Why a Workplace Tribunal Banned AI From Employee Correspondence

In late July 2026, Australia’s Fair Work Commission ruled on a workplace bullying complaint brought by a caretaker employed by a Queensland body corporate. The commissioner found the caretaker had, in fact, been bullied. She also found something else worth reading closely: the caretaker’s own correspondence, which appeared to have been prepared with an AI tool, had made the situation harder to resolve, not easier.

The order that followed banned AI from the parties’ future correspondence entirely. Not a disclosure requirement, a ban.

That is a different problem than the one most “should you use AI at work” posts argue about. Nobody here was accused of hiding AI use or misrepresenting authorship. The objection was to the text itself: what it said, how long it took to say it, and the tone it defaulted to.

What the Commissioner Actually Objected To

Commissioner Sarah McKinnon described the caretaker’s AI-assisted correspondence as “lengthy, wide-ranging, replete with generalisations, repetitive, and often couched in accusatory language,” according to ABC News’s report on the ruling. She added that it was “unsurprising” the employer had eventually stopped engaging with the messages at all.

Going forward, she ordered that correspondence between the two parties be “accurate and respectful,” kept “brief and to the point,” with “no use of artificial intelligence tools in the preparation of correspondence between the parties.”

Read that list again. None of the four complaints are about detection. They are ordinary writing failures: length without density, sweeping claims standing in for specifics, the same point restated across paragraphs, and a tone that escalates a dispute instead of describing it. A good rewriting tool exists to remove exactly those failures. Here, whatever tool was used appears to have added them instead.

This Is a Drafting Problem, Not a Disclosure Problem

Most coverage of AI at work centers on whether readers can tell, and whether that matters. We covered that question in should you disclose that you used AI to write an email. This case does not turn on any of that. The commissioner was not weighing authenticity. She was assessing whether the messages were useful, accurate, and proportionate to the dispute, and found they were not.

That distinction changes what you should check before sending. Disclosure is a policy question with a yes or no answer. Whether a message pads, generalizes, or escalates is a quality question you have to read for every time, regardless of whether you plan to mention a model was involved.

Why General-Purpose AI Drafting Produces This

Ask a general chat interface to “write an email explaining that my employer keeps changing my duties and it’s affecting me,” and it tends toward three defaults: it over-explains to sound thorough, it restates the complaint in different words across several paragraphs, and it drifts toward the emotional register of the prompt rather than a measured one, because that is the register it was handed. None of this is malicious. It is what a model built to sound comprehensive does with an open-ended, emotionally loaded instruction.

Before:

I am writing to formally address an ongoing and deeply frustrating issue that has been significantly impacting my ability to perform my role effectively. Despite multiple attempts to seek clarity, I have consistently been met with vague, inconsistent, and at times contradictory direction regarding my actual responsibilities. This lack of communication is not only unprofessional but has created a hostile and confusing working environment that I believe warrants serious attention.

After:

On three occasions this month (July 4, July 11, and July 19) I was given conflicting instructions about weekend gate access without a written update to the duty list. I am asking for one current written description of my duties, confirmed by [date], so future instructions can be checked against it.

The first version contains an opinion about the workplace and no facts a reader could act on. The second contains three dates, a specific recurring problem, and a concrete request. A reader can resolve the second one. They can only argue with the first.

The Compliance Backdrop Is Also Tightening

This ruling is not an isolated data point. Weeks later, the Fair Work Commission’s president published a guidance note on the use of generative AI in Commission cases, effective October 20, 2026. From that date, anyone using tools such as ChatGPT, Claude, Copilot, or Gemini to prepare an application, submission, or witness statement must disclose that use and confirm they checked the document’s accuracy, with particular attention to case references, a direct response to fabricated citations turning up in filings.

That rule is narrower than the bullying-case order: it covers formal documents filed with the Commission, not everyday workplace email. But the direction is the same. Regulators reading AI-assisted text are increasingly willing to name what AI writing gets wrong, rather than treating “AI was involved” as the whole issue. As of this writing in September 2026, no equivalent rule exists for ordinary internal correspondence, and one case plus one Commission’s procedural rule is not yet a general legal standard for work email.

How to Draft a Hard Workplace Message Without the Same Failure

The fix is not avoiding AI for difficult messages. It is giving it an instruction that forbids the specific things this ruling objected to, instead of a generic “make this sound professional.”

A Wrivio Context for a workplace complaint or dispute message could say:

Rewrite this as a factual account of a workplace issue. State only what happened, when, and what response is being requested. Keep every date, name, and quoted instruction exactly as written. Do not add adjectives describing the other party’s conduct, restate the same point in more than one paragraph, or generalize beyond what is stated. Target under 150 words.

Press Ctrl+Shift+Space, paste the draft, and check the diff for what changed rather than trusting the final text on sight. A rewrite that added a sentence of context you did not write, or softened a date into “recently,” is worth catching before you send it. How to review AI rewritten text covers what to look for in that diff. If the dispute involves details you would rather not send to any server, Local mode keeps the rewrite entirely on your device, which matters for anything covered in is it safe to paste work emails into ChatGPT. If the draft you are starting from is already angry rather than just long, how to rewrite an angry work message is the more direct starting point.

Common Questions

Did the tribunal ban AI from workplace email in general?

No. The order applied only to future correspondence between the two specific parties in that bullying case, and no broader Australian law currently bans AI-assisted workplace email.

Does this mean I shouldn’t use AI to write a workplace complaint?

No. It means checking the result for the specific failures the commissioner named: padding, generalized accusations, repeated points, and escalating tone. A narrow, fact-preserving instruction is far less likely to produce those than an open-ended request to “make this sound professional.”

What is the Fair Work Commission’s new AI disclosure rule, and does it cover normal emails?

From October 20, 2026, it requires anyone using generative AI to prepare an application, submission, or witness statement filed with the Commission to disclose that use and confirm they checked its accuracy. It applies to documents filed in Commission cases, not to everyday workplace email.

How can I tell if an AI-drafted message is padded or overgeneralized?

Compare its length to what a plain, factual account of the same events would need. If a claim is a sweeping description rather than a specific instance with a date, or if the same point appears in more than one paragraph, that is the pattern the commissioner criticized.

Can a rewriting tool avoid these problems automatically?

Only if the instruction tells it to. General-purpose chat interfaces default toward elaboration and a tone that mirrors the prompt. A Context that forbids added adjectives, restated points, and generalization, and that caps length, produces output much closer to what a written record can actually be judged on.

Download Wrivio for Windows to keep a hard workplace message factual and brief instead of long, repetitive, and easy for a reader to stop engaging with.