The Australian Government Wants AI Built With 'Australian Values.' That's Not a Spec.
The Australian government wants AI built with "Australian values."
I'm Australian. I've been in tech for 20 years. I have no idea what that means as a product spec.
What was actually said
The Albanese government told OpenAI, Anthropic, and every major AI lab to bake Australian values into their products or face a decade of regulation.
The intent is right. AI shouldn't be a black box that ignores local context. The products people use every day should reflect the communities they serve. Cultural sensitivity in AI outputs matters. Data sovereignty matters. Local language and context matter.
But none of that is what was communicated. What was communicated was a mood.
Values aren't requirements
"Australian values" isn't a requirement. It's a mood board.
Is it the interface? The model's refusal patterns? How Indigenous data is handled? What it does with health records? How it treats Australian English versus American English? Whether it knows that football means AFL in Melbourne and NRL in Sydney?
Nobody has said. And that's the problem.
If you've ever written a product brief, you know the rule. If the spec isn't specific enough to build against, it isn't a spec. It's a wish. And wishes don't ship.
What the labs will actually do
Without a real spec, the labs will nod, publish a blog post about their commitment to Australian users, and change nothing meaningful. Maybe they'll add a few Australian examples to their training data. Maybe they'll hire a policy person in Sydney. Everyone moves on.
This isn't cynicism. It's how product development works. Engineers build to specifications. When the specification is a vague cultural aspiration, the implementation is a vague cultural gesture.
The labs aren't ignoring Australia out of malice. They're responding rationally to an unactionable request. Give them a clear spec and they'll build to it. Give them a vibe and they'll perform compliance without delivering substance.
The gap between policy and product
This is what happens when policy teams try to regulate products without anyone in the room who builds them. It's the same failure mode behind Australia's age verification law and the $64 million bus app.
Policy people think in principles. Product people think in specifications. The gap between the two is where regulation goes to die, or worse, where it survives but achieves nothing.
A useful AI regulation for Australia would specify things like: models must be tested against Australian English comprehension benchmarks. Health-related outputs must comply with TGA guidelines. Indigenous cultural knowledge must be handled according to specific protocols developed with Indigenous communities. User data from Australian residents must be stored and processed according to defined data sovereignty standards.
Each of those is specific enough to build against, test for, and audit. Each of them advances the intent behind "Australian values" without relying on the labs to guess what that means.
You can't regulate what you haven't defined
You can't design to a vibe. You can't build against a mood. And you definitely can't regulate compliance with a concept that hasn't been operationalised.
If Australia wants AI that reflects its values, it needs to do the hard work of defining what that means in terms specific enough to implement, test, and enforce. That's not the labs' job. That's the government's job.
Until they do it, every announcement is just that: an announcement. And the AI products Australians use every day will continue to be designed in San Francisco by people who think a flat white is a paint colour. Meanwhile, $2.5 trillion is being spent on AI with almost none of it going into making any of it usable.
---
