpaid, continues and is relisted for 1 February 2027.

The vacation came out of the plaintiffs' application under s 50 of the Evidence Act 1995 (Cth), which concerned the mode of proof of approximately 7,627 underlying documents. On 11 August 2026 the plaintiffs served summaries of more than 7,500 documents, which the judgment says were prepared using generative AI and a process of verification by the plaintiffs' solicitors. The defendants found errors by spot-checking. The plaintiffs accepted that errors existed but disputed their materiality. Updated summaries followed on 14 August, 3 September and 21 September 2026.

At the amendment hearing on 2 April 2026, the plaintiffs had foreshadowed summaries of some 500 contracts. The judge found that the exercise grew to more than 7,500 documents, that the increase was not disclosed before timetable orders were made by consent on 11 August 2026, and that a progress email of 30 July 2026 did not reveal it. Under s 37M of the Federal Court of Australia Act, Cheeseman J said: "If the scale of a proposed evidentiary exercise materially increases, the party undertaking it should disclose that development promptly so that the timetable may be made, or reconsidered, on an informed basis."

Why did GPN-AI not supply a separate costs basis?

The defendants argued that a costs order was consistent with the Court's Use of Generative Artificial Intelligence Practice Note (GPN-AI), which, as the judgment records their submission, says at [4.4] and [5.1] that presenting inaccurate AI-generated material to the Court is unacceptable and may attract adverse costs consequences. Cheeseman J did not accept that this added anything in this case: "I do not accept, however, that GPN-AI provides a separate basis for an adverse costs order in this case."

The reason given was the plaintiffs' process. They used AI to extract information and then, in the judge's words, undertook a substantial process of manual review and correction. "The vice identified by the defendants is not the responsible use of AI for a mechanical task involving voluminous material. It is the late disclosure of the scale of the exercise and the resulting effect on the defendants' ability to prepare."

Cheeseman J cited the observations of Lee J in Rogers v McDonald's Australia Ltd (AI-use) [2026] FCA 1264 at [11]-[13] for the proposition, in her words, that "the use of AI for collation and synthesis may further the overarching purpose where adequate verification, human supervision, and traceability to source material are maintained." We report that citation as Cheeseman J made it. We did not open Rogers.

The judge also said the same reasoning applies to the defendants. If the plaintiffs could properly use AI, subject to verification, there was "no principled reason" why the defendants should not use equivalent technology to speed their own review. The defendants told the Court they would do that. None of this, the judgment says, diminished the defendants' entitlement to a fair opportunity to check the summaries. The ruling does not amend GPN-AI, and it says nothing about cases where AI-generated material reaches the Court unverified.

What did Wheelahan J find in Raghib v Stantec?

Raghib is a costs ruling in the Fair Work Division, again by a single judge at first instance and binding on the parties. On 15 September 2026 Wheelahan J had dismissed the self-represented applicant's judicial review of two Fair Work Commission decisions. Stantec, the first respondent, then sought costs. Under s 570 of the Fair Work Act, costs can be ordered in such matters only in limited circumstances, and the judge found two of them engaged: s 570(2)(a), proceedings instituted without reasonable cause, and s 570(2)(b), an unreasonable act or omission causing costs.

The AI observation sits in the judge's account of how the case was presented. The applicant's claims, the reasons say, "contained apparent hallucinations, including occasions where the applicant made reference to written submissions that did not exist". The judgment records the applicant's own narrower account: "The applicant stated at the hearing of the matter that he had used artificial intelligence, but only to check the grammar and spelling of documents he otherwise prepared." The judge's view followed: "In my view, there is a likelihood that some of the anomalies in the applicant's filed documents were the product of the indiscriminate use of artificial intelligence."

That is a likelihood, stated by the judge. It is not a finding that AI caused any particular error, and the reasons do not treat AI use as a costs ground in itself. The conclusion at [23] is that the proceeding "was always doomed to fail because it lacked merit" and was instituted without reasonable cause. The judge also treated a subpoena pursued without legitimate forensic purpose as conduct that the Australian Government Solicitor's referral did not make reasonable.

Stantec asked for indemnity costs of $84,345 or party and party costs of $56,230. Wheelahan J refused indemnity costs, noting the applicant had generally complied with the timetable, and fixed party and party costs at $45,000 excluding GST. Part of the reduction came from errors in Stantec's own costs affidavit, which the applicant had identified. A second affidavit, served on the morning of the hearing to fix them, was not allowed to be read.

How do the two rulings fit together?

What follows is our reading, not something either judge said about the other case. Both judgments are first-instance costs decisions and neither cites the other.

In Jahani the party using AI had a documented verification process and was still ordered to pay costs, because of what it failed to disclose about scale. In Raghib the judge thought indiscriminate AI use likely, and the costs order rested on statutory findings about merit and conduct. Neither judgment states AI use as the basis for its costs order.

There is a smaller point in Raghib that document-heavy teams may notice. The costs evidence that drew criticism was the respondent's own affidavit, and the judgment does not say how that affidavit was prepared. Errors cost Stantec part of its recovery regardless of how they arose.

What would litigation and e-discovery teams take from this?

These points are inferences from two single-judge rulings, and they bind nobody beyond the parties. On Cheeseman J's reasoning, a team running AI extraction over a large document set would want a record of its human review and correction, because that record was what separated responsible use from the conduct the defendants complained of.

Scale is the second point. The Jahani plaintiffs were best placed, the judge said, to see that the exercise had grown, and their failure to correct the earlier estimate before seeking consent timetable orders was inconsistent with s 37M. A team whose AI-assisted review grows by an order of magnitude would, on that reasoning, flag the change to the other side and the Court before timetables are fixed around the old number.

The third concerns the receiving side. Cheeseman J said the overarching purpose is better served by both parties making appropriate use of available technology than by treating such tools as grounds for criticising the user or excusing delay by the non-user. A party served with AI-prepared summaries would, on that view, struggle to rely on the volume alone without using comparable tools to check them, while still keeping its entitlement to a fair opportunity to do so.

What we did not verify

What we opened: the full published reasons and orders in Jahani v Qiu (costs) [2026] FCA 1419 (21 paragraphs, certified 25 September 2026) and Raghib v Stantec Australia (Costs) [2026] FCA 1415 (36 paragraphs, certified 24 September 2026), both as published on the Federal Court of Australia judgments site. Every fact, figure and quotation on this page comes from those two texts.

What we did not open: Rogers v McDonald's Australia Ltd (AI-use) [2026] FCA 1264, which we report only as cited by Cheeseman J; the text of GPN-AI itself, including [4.4] and [5.1], which we report only as the defendants' submission recorded in Jahani; the earlier Jahani amendment ruling [2026] FCA 398; the principal Raghib judgment [2026] FCA 1354 and the subpoena ruling [2026] FCA 1322; and any notice of appeal in either matter.

What we refuse to claim: that responsible AI use gives immunity from costs; that the Jahani summaries were error-free (the plaintiffs accepted errors existed); that either judgment creates an automatic AI costs penalty; that GPN-AI has been amended or reinterpreted with general effect; that either ruling is appellate or binding precedent beyond the parties; or that AI use caused any specific defect in Raghib, where the judge expressed a likelihood. We make no claim that these are the first rulings of their kind.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• Australian Federal Court Practice Note on Generative AI Use
• Australia's Fair Work Commission drafts rules on disclosing and verifying AI-prepared filings
• NSW Law Society guide maps solicitor AI duties to conduct rules
• Asia-Pacific AI regulation tracker
• AI Regulation News hub