Australia Just Put the People Who Broke the Weather in Charge of AI
Australia just put the people who broke the weather in charge of AI.
Yesterday the Prime Minister announced an Office of AI. National standards. A mandatory framework. World-first, apparently.
This is the same government whose $96 million weather website drew 750,000 complaints. The old site is still running because people refuse to leave it.
Here's the part nobody is quoting.
The PM said the quiet part
The PM said the real issues AI presents are not technical ones.
Correct.
And that's exactly why this announcement worries me.
If the real issues aren't technical, then an office that writes standards and frameworks is aimed at the wrong layer. Standards govern the technology. The failures that actually hurt people happen in the experience, and this government has a fresh, expensive, well-documented track record there.
The weather site wasn't a technical failure
The BOM website wasn't a technical failure. The site was live. The weather data was there. The servers held. Every box an engineering review would check got checked.
It failed because nobody had the authority to say this isn't ready for real people.
Farmers couldn't find the radar loop they'd used every morning for a decade. The information architecture served the organisation's structure, not the public's questions. That's what happens when nobody designs for the ten-thousandth hour of use, for the person who opens the same page every morning as part of their job. And the response to 750,000 complaints was to keep the old site running alongside the new one, which is the most honest usability verdict a government has ever published.
I've written about this pattern before. The $64 million bus app that didn't work on launch day had the same anatomy: a design failure wearing a technology budget. So did the age verification law, where the failure was baked in at the policy layer before a single screen existed.
An office can't stop a launch
You can't fix that with an office. You fix it with someone whose job is the experience, not the announcement. Someone who can stop a launch.
That's the entire test. Not the org chart, not the framework document, not how many stakeholders signed off. When the thing is live, the data is flowing, the minister's speech is drafted, and a person in the room says real people can't use this yet, does the launch stop? The private sector fails this test constantly too. Sora shipped as a demo with a download button because nobody could slow the announcement down. Government just fails it with your tax money.
At the BOM, the answer was no. Nothing in yesterday's announcement changes that answer. An Office of AI adds a compliance gate, and compliance gates check what's measurable: data handling, model risk, documentation. Nobody fails a compliance review because a farmer can't find the radar.
Values are not a spec, and neither are standards
This is the second time in a year the government has reached for governance language where a design capability should be. Last time it was AI built with Australian values, a phrase no engineer can implement and no reviewer can test. Now it's national standards and a mandatory framework.
Neither answers the only questions that decide whether government AI helps or humiliates people. Who watches a real citizen try to use it before launch? Who owns the moment the model is wrong and a human needs to override it? Who has the standing to delay a minister's announcement because the experience isn't ready?
It's the public-sector version of what the whole industry is doing right now: $2.5 trillion on AI, almost none of it on making it usable.
Standards are the press release
Standards are the press release.
Design is the product.
Australia keeps shipping the press release.
An Office of AI could genuinely matter, if it hired people whose job is the citizen's experience and gave them launch authority. That would be a world first worth announcing. Until then, we've put the people who broke the weather in charge of something much harder to get right, and much harder to run the old version of when it fails.
---
