Australia's AI copyright reform: what builders need to know about the data centre investment standoff

Australia's 1968 copyright law, written before the moon landing, has become the central barrier to tens of billions in AI investment. Anthropic has signalled interest in building its first data centres outside the US in Australia, but a Treasury briefing obtained under FOI confirms any major investment would be contingent on clarity of copyright settings. For AI builders, the outcome will determine training data access, licensing complexity, and whether Australia becomes a viable hub for model development or remains a market that consumes models trained elsewhere.

What happened

Prime Minister Anthony Albanese is set to deliver a major AI policy speech as the government weighs reform. Internal briefings show Anthropic cited Australia's stability, renewable energy potential, and close US ties as reasons to host data centres locally. The Australian Financial Review reported the deals discussed could be worth tens of billions.

UNSW copyright law professor Kathy Bowrey described Australia's copyright system as "incredibly complex and complicated" and "basically dysfunctional" for modern AI training. Each stage of training (data collection, copying, training, testing) could constitute a separate infringement. Australian law offers fewer defences than US fair use, and no AI copyright case has reached an Australian court.

Rights holders argue for control and fair value. APRA AMCOS, which handles royalties for Australian and New Zealand artists, says no major AI platform has made a genuine attempt to negotiate with Australian rights holders. Tech and business councils, including Atlassian co-founder Scott Farquhar, argue reform could unlock "billions of dollars of foreign investment".

Why AI builders should care

For developers building AI products in Australia, the policy risk is direct. If training on copyrighted local data remains legally risky, builders may develop proofs of concept locally then move offshore for scale, as Bowrey noted. Australians face no barriers using models trained overseas, meaning jobs and infrastructure investment flow out while local creators receive nothing.

A permissive reform could unlock access to Australian training data through licensing deals or a text-and-data-mining exception. That would attract data centre builds and reduce the friction of aggregating local content for domain-specific models. A restrictive outcome would push builders toward copyright-free data strategies or offshore hosting.

Practical implications

Australian company Maincode built its chatbot Matilda using only copyright-free material and customer-supplied data. This shows a viable path for specialist models, but the company acknowledged it is unclear whether a general-purpose system like ChatGPT or Claude could be built the same way.

The government has repeatedly ruled out a blanket text-and-data-mining exception. Collective licensing is under discussion, where one or more organisations negotiate payments for groups of rights holders. The media union MEAA has proposed guaranteeing an ongoing share of licensing payments to creators, regardless of copyright transfer. AI safety charity Good Ancestors proposed a permit system where companies purchase the right to train on copyrighted material, with a separate fund supporting creatives.

Caveats

The policy landscape is evolving. The prime minister is not expected to announce a decision in his upcoming speech. The US Copyright Alliance has warned that any Australian licensing deal could be cited in American courts as proof that licensing is feasible, creating global precedent. Bowrey cautioned that the debate risks asking copyright law to solve problems beyond its purpose, noting that "copyright has never been a cultural policy".

Questions around likeness protection, voice imitation, privacy, misinformation, Indigenous cultural rights, and AI safety remain unresolved. A copyright deal would settle only one part of the broader AI governance puzzle.