Australian tribunal orders costs over 'plain wrong' AI legal advice in dismissal case

Australia's Fair Work Commission has publicly condemned a dismissed supermarket worker's reliance on AI-generated legal advice, ordering him to pay part of his former employer's legal costs after his failed challenge to his sacking. According to ABC News, the decision is being read as a landmark caution for self-represented litigants who lean on chatbots as substitutes for legal judgment.

A rare costs order over AI reliance

The case centred on Sadnan Khan, dismissed by the supermarket chain ALDI, who took his case to the tribunal while representing himself. ABC News reports that despite repeated warnings his case had no substantial prospects of success, he continued with AI-assisted filings. Deputy President Michael Easton took what he called the "very rare step" of ordering Khan to pay $1,230 toward ALDI's legal fees, on the basis that his unreasonable conduct saddled the employer with avoidable costs. Easton added that if Khan had properly read his own AI-generated submissions, he would have recognised his case was doomed.

Khan told ABC News he had used a paid version of ChatGPT to help him unpack the legal cases cited by the tribunal and the company, arguing that a non-lawyer needs somewhere to do that study. He acknowledged errors, including forgetting to strip AI chat instructions out of documents he filed, and observed that the AI's manner of argument did not match how Australian courts communicate. A report commissioned by the tribunal reportedly shows a sharp rise in cases involving generative AI.

Despite the rebuke, Khan said he intends to appeal, planning to run a mix of Claude, ChatGPT and other agents to compare outputs. His appeal window closes before the tribunal's new rules come into force.

The tribunal did not condemn AI itself

The Commission was careful to separate the outcome from the tool. It acknowledged that AI is improving access to justice for applicants with genuine claims who might otherwise be deterred, and said AI use by applicants is not inherently problematic, noting that sophisticated users approach it critically as one tool among many.

ABC News points to Gregory Baker, a computer science lecturer at Macquarie University, who this month became the first person to successfully challenge casual employment laws using a team of AI agents. Baker described running his claim like a software project: a repository for his filings and a checking process that verified citations and logical coherence. He said asking ChatGPT to act as a bare oracle with no context does a terrible job, and that even his AI-prepared cross-examination practice did not stop the oral hearing from going unexpectedly. Macquarie University has until the following Wednesday to appeal his win.

Baker also estimated a gap of roughly five years between well-prompted frontier models and the weakest free ones, warning that pricing could entrench inequality because many people cannot spend hundreds of dollars a month on AI.

Expect tighter tribunal standards

Genevieve Grant, a law professor at Monash University, told ABC News that AI helps most when the user already understands the specific legal question, and that inexpert prompts produce inexpert results. She noted AI's sycophantic streak can convince people that weak claims have merit, and said Khan appeared to argue the wrong legal question despite the tribunal's attempts to redirect him. She expects courts and tribunals to deploy a widening set of tools, including costs penalties, and to raise the standards and reduce the latitude afforded to litigants, potentially deterring or filtering weak claims earlier.

The Commission has already introduced a template for AI-dependent litigants, encouraging them to copy prompts into their agent. Under guidance announced in the same week, applicants must from 20 October disclose any AI use, verify facts, authorities and hyperlinks, and confirm witness evidence, with consequences flagged for non-compliance.

Why it matters

This is a concrete, financial penalty attached specifically to how a litigant used AI, not merely a warning. It draws a line between critical, verified use and blind reliance, and it comes with enforceable disclosure rules that will shape how self-represented parties prepare filings. The split outcomes in the same tribunal, one applicant penalised and another victorious, suggest the dividing factor is process and verification rather than the technology itself. For anyone building or using AI assistants in high-stakes domains, it is an early signal that institutions will regulate the user's conduct, and that the cost of getting it wrong now has a price tag.