What the leaked slides reportedly propose
In mid-September 2026 the ABC reported on Attorney-General's Department slides titled "AI on Australian Terms", shown to rights-holder groups. As described by Region Canberra (17 September 2026) and IBTimes UK (20 October 2026), both options would let AI firms train on online material whose owners have not opted out. Option one has firms pay a central body or royalty collecting societies. Option two requires only a set number of deals with rights holders, with no further payment to creators. The slides also reportedly promise penalties for ignoring opt-outs and protections for Indigenous cultural knowledge. We have not seen the slides themselves, and the department has not published them. Everything here rests on press accounts of a leak.
The case for the government's approach
The strongest argument for a structured opt-out is practical. Frontier developers need legal certainty about training data, and Australian law has no clear answer. A scheme with a collecting-society payment route would give creators a revenue stream and give developers a predictable compliance path. Penalties for ignoring opt-outs would give rights holders a real remedy. Indigenous cultural knowledge protections address a gap that general copyright law has never handled well. If the alternative is years of litigation and no investment, a negotiated framework is not unreasonable.
Why it contradicts the October 2025 pledge
The problem is the sequence. On 27 October 2025 Attorney-General Michelle Rowland said the government "will not provide exceptions to text and data mining companies in any revision of copyright laws", adding that there were "no plans to weaken copyright protection when it comes to AI". She pointed to licensing and the Copyright and AI Reference Group as the way forward. An arrangement that allows training on anything not opted out is, in legal substance, a permission to copy that rights holders must actively revoke. Shadow Attorney-General Michaelia Cash called it "a text and data mining exception with an opt-out bolted on". That description is hard to rebut on the reported design alone.
Ministers deny weakening copyright, and the Deputy Prime Minister has denied that OpenAI issued an ultimatum. Reporting says OpenAI and Anthropic are tying Australian investment to clear copyright rules. That reporting is unconfirmed, and the denial stands. If it is accurate, it explains the timing but not the justification. Investment pressure is a reason to settle the rules quickly. It is not a reason to settle them quietly, or to settle them against a promise made eleven months earlier.
Option two is the weak link
The two options are not equal. Option one at least routes money to creators. Option two, as reported, makes training lawful once a firm has struck a set number of deals, with no further payment to the creators whose work is used. A rule like that rewards signing with large rights holders who can negotiate. Independent writers, photographers, musicians and small publishers, whose material is online but who sit outside any deal, would be covered by the permission without sharing in the payment. A pro-innovation stance does not require that. A functioning licensing market works because developers need to pay for access. A threshold of deals that substitutes for payment removes that incentive.
The opt-out itself is also a burden. It assumes every creator can find out that their work is being collected, knows how to signal a reservation, and can enforce it against offshore crawlers. The penalty promise helps only if enforcement reaches foreign firms.
Lessons from Productivity Commission and the UK
Australia has already been through one version of this fight. The Productivity Commission's interim report floated a TDM exception, and its final report, released on 19 December 2025, did not recommend one. It advised the government to wait and monitor for three years, noting uncertainty about licensing markets and the effect on creators. That is a sensible, evidence-led position. A decision this year to adopt an opt-out regime would pre-empt the monitoring period the Commission recommended.
The UK offers a closer comparison. Its consultation had presented a TDM exception with rights reservation as the government's preferred option. The impact assessment published on 18 March 2026 states that it "does not establish a preferred option". City AM reported Technology Secretary Liz Kendall saying the opt-out had been "overwhelmingly rejected" and that the government "no longer has a preferred option". The UK has not solved the problem either: its assessment does not monetise costs and benefits, and the underlying question of how creators get paid remains open. But it shows that an opt-out model can fail politically, and it did so after a public consultation.
What would work better
The fix is procedural before it is substantive. First, publish the options. A policy that shapes how every Australian creator's work is used should be consulted on openly, with the department's own economic analysis attached, not previewed to select groups. Second, if an opt-out is chosen, pair it with a remuneration right that cannot be bargained away, so that the money reaches individual creators and not only large rights holders. Third, require developers to disclose what they trained on, because neither opt-outs nor penalties can be enforced without it. Fourth, set a review date tied to the Commission's three-year monitoring window.
The government can keep AI investment and keep faith with creators. It cannot do both while describing an opt-out system as the absence of an exception. If ministers want to change their position, they should say so and defend it. Having ruled out the exception in October 2025, they owe the public a plain account of why a version of it is now on the table.