AI-drafted filing costs A$1,230 as disclosure rules land

A tribunal just put a dollar figure on unverified AI output. Australia's Fair Work Commission ordered a former ALDI employee to pay part of the company's legal bill after he ran a dismissal case built on ChatGPT-drafted submissions that got the central fact wrong. Seven weeks later, the same body's AI disclosure rules take effect. If you use a model to draft anything that leaves your building, both halves of this story are your problem.

What actually happened

Deputy President Michael Easton ordered Sadnan Khan to pay A$1,230 toward ALDI's costs, writing that "his unreasonable conduct caused his former employer to incur those costs unnecessarily." Costs orders are rare in unfair dismissal matters — each side normally pays its own way regardless of who wins.

The detail that made it unreasonable: Khan's case turned on his employment clearing six months, and his AI-assisted submissions argued the wrong dismissal date. He was told twice, in plain language, that the date on the record was the one that counted, and invited to withdraw. He continued. He used a paid tier of ChatGPT to interpret case law, and left the AI chat instructions sitting in his filed submissions.

Separately, the Commission published its final Guidance Note on generative AI. From October 20, 2026, anyone who uses ChatGPT, Claude, Copilot, or Gemini to prepare an application, submission, or witness statement in an individual rights matter — unfair dismissal, general protections — must disclose that use and verify the accuracy of the result, including facts, citations, and witness evidence. The framing is not prohibition. It's attribution plus a verification duty.

Why AI disclosure rules matter for your business

You are probably not filing in Australia. You are almost certainly shipping model output to someone who can penalize you for it.

The pattern generalizes fast: quotes, scopes of work, insurance paperwork, warranty responses, vendor questionnaires, W-9 chases, anything with a date or a number in it. A model will produce a confident March-versus-April error and no part of your stack will flag it, because nothing in a chat window knows which fields are load-bearing.

So decide which ones are, and put a check there. For us that means a drafting step and a verification step are separate things in the pipeline — the model writes, a deterministic check or a named human confirms the fields that carry consequence, and the artifact records which happened. Dates, amounts, party names, and citations get verified. Tone does not.

The disclosure half is the cheaper lesson. Regulators are converging on "say you used it, then prove you checked it," and shops that already log which documents came out of a model will answer that in an afternoon. Shops that don't will be reconstructing provenance from memory.

Sources: ABC News, Fair Work Commission.