A model can show you what your content does. It cannot decide what your content should say. Everything below sits on that line.
This has a deadline. Every non-corporate Commonwealth entity has until 15 December 2026 to have mandatory AI training in place for all staff. That comes from version 2.0 of the Policy for the responsible use of AI in government. The policy encourages corporate entities rather than binding them. The defence portfolio and the national intelligence community sit outside it.
The Digital Transformation Agency (DTA) built the foundational module, and it takes about 20 to 30 minutes. It is worth having. Before the policy existed, a Community and Public Sector Union (CPSU) survey found 92% of staff had received no AI training at all. It covered almost 2,000 people across almost 90 agencies.
That module is a starting point, not your training. It teaches responsible use across the Australian Public Service (APS). It does not teach you how to work on a page somebody reads to find out whether they qualify for a payment.
No policy mandates the second gap. So it lands on you. Here is what to do with it.
Before you open the chat window
Know which tool you are in
The DTA's staff guidance on public generative AI permits public tools for OFFICIAL information. Your agency's own policy still applies on top. Nothing at OFFICIAL: Sensitive or above goes in. Personal information and third-party copyright material stay out too.
The guidance names the trap itself. Microsoft 365 Copilot is an enterprise tool some agencies use. Microsoft Copilot is a public web tool aimed at individuals. They look almost identical.
If you cannot tell which one is open, you cannot tell what you may put in it.
Know where the decision sits
The same guidance is blunt. Generative AI must not make final decisions on government advice, services or outputs.
In content terms, that is the line between asking for 5 ways to phrase a heading and asking what an eligibility rule should say.
Know how to report it when it goes wrong
The policy makes every agency set up 2 reporting pathways by 15 December 2026. One for staff to report AI safety concerns and incidents. One for the public to do the same.
If nobody on your content team can name your agency's pathway, your team does not have one.
What to try first
Start where you are still the editor and the model is reacting to something you have already made.
Ask it to summarise a page you wrote, then check whether the summary matches what you meant the page to say. If the emphasis comes back wrong, your structure is telling readers something you did not intend.
Ask it to list the questions a reader would still have after finishing the page. Ask for 6 alternative headings for a section you are stuck on. Ask it to flag passive constructions and long sentences. Then make the calls yourself.
These are diagnostic uses. The model is showing you what your content does, not deciding what it should say.
What to leave alone
In a public tool: nothing above OFFICIAL. No personal information. No third-party copyright. That one is not a judgement call.
Leave alone the final wording on who qualifies, who pays, and what happens if they do not. Take the plain language help. Then have the person who owns the policy confirm the meaning survived the rewrite.
Leave accessibility sign-off alone. A model will produce plausible alt text and confident heading structures all day. The Web Content Accessibility Guidelines (WCAG) 2.2 come down to what assistive technology can parse. Fluent output that reads as though it should be fine does not meet them.
Leave alone content for communities you cannot check with a person from that community.
What decides whether any of this works
The teams I have watched use AI well in content work are not the ones with the best tools. They are the ones who could already tell a wording problem from a structure problem before they opened the chat window.
The strongest predictor of good AI output on a content team is whether the person prompting can already recognise good content.
Take a practitioner who knows the Australian Government Style Manual, and who can spot a page organised around an org chart instead of a task. They will get useful work out of a model. A practitioner who cannot will get fluent output and no way to judge it.
That is the gap the mandatory module never set out to close. It is also the one that decides whether AI helps your content or degrades it.
What good looks like
Everyone on your team can say which tool they are in, and whether it is public or enterprise.
Nobody on your team has asked a model what an eligibility rule should say.
Your team can name your agency's AI incident reporting pathway without looking it up.
The last thing a model produced for you was a diagnosis, not a final wording.
A person who owns the policy confirmed the meaning survived your last AI-assisted rewrite.
So here is the question for your next team meeting. When somebody gets fluent output back from a model, who on your team can tell whether it is right?
Photo by Viktor Forgacs on Unsplash.