Australia wants artists to block AI themselves

An Australian musician doesn't want her catalog used to train a model. Under the plan the federal government has been circulating since early September, here's what she'd have to do: install a technical block on the pages where her music lives, so crawlers can't get in.

Her songs aren't on her own website. They sit on streaming platforms, on her label's pages, on accounts she doesn't control. The keep-out sign she's being asked to post has to go up on land that isn't hers.

The document is real, and it's been public since Tuesday

Independent senator David Pocock tabled a document in the Australian Senate on Tuesday, 15 September, at 5.07pm: three pages titled "AI on Australian terms - Consultation proposals." Stamped confidential, it was shown to rights holder groups in early September, according to the ABC. We downloaded it from the Australian Parliament's tabled documents database and read it in full.

The first bullet on the first page states the goal: encourage frontier AI investment in Australia. Legal certainty for AI companies follows, then their concern spelled out in plain English, the "long tail": the mass of online content for which striking deals one by one is, the document says, neither realistic nor possible.

The document lays out two options. Both keep voluntary deals between rights holders and AI companies in place. Both add a statutory license to train on what the text calls "unprotected" material. And both offer creators the exact same way out.

"Digital protection" means robots.txt

This is the part the news coverage missed, and it's written twice. Rights holders who refuse training can opt out by adopting digital protection that stops AI from crawling and training on their content. In its list of strengthening measures, the document gives the example: "Increased use of digital protection for online content (eg robots.txt) to enable 'opt out' from training."

Robots.txt is a few lines of plain text sitting at the root of a website, telling crawlers what they're allowed to visit. It dates back to the 1990s. It carries no legal force on its own, and above all it assumes one thing: that you control the server.

That's where the method runs into reality. A novelist published by a house, a staff reporter, a photographer whose images live on a client's site: none of the three has any say over the file meant to carry their refusal. They own the rights. They don't own the server.

So "unprotected" doesn't mean "free to use." It means "not technically blocked." The Guardian reports that rights holders call the measure unworkable. It's easy to see why.

What each option pays for, and what it doesn't

Option A offers either payments to a central body that redistributes to registered rights holders, or a statutory extension of the licenses Australian collecting societies already issue for their members, stretched to cover non-members' material too.

Option B runs on a quota. An AI company earns the right to train on unprotected material once it has struck deals with a minimum number of businesses, for a minimum length of time. The slide doesn't dress it up, and the ABC quotes it verbatim: no extra money would be owed to creators beyond whatever was negotiated to hit the quota.

The rest of the document is more favorable to creators than other coverage suggested. It sets a steep penalty for circumventing protections, specific safeguards for Indigenous cultural and intellectual property, a duty to make serious efforts to keep pirated content out of training sets, plus transparency, oversight and audit requirements. There are real trade-offs in the text. What it moves is the starting point.

Europe signed off on the same mechanism in 2019

This Australian story speaks directly to anyone publishing in Europe. The EU's Directive 2019/790 contains, in Article 4, a text and data mining exception. Paragraph 3 sets the condition for it to apply: that the use of the works "has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online."

Machine-readable means. In practice, robots.txt and its cousins. It's the same mechanism as the Australian slide, voted into law seven years earlier.

The difference doesn't favor Europe. Article 4 provides for no payment at all. No central body, no deal quota, no extended collective license. Australia is at least arguing over what it gets in return for the shift. Europe voted it through and never opened a tab.

What Canberra had ruled out in October 2025

On 26 October 2025, Attorney-General Michelle Rowland issued a statement ruling out a text and data mining exception, saying the refusal was about "providing certainty to Australian creators." The same statement flagged what came next: examine a paid collective licensing framework, or stick with voluntary licenses.

The government kept that promise. It ruled out the exception, and it is consulting on licenses. What it never said out loud is that the starting point had switched sides. Pocock, quoted by the Guardian, draws the conclusion: it shouldn't fall to Australians to defend rights they already hold over what they create.

And the contradiction runs straight through the government itself. Assistant technology minister Andrew Charlton, asked about it on News24 on Tuesday, said he wouldn't trade the policy away: "So if we lose some datacentres over this, so be it."

The data center announced the next day isn't the one everyone's talking about

On Wednesday, the Guardian reports, Queensland premier David Crisafulli told his state's Parliament about a $31.9bn Anthropic data center, powered by a coal plant, on a site the local council hasn't yet approved. The state says it secured an exemption from the national renewable energy requirement.

The connection is tempting, and it would be wrong. The same article notes the facility will handle inference, answering user queries, not training the model. Training, and training alone, is what runs into copyright.

It was a training facility that Ann O'Leary, OpenAI's vice-president of global policy, was talking about in an interview with The Australian: without a change to copyright law, she said, her company won't be able to build one on Australian soil. The paper read that as an ultimatum. Deputy Prime Minister Richard Marles thinks that's overreading it, and OpenAI asked the paper to correct its headline.

Meanwhile, APRA AMCOS, which manages the rights of Australian and New Zealand artists, tells the ABC that no major AI platform has ever seriously tried to negotiate with the country's rights holders.