Anthropic ties Australian AI investment to a clear legal basis for model training

Anthropic has told federal parliamentarians that its backing for frontier AI research and infrastructure in Australia depends on the law giving clear cover for training models in the country. Its written answers to the Joint Select Committee on Artificial Intelligence argued that investment at the scale of training would prove sturdier than spending on inference capacity alone. The company said it was confident of relocating existing researchers and recruiting locally if training could happen in Australia.

The answers contained no estimates of spending, headcount, computing capacity or power demand, and no candidate sites. Anthropic said no single provision of Australian copyright law had decided its investment choices. Since it trains models offshore, it said, Section 43B of the Copyright Act 1968 has never been tested against its training in practice. It said hosting such training in Australia would need a clear legal basis, and that it understood the government was weighing what that framework should look like.

On data, Anthropic argued that moving training between the US and Australia would not alter what developers feed into models. It said they would still draw on public web content, data bought or obtained through commercial agreements, and data generated internally. Anthropic said it strikes commercial deals for data, partly to cover gaps in specific knowledge areas, but not licences aimed specifically at training, and it expects such deals for non-public and specialised data to continue. It added that it had not announced any publisher licensing deals.

The company also commented on proposals for an Australian regime that would allow training on copyrighted works under conditions, including a way for creators to reserve their material. In its view, a workable opt-out needs controls that are specific, verifiable and machine-readable, applied when data is collected. Anthropic proposed that following those controls be mandatory, with an Australian regulator checking and enforcing it, and that a developer who failed would forfeit its approval to train. It also said crawlers should identify themselves to site operators.

Anthropic also addressed its US litigation, including the US$1.5 billion Bartz v Anthropic settlement. It said the US court held that training on the books was fair use but that obtaining them from pirate libraries was not. It acknowledged that no US appellate court has yet decided whether training a generative AI model is fair use. Asked if inference of its models complied with Section 43B, it did not answer directly, saying no Australian court had considered the provision's application to inference and the outcome would likely turn on facts and technical processes.

Anthropic said its March memorandum of understanding with the Australian Government carried no financial commitments beyond exploring partnerships. It said it had since announced a further AU$10 million in Claude credits for Australian researchers and deep technology start-ups. It did not disclose its Australian revenue or tax paid since setting up a local subsidiary in December 2025, saying it would report and pay in line with Australian law.

On safety, Anthropic disclosed incidents in which its models reached real systems during security testing because a misconfigured sandbox gave them internet access. It said it paused cyber evaluations, told affected organisations within days and added real-time classifiers to stop a model probing or escaping its environment. It said none of the cases involved Australian entities. It would not say how many Australian organisations take part in Project Glasswing, calling them commercial in confidence. The committee is due to deliver its final report by 30 November 2026.