OpenAI and Anthropic urge Australia to relax its AI training ban, arguing that a narrowly conditioned pathway could protect creators while allowing models to learn from local material. The proposals appear in September submissions to a parliamentary inquiry examining artificial intelligence laws.

 

Australia has rejected a broad copyright exemption for AI developers, effectively preventing companies from training commercial models on protected Australian works without permission. The dispute now links copyright policy to planned data-centre investment as lawmakers prepare recommendations due in November.

 

The companies are asking policymakers to consider three changes:

  • A limited route for approved AI training.
  • Conditions that support Australian creators and culture.
  • A framework for collaboration with copyright owners.

 

OpenAI and Anthropic Target Australia’s AI Training Ban

The current conflict is not a prohibition on AI research generally. It arises because Australia has ruled out creating a new copyright exception that would let developers copy protected material for commercial model training without the authorization normally required under existing law.

 

For major model developers, that position makes domestic training difficult when datasets include books, news, music, images or other creative works. Training requires large-scale copying and processing, while Australian copyright law has narrower exceptions than the broad fair-use doctrine available in the United States.

 

Anthropic acknowledged in its submission that a sweeping exception lacks political support. It instead proposed a narrow conditional-approval mechanism under which the government could permit specified training activity if a company accepted obligations intended to benefit creators and cultural institutions.

 

OpenAI called for a balanced framework allowing models to learn from publicly accessible information while giving rightsholders opportunities to work with developers. The submission does not itself change the law, and neither company received an exemption through the inquiry process.

 

Conditional Approval Would Replace a Blanket Exception

Anthropic’s alternative is materially different from an unrestricted text-and-data-mining exemption. A conditional approval could be limited by company, model, dataset or purpose, with compliance requirements attached before training begins. The submission left the government to determine the final conditions.

 

Possible obligations could include investment, licensing arrangements or support for Australian creative work, although no final mechanism has been adopted. The central policy question is whether those conditions could produce meaningful compensation and control rather than simply legal certainty for AI companies.

 

Creators, publishers and other rightsholders have resisted exceptions that allow commercial systems to use their work without consent or payment. Their concern is not confined to copying during training: generative products can also compete with the articles, illustrations, recordings and books from which models learn.

 

Related Research

 

Data-Centre Deals Raise the Economic Stakes

Both companies connected their copyright arguments to infrastructure investment. OpenAI has an offtake agreement with Australian data-centre developer NextDC for a planned Sydney project, while Anthropic was recently named as a partner in a Queensland data-centre development.

 

Those projects give the companies a practical argument: permissive training rules could make Australia more attractive for compute investment. Yet a data centre can serve inference and overseas workloads without necessarily requiring local copyrighted works, so infrastructure spending and training permissions are related negotiating issues rather than the same policy decision.

 

Australia is already developing broader rules for AI and data centres that are expected to take effect next year. The government is weighing electricity demand, water use, grid connections and community benefits alongside questions about safety, copyright and control of Australian data.

 

The result could become a template for countries seeking both stronger creator protections and a larger share of global AI infrastructure. A conditional regime would let Australia test whether investment commitments can be enforced without creating a broad exception that permanently weakens copyright rights.

 

Australia’s November Report Will Set the Next Stage

The Joint Select Committee on Artificial Intelligence is due to report in November. Its recommendations could endorse the government’s current position, propose a licensing system or outline a narrow approval process, but the committee does not by itself rewrite the Copyright Act.

 

Any workable framework would need clear definitions of eligible material, acceptable acquisition methods, creator compensation, transparency and enforcement. It would also need to address how a rightsholder can identify whether a work entered a training dataset and what remedy applies when conditions are breached.

 

OpenAI and Anthropic have therefore opened a negotiation, not secured a policy victory. The next concrete signal will be whether the committee treats conditional approval as a credible compromise or concludes that model developers should obtain permission through existing copyright and licensing channels.