Leaked proposal would make Australian copyright owners opt out of AI training

Opt-out and opt-in are two different rules. They differ in who pays for inaction, and here also in whether the remedy still works at the moment it is used. The phys.org analysis puts its weight on the second point: by the time an owner gets around to opting out, the material has probably been inside the neural network for a while, and the system cannot unlearn it ^[11]. Australia's Copyright Act gives the owner of a piece of writing or a photo the right to exploit those rights as they choose ^[9], and the analysis argues the burden should run the other way, with AI companies asking permission ^[10]. Its broader claim is that the dispute is about output quality and not only about individual works, because better training data produces better outputs ^[22].

The argument here is built on a thin public record. Deputy Prime Minister Richard Marles said, "The opportunity for Australia economically is enormous in collaborating with frontier companies" ^[5]. Minister for Industry and Innovation Tim Ayres has denied a reduction in copyright protection ^[6]. The details of the plan are not publicly available ^[7], so a notice period, a registry and a royalty are all still compatible with the same leaked document.

The copyright argument and the replacement argument rest on evidence of different strength. The piece cites studies suggesting readers prefer AI-produced works to those of human writers ^[13] and a recent study in which law professors preferred AI answers to those written by human colleagues ^[14], and it gives neither the number of participants nor the size of the preference gaps ^[20]. A preference in a rating task tells you what raters chose when asked, not what a publisher commissions.

The labour evidence is one association plus a set of accounts. A recent government report noted that occupations at risk of AI replacement had slower employment growth than those not at risk ^[15], and listed advertising and marketing professionals among the most exposed ^[16]. Exposed and unexposed occupations differ in many things besides exposure, and a comparison of their growth rates leaves AI mixed in with those other differences. The accounts are more specific. Visual artists, graphic designers and illustrators have reported lower wages and, in some cases, complete replacement ^[19]. Translators have been moved off primary translation and onto checking the accuracy of AI translations, with a decline in job satisfaction ^[17]. Graduate lawyers are being shifted onto verification of AI-generated content ^[18].

In my view the design question can be settled without any of that. Whether an Australian author's default is control or exposure is a choice about where the statute puts the burden, and the preference studies leave it open either way. If an opt-out scheme does arrive, the first thing to read in it is notice: how an owner learns the scheme exists, and how long they have before training runs.