The expensive line in an AI budget is rarely the invoice for compute. The costly assumption is that someone had a right to train on the work inside the model. Australia is a useful place to watch that assumption get tested, not in a lab, but in print shops and studios in Grayndler, the inner-west Sydney electorate Anthony Albanese has long represented.
What the October record actually shows
On 26 October 2025, the Attorney-General’s Department issued a release titled “Albanese Government to ensure Australia is prepared for future copyright challenges emerging from AI.” Digital Policy Alert, citing that release, logs the announcement as a consultation on potential copyright updates for machine learning, and its event title records a rejection of a text and data mining exception. The tracker’s summary of the release is cut off after “The Government confirmed it will not introduce a…”. The operative sentence is not fully visible in that record.
Treat that as a secondary account of an official release. You cannot quote it chapter and verse like a statute. It is not a promise that every future amendment is off the table. Consultation on other updates was still open. What the record does not show is a tabled rule giving AI companies access to Australian creative work by default, or an official adoption of an opt-out scrape.
That distinction matters, because the public argument keeps collapsing three different things. Current copyright law is one. The October announcement is another: a consultation, plus a reported refusal to create a mining exception. A feared later compromise, in which online work would be available unless a creator objects, is a third. Only the second has a dated official footprint in the material that can be checked. The third is what working artists say they are fighting. Past training, here and elsewhere, is still unresolved.
A mining exception is a shortcut, not a courtesy
A text and data mining exception is the device that lets a developer copy and analyse copyrighted material to train a system without a licence from each rights holder. It is how “we scraped the web” gets dressed up as a permission.
The European Union already has a specific version, and it is worth naming rather than waving at a continent. Under the DSM Directive, Article 3 is a research exception. Article 4 allows commercial text and data mining of lawfully accessed works unless the rights holder opts out, including by machine-readable means for work published online. The United Kingdom is not that regime. It looked at a broad commercial mining exception and retreated.
Australia’s reported position in October 2025 pointed the other way: no mining exception, consultation still live. Rejecting one exception does not freeze the Copyright Act. It does mean the documented signal, as far as that truncated release goes, was not a settled right for AI companies to take Australian work unless a creator objects.
Opt-out moves the bill onto the person who already worked
The fight is about the default, and defaults are where money hides.
A licence, or any real opt-in, puts the burden on the developer: find the rights holder, agree a price or a permission, then train. Opt-out reverses it. The work is available unless the creator discovers the use and objects in whatever form the law, or the company, happens to recognise. For a novelist with a lawyer, that is a chore. For a freelancer who posted a portfolio five years ago and moved on, it is a scavenger hunt with no map.
Technology companies have long said the licence route is too slow and too fragmented for web-scale training. The difficulty is real. A model does not become lawful because the alternative was inconvenient. “Too hard to ask” is a business-model claim. It does not transfer title.
None of this clears what has already been trained. A vendor who shrugs and says the data was “publicly available” has a hope. They do not have a clearance. Work on the open web is often still copyrighted. You can reach a file without having any right to copy it.
What to ask before the renewal
If you are putting an AI agent into sales, service, or production, you inherit the training question whether or not you trained anything. An agent that drafts a pitch, pulls a file, and files the reply is only as clean as the model under it and the documents you feed it.
Three distinctions keep the purchasing conversation honest.
- Use is not training. A designer running a tool on her own brief is not the same act as a company copying a nation’s portfolios to build the tool. People can use these systems every day and still reject unlicensed extraction. Collapsing that into one “pro-AI” or “anti-AI” camp is how bad policy gets written, and how bad contracts get signed.
- Jurisdiction is not a sticker. An EU-style commercial opt-out, an American fair-use argument still being tested in court, and an Australian refusal to create a mining exception are three different permissions. Buying API access to a model trained elsewhere does not wash the question out of your contract. It moves the question into indemnity, venue, and what you will do when a rights holder writes to you rather than to the vendor.
- “Publicly available” is not a licence. If your vendor cannot say which is which, you have bought a dispute with a user interface.
Before the next renewal, ask four things in plain language. What categories of data was the model trained on, and under which legal theory? Does the indemnity cover training-data claims, or only what the model emits? Which venue decides if a claim lands? And if you fine-tune on customer files, employee writing, or your own marketing archive, do you actually hold those rights? Automating a process you do not own is a fast way to scale someone else’s claim.
There is a serious counterpoint, and it should not be waved away. Firms that want to train locally will say a strict licence regime pushes compute, jobs, and model development offshore, and that Australia then imports systems built on everyone else’s unresolved mess. A country can want a domestic AI sector and a consent rule at the same time. Pretending the second is a tax on the first is how a protection pledge and a studio revolt end up in the same electorate.
Compensation schemes, enforcement tools, and a clear list of covered material are not in the public summary of that October announcement. Do not let a vendor, or a minister, imply they exist because a consultation has been opened.
Why Grayndler hears it in the studio
Grayndler is an awkward place to have this argument in the abstract. Reporting on the electorate, drawing on a Creative Australia profile, describes more than 3, 000 cultural and creative businesses, and a workforce in which 10 per cent of workers are in cultural and creative fields, against 4.2 per cent nationally. Albanese has, in the past, spoken proudly of the seat’s live music, theatre, and artist spaces. He has also been reported as promising “the strongest possible protection, ” and as warning that unpaid use of creative work would be “theft.”
That language sits more easily beside a rejected mining exception than beside a scrape-first law. It has not settled the people who make the work. Their objection is the fear that access to Australian creative work online becomes the default, with the creator left to opt out. Hold that as their characterisation of the risk, not as settled policy. The October record does not show the government adopting it.
There’s obviously an AI gold rush at the moment, it’s the wild west and we really need to push back. It’s hard enough being a starving artist these days without having all your work taken and used in these nefarious ways.
That is Pete Strong, a screenprinter whose business is based in Grayndler, as quoted in coverage of the local backlash. He has also said AI could be “good for humanity” and is now an “extraction industry mining for profit for the few.” The gold-rush line is a metaphor. The consent point under it is the part that holds.
Inner-west designer Lu Campbell-Smith has said she finds AI useful in practice and still calls having to opt out “outrageous.” Her question is the one a finance director should be able to repeat without a briefing note: “Why would anyone want to make anything and post anything online if … you’re creating free content for an AI machine?” She has also described Australia as a test market, “a little resource to be exploited.” That is her view of the power imbalance. It does not need to be promoted into a theory of history.
At Tortuga Studios and the aerie, an artist-led space in an old brick factory in St Peters, author and curator H Morgan-Harris has said the creative ecosystem could be “decimated, ” that the conversation “always comes back to money, ” and that artists are “the first to be thrown under the bus.” Her own copywriting trade, she has said, has already been hit hard. Jason Zoidberg, co-founder of Monster Mouse Studios in Marrickville, argues the industry has already, “without consequence, ” taken artists’ “collective labour, ” and that it would be “utterly crazy” to authorise that after the fact. He has also noted a local twist: workshops at his studio are drawing more participants amid an inner-west backlash against social media and AI. He called it a niche, “a little inner west bubble.”
On the reporting, this is not a single letterhead campaign. Working artists in different rooms are reaching for the same objection: consent, payment, and the fear that they will carry the cost of someone else’s model.
Questions worth asking before the next draft drops
Did Australia legalise unlicensed scraping of creative work?
Not on the record dated 26 October 2025. Digital Policy Alert, citing the Attorney-General’s Department release, logs that announcement as including a rejection of a text and data mining exception, with consultation on other copyright updates still open. The summary is truncated, so the full wording is not visible. Default or opt-out access is what Grayndler creatives say they are opposing. It is not a law you can cite.
Why do creators treat opt-out as a trap rather than a compromise?
Because the burden sits on the person who already did the work. They have to find the scrape and object, often years after the file went online. A licence or an opt-in rule puts that burden on the developer who wants the data.
Does rejecting a mining exception end the fight?
No. The same announcement left consultation open. Refusing one exception is not a ban on every later amendment, and it does not settle training that has already happened. Artists quoted from Grayndler are reacting to the prospect that a later compromise will let companies take online work unless a creator objects, not to a finished statute. That prospect and the October refusal are different things.
Is this only an arts story?
No. Any business deploying agents or automating sales and service work is downstream of the same consent design. If the model’s inputs were taken on a contested theory, the risk does not stay in the studio. It shows up in your contract, your customer files, and whatever you use for fine-tuning.
What is still unknown?
Who would be paid, how an objection would be enforced, which material would be covered, and whether a later compromise reopens a mining right the government was reported as declining to create in October 2025. Anyone who fills those blanks with confidence is selling you a mood. You still do not have a policy.
When this consultation hardens into text, the companies that look competent will be able to say, in a sentence a creator and a general counsel would both accept, whose work the system learned from, and why that was allowed. The loudest slide about innovation will not get them there.