Submission 323
Ms Janet Egan
Submitter group: Individuals.
Position labels are not shown for individual submitters.
An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.
Summary
Read the original submission for the full text.
Writing personally as a US-based think-tank researcher, the author argues Australia should enable AI training and clean data centres to gain frontier supply-chain influence and access, and discusses copyright reform as an option. Recommends substantially funding the AI Safety Institute and verification science, and screening synthetic DNA imports. Says Australia should not try to train truly sovereign frontier models.
Key points
Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.
Argues that if Australia can host a significant portion of global AI training compute, it stands to become a relevant AI player.
But if Australia can host a significant portion of global AI training compute, it stands to become a relevant AI player.
Argues Australia should not try to train truly sovereign frontier models, which the author calls a pipe dream given the costs.
This does not mean Australia should train truly sovereign frontier models—that remains a pipe dream. The costs of frontier AI development have become a barrier to entry.
Argues that Australian copyright law effectively prevents AI training in Australia, because paying for a licence could weaken the US fair use defence.
Australian copyright law effectively prevents AI training in Australia. By paying for a copyright license in Australia, an AI company effectively acknowledges that a market for training data exists—an admission that could weaken the US fair use defence that has supported AI development worldwide.
Says funding for the AI Safety Institute (less than $8m a year on average) pales in comparison to the challenge and to the UK equivalent.
But the magnitude of funding dedicated to this outfit (less than $8m a year on average) pales in comparison to the challenge. For comparison, the UK AI Security Institute is funded with over A$120 million per year in funding
Argues that no equivalent verification capability exists for AI: governments cannot currently verify how a model was trained or what compute was used.
But no equivalent capability exists for AI. Governments cannot currently verify claims about how a model was trained, what compute was used, or whether agreed conditions were observed.
Argues that attempting to control the flow of knowledge is difficult and can undermine legitimate research, and that requiring security screening of synthetic DNA orders is a much more feasible approach.
Moreover, attempting to control the flow of knowledge is a difficult and complex endeavour, and can undermine valuable legitimate research activity in Australia. Requiring security screening of synthetic DNA orders is a much more feasible approach.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Unlock AI training in Australia and enable a green AI data centre buildout.
Unlock AI training in Australia and enable a green AI data centre buildout, to earn Australia a role in the frontier AI supply chain.
Could use leverage from hosting AI training compute to shape frontier AI governance, for example by requiring incident reporting, government and third-party testing, or other risk management measures.
The Australian government could use its leverage to shape frontier AI governance and share benefits. For example, the government could require incident reporting, testing by the government and third-party organisations, or incorporate other risk management measures deemed appropriate.
Ensure deals with AI companies explicitly require AI engineers to be based in Australia, which the author says may be beneficial.
it may still be beneficial for the Australian government to ensure that deals with AI companies explicitly incorporate a requirement to have AI engineers in Australia.
Address the AI-training copyright barrier, for example through an existing proposal for companies to pay for training privileges, funding a Public Benefit Fund that improves creators’ position over the status quo.
There are already proposals on how to move forward: the government could reform copyright so that AI companies pay for the privilege of training in Australia, and use this revenue to set up a Public Benefit Fund to ensure Australian creators are better off than the status quo.
Invest in a world-leading AI Safety Institute to attract talent, access computing resources and undertake an ambitious research and testing agenda informing national and global AI policy.
Australia has an opportunity to build a world-leading AISI, whose expertise informs strong national and global AI policy, but this would require real investment—to attract the right talent, access computing resources, and undertake an ambitious research and testing agenda.
Invest in verification science to make AI agreements verifiable.
Irrespective of hosting AI training, Australia can contribute to better AI outcomes by investing in verification science—developing, testing and refining the technology needed to make AI agreements verifiable.
Require security screening of synthetic DNA orders, which the author says can be applied to all synthetic DNA imports under the Director of Biosecurity's existing import-permit powers without new legislation.
Requiring security screening of synthetic DNA orders is a much more feasible approach. Synthetic DNA is a necessary input to most pathways of concern in pandemic-capable pathogen development, and its production is concentrated in a small number of companies and countries. Australia can move quickly: the Biosecurity Act 2015 already empowers the Director of Biosecurity to impose conditions on import permits, so a screening requirement can be applied to all imports of synthetic DNA without new legislation.