The Prime Minister appears poised to make a major statement about the social license of artificial intelligence (AI) on Wednesday.
It’s hard to see how he announces anything meaningful without first resolving the festering copyright question, because under our current settings nobody, local or foreign, can train a meaningful AI model in Australia without taking on enormous legal risk. That is the issue blocking serious AI investment here, and with it any chance of Australia earning a seat at the global AI table.
So what can the PM actually do?
It was reported last week that the AI companies, with Anthropic at the centre, offered a package to break the deadlock: a text-and-data-mining exemption in exchange for annual payments into a creators’ fund and tens of billions in data centre investment.
The rights holder bodies rejected it flat out. Their position, put plainly by News Corp’s Michael Miller among others, is that AI companies should instead negotiate directly with content owners.
Direct negotiation is a dead end. No AI lab, anywhere, has cleared training rights holder by rights holder. They sign content deals with large publishers, News Corp’s own arrangement with OpenAI is reportedly worth more than US$250 million over five years if the Wall Street Journal is to be believed, but nobody has ever done, or offered, the work-by-work deal the Australian creative lobby is holding out for.
The reason is simple. Agree individual terms with one rights holder and you’ve conceded the principle to every rights holder on earth, in every jurisdiction, each wanting their own slice.
It’s a can of worms no lab could ever close, and the imbalance of power is stark: Anthropic’s valuation has climbed from US$61.5 billion to a reported US$965 billion in a little over a year, and nothing requires it, or any lab, ever to train a model in Australia.
But suppose, for argument’s sake, a lab signed generous agreements tomorrow with APRA AMCOS and the Copyright Agency. It still hasn’t bought safety, because collecting societies only represent their members.
Every Australian rights holder outside those bodies remains a live claim, and the recent Anthropic settlement in the United States has effectively priced that risk at a little over US$3,000 per work.
No eggs unscrambled
That settlement is worth understanding properly, because it shows where the hard boundary sits. The court found that training on lawfully acquired books was fair use; what Anthropic agreed to pay US$1.5 billion for was downloading and keeping pirated libraries. The remedy was money and the destruction of the pirated source files.
Nobody proposed removing the influence of the 482,460 books from the models themselves, because it cannot be done. Models aren’t trained by picking one book and one song. They ingest vast pre-assembled snapshots of the web, with no map connecting each fragment back to a rights holder.
An author’s words live on in reviews, quotes, fan forums and pirated mirrors on domains they’ve never controlled. Even a “clean” new model inherits the old content’s influence through its starting weights and the synthetic data earlier models generate. You cannot unscramble the egg.
Others have already run the experiments for us. The UK proposed an opt-out scheme and formally abandoned it in March after concluding it couldn’t be made to work. The EU has an opt-out that creators say they cannot actually operate.
OpenAI promised authors a per-work opt-out tool back in 2024 and then quietly never shipped it. Some rights holders in, some out, each cutting their own deal fails technically, because works can’t be unpicked, and legally, because partial coverage leaves partial liability.
Meanwhile the pervasiveness of AI keeps making the debate look faintly absurd. On 1 July, artists rallied at Parliament House demanding the government hold firm.
The same day, quietly, the new whole-of-government Microsoft agreement commenced, giving every federal agency access to Copilot, built on models trained on the very content the artists say was taken.
Senator David Pocock calls the mooted arrangement “the ultimate dirty deal”, yet published AI deepfake videos of Albanese and Dutton as a 2024 stunt and is an open user of these tools. Senator Sarah Hanson-Young says Labor “must not sell out Australian creators, journalists and our environment for the profits of big tech and the AI juggernauts”; presumably then the Greens’ offices have stopped using ChatGPT, Claude and Gemini in solidarity?
I’m being tongue in cheek, and to be clear I’m not asking them to stop.
Nobody can; not the senators, not the public service, and certainly not me, my industry runs on these tools. The only question left with any life in it is whether the Australians whose work sits inside it ever see a dollar.
So the honest design space is brutally narrow. It’s all in or all out.
The Copyright Act as solution
Happily, the all-in mechanism already exists in our own Copyright Act. Australian schools have copied protected material for decades under a statutory licence: everyone is covered by law, nobody signs anything, nobody forfeits anything, and when the parties can’t agree on price, the Copyright Tribunal sets one.
Extend that model to AI training. Every Australian rights holder covered automatically – member of a collecting society or not – with the rate benchmarked against real global deals and the money actually flowing to creators. The labs get certainty under Australian law, which is the only certainty Canberra can offer, and Australia gets the data centres and the capability.
The government should still reject a free carve-out. The work has enormous cultural value, the kind that will never feel properly compensated regardless of what the AI companies pay.
And they will pay.
But there’s a limit, and the creative lobby needs to accept that the mechanism can’t be bespoke, opt-in or run work-by-work. One scheme, everyone covered, real money.
The alternative isn’t protection. It’s unpaid.
- Adam Barty is CEO of digital and AI consultancy Revium.
