Responsible AI use in government content isn't a policy question. It's a workflow question.
Every Australian government agency I have worked with in the last twelve months either has an AI use policy or is writing one. The policies almost always say the same things. Use AI ethically. Be transparent. Maintain accountability. The policy is approved, posted to the intranet, and a content officer sits down on a Tuesday afternoon to write a fact sheet — and has no idea what any of it means for what they are about to do.
Responsible AI use in government content is a daily workflow problem, not a policy problem. Through my work with ACT Government, the same pattern showed up everywhere. The governance documents were written for executives. The people doing the work needed something different — concrete decisions they could apply page by page.
Here is what responsible AI use actually looks like once the policy is on the wall.
Drafting aid, not decision maker
The single most useful distinction I have found is the line between using AI to help draft something and letting it make a content decision.
Drafting support is low-risk. Asking a model to suggest five rewrites of a paragraph, summarise a long document into bullet points, or generate alternative headings is a productivity gain. The human is still the editor, the decision maker, and the accountable party.
Letting AI decide what to publish, what to remove, or how to describe a sensitive issue is a different category of risk entirely. That is no longer drafting. That is delegating editorial judgment to a system that cannot be held accountable. Treat the two cases differently in your guidance and in your review process.
Transparency that is operational, not aspirational
Most AI policies include a clause about transparency. Few define what transparency actually requires of the content team.
In practice it means three things. Your team knows internally which content has been AI-assisted and how. Your reviewers know that an AI-assisted draft needs a different kind of review than a human-only one — not better or worse, just attentive to different failure modes. You have decided whether and how AI assistance is disclosed externally, and the decision is documented.
Transparency that is not written down is not transparency. It is good intentions waiting to be forgotten. The Digital Transformation Agency's guidance on AI is a useful starting point. It is general by design — your team has to translate it into a workflow.
Quality checks belong to the human
The Style Manual and WCAG do not change because AI helped produce the content. The Australian Government Style Manual still requires plain language, structured headings, and accessible link text. WCAG 2.2 still requires meaningful alt text, logical reading order, and keyboard accessibility.
AI tools are improving at flagging plain language problems and accessibility gaps. They are not yet reliable at fixing them, and they should not be the final check. Plain language depends on whether your reader can act on the content. Accessibility depends on whether real assistive technology can parse it. Both of those are human judgments grounded in user need, not text patterns.
Accountability sits with a person, not a system
If AI-assisted content causes harm — wrong information, inaccessible structure, language that excludes a community — accountability cannot sit with the model. It sits with the person who approved the publish.
This is the part of responsible use that quietly disappears in most policy documents. Naming the accountable role is more useful than naming the principle. In the work I have led, the most reliable governance change was making the clearance owner explicitly responsible for the AI-assisted parts of the content too. Not the AI lead. Not the policy team. The person whose name goes on the publish.
A short practical check before publishing
Before AI-assisted content goes live, the person clearing it should be able to answer four questions.
Have I checked the facts and citations against a primary source, not the AI output?
Does the content meet plain language and accessibility standards on its own merits?
Have I documented internally that AI was used, and how?
Am I comfortable being the named accountable person for what is being published?
If any of those is uncertain, the content is not ready. That is what responsible use looks like on a Tuesday afternoon at the desk where the content gets made.
Photo by Google DeepMind on Unsplash.