AI Inquiry Atlas

An unofficial, independent guide to the public records of the Joint Select Committee on Artificial Intelligence. Not connected with the Parliament of Australia or the committee. Its makers also made a submission (no. 481; see About). Summaries and scores are written by AI and can be wrong: check the original before relying on them.

Submission 305

Anthropic

Submitter group: Companies.

An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.

Open in the interactive guide

Summary

Read the original submission for the full text.

An AI developer describes opportunities and risks from AI for Australia, covering its planned regional data centre investment, sovereign capability, workforce monitoring, a narrow conditional approval for AI training under copyright, frontier safety transparency, scams, and cyber and national security, and recommends measures in each area.

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.

  1. States that sovereign AI capability involves trust and control: assured access to capable systems and training, running, evaluating and securing them under Australian law and on local terms.

    In our view, sovereign capability in AI is a matter of trust and control. This includes assured access to the most capable systems from trusted partners; the ability to train, run, evaluate and secure those systems under Australian law and on local terms;

  2. Says it prefers building compute infrastructure in regional, rural and remote Australia.

    These factors and others, such as access to land and energy, mean that Anthropic’s preference is to build this compute infrastructure in regional, rural and remote Australia.

  3. Says it accepts that broad copyright exceptions have not received widespread support, and proposes that government could permit AI training through a narrow conditional approval.

    Anthropic accepts that broad exceptions from copyright, such as a text and data mining exception, have not received widespread support and have been ruled out by the Government. As an alternative, we believe that the Government could permit AI model training through a narrow form of conditional approval.

  4. Says AI could deliver scientific and productivity gains but also reduce demand for human work; longer-term impacts are harder to assess.

    Longer term impacts are harder to assess. Anthropic does believe that AI can deliver incredible progress in scientific discovery and workplace productivity but that these gains could also come with a reduction in demand for human work.

  5. Describes an essay arguing that frontier capabilities are outpacing work to keep models under control, so industry should slow capability improvements.

    The essay argues that frontier AI models are becoming capable enough that developers can lose control of them, as the OpenAI/Hugging Face incident showed. It says capabilities are now moving faster than the technical work needed to keep models under control, so the industry needs to slow the pace at which AI capabilities improve, which Dario calls pacing the frontier.

  6. Says that while AI is enabling more malicious activity, it is also the best tool Australia has for defending its people from these risks.

    While AI is clearly enabling an increase in malicious activity, AI is also the best tool Australia has for defending its people from these risks.

Recommendations

Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.

  1. Establish the conditions for safe and responsible AI investment and development in Australia.

    we firmly believe that Australia should establish the conditions for safe and responsible AI investment and development in this country.

  2. Prioritise trusted-partner frontier-model access, choice and control on acceptable terms, complementary to national AI initiatives; enable Australian frontier training to strengthen access, talent and supply-chain agency.

    Access to frontier models from trusted partners. For most countries the practical question is not whether to build a domestic frontier model but whether they can access, choose between and control the most capable models on acceptable terms. National and sovereign AI initiatives are complementary to, not competitive with, the work of frontier developers. There will not, and should not, be a single model that serves everyone. The ability to train frontier models in Australia will increase its agency over ongoing access, increase the pool of talent and elevate its position in the frontier model supply chain.

  3. Prioritise compute and energy, recognising training-scale compute scarcity, energy access and planning approval speed as constraints.

    Compute and energy. Training-scale compute is the scarcest input, and access to energy and the speed of associated planning approval processes are the main constraints on compute.

  4. Prioritise talent and research capacity: researchers, engineers, specialists and the universities and institutes that produce them.

    Talent and research capacity. The researchers, engineers and specialists who build, adapt, evaluate and secure AI systems, and the universities and institutes that produce them.

  5. Prioritise data and content, noting TPDI's assessment of strong Australian dataset agency but a disconnect with data-lifecycle maturity.

    Data and content. TPDI rates Australia's agency over its medical, geospatial, environmental, demographic and infrastructure datasets as very high, but finds "a substantial disconnect" between those strengths and Australia's maturity in managing the data lifecycle.

  6. Prioritise predictable rules for AI training, deployment and use, and domestic evaluation and audit capacity through the AI Safety Institute.

    Legal certainty and assurance. Predictable rules for training, deploying and using AI, and the domestic capacity to evaluate and audit systems, which Australia is starting to build through its AI Safety Institute (AISI).

  7. Prioritise trusted-partner alignment through Five Eyes, AUKUS and democratic institutions as a basis for sharing capable systems.

    Trusted-partner alignment. Australia's membership of the Five Eyes and AUKUS partnerships, and its democratic institutions, are themselves a source of agency: they are the basis on which the most capable systems can be shared.

  8. Support the ABS, Jobs and Skills Australia and DEWR to track potential AI-driven labour market disruption and to be able to trigger policy responses.

    Government statistical agencies like the Australian Bureau of Statistics, Jobs and Skills Australia and DEWR should be supported to track potential AI-driven labor market disruption and ensure they have the ability to trigger policy responses.

  9. Ensure the training and skills system can adapt quickly to role augmentation and, longer term, potential displacement.

    The training and skills development system must be able to adapt quickly to role augmentation and in the longer term, potential displacement.

  10. Investigate barriers to workers retraining or earning new qualifications, such as strict recognition of prior learning requirements.

    that create barriers to workers retraining or earning new qualifications – such as strict requirements for recognition of prior learning (RPL) – should be investigated.

  11. If impacts are seen across the labour market, consider incentives for firms that retrain, retain and redeploy workers.

    If impacts are being seen across the labour market, governments should consider incentives for firms who retrain, retain and redeploy their workers.

  12. Consider targeted AI-training approval for material potentially copyrighted in Australia, with possible conditions including verified technical protection measures or robots.txt controls and proposed AI standards; Anthropic is open to creator-supporting investment or other conditions.

    Through this mechanism the Government could provide a targeted approval for AI model training on material that could be subject to copyright in Australia, as long as the organisation undertaking the training met a series of conditions. These conditions could include respecting and verifying controls on content like technical protection measures or “do not crawl” signals through robots.txt and meeting the proposed set of AI standards. We understand that the Government may choose to require investment or other conditions designed to support the future of Australian creators and cultural endeavours, and we are open to considering those conditions.

  13. For the US: enact national frontier-AI testing and auditing law; empower the US Government to block or deter models posing a significant risk of catastrophic harm.

    Anthropic has called on the US Congress to pass a national law that requires testing and auditing of frontier AI. Given America’s key role in developing this technology, the US Government should have the authority to block or deter models that pose a significant risk of catastrophic harm.

  14. Lead diplomatic efforts towards verifiable international agreements, starting with a ban on using AI to make biological weapons.

    Protect the democratic lead and work toward verifiable international agreements. As an active, influential democratic ally and middle power, Australia can help lead this international diplomatic effort, starting with a ban on using AI to make biological weapons.

  15. Require the most capable model developers to publish safety frameworks, regularly report catastrophic-risk evaluations and notify Government of critical safety incidents, along the lines of SB 53 and the Advanced AI Framework.

    Make transparency on safety a core obligation on frontier model developers. We support requirements (explored within the Office of AI’s consultation paper on the proposed Australian Standards for AI13) on developers of the most capable models to publish a safety framework, report on catastrophic-risk evaluations at regular intervals and notify the Government of critical safety incidents, along the lines of California's SB 53 and our own Advanced AI Framework14.

  16. Appropriately resource regulators and enforcement agencies, including the ACCC, AFP and eSafety Commissioner, for scams, deepfake and online safety laws.

    the Government must ensure that regulators and enforcement agencies covering harms that have an existing legal framework are appropriately resourced. This includes the ACCC, the AFP and the eSafety Commissioner implementing laws such as the Scams Prevention Framework Act 2025, the deepfake offences in the Criminal Code and the Online Safety Act.

  17. Seek interoperability of Digital Duty of Care risk assessment and safety-control standards with other international safety requirements focused on AI developers.

    Ensure the international interoperability of the Digital Duty of Care. As the Australian Government develops the framework for risk assessments and safety controls under the proposed Digital Duty of Care legislation, it should seek to ensure that these standards are interoperable with other international safety requirements focused on AI developers.

  18. Support strict export controls on advanced chips and semiconductor equipment, close third-country loopholes, and adopt legal and threat-intelligence measures to deter distillation.

    this effort should be supported by strict export controls on advanced chips and semiconductor manufacturing equipment, the closing of loopholes that allow controlled hardware to be reached through third-country facilities, and legal and threat-intelligence measures to deter distillation.

  19. Position Australia as a preferred trusted jurisdiction for allied frontier compute and contribute to Five Eyes cooperation against chip smuggling and model distillation.

    Position Australia as a preferred trusted jurisdiction for allied frontier compute and contribute to Five Eyes cooperation against chip smuggling and model distillation.

  20. Accelerate adoption of frontier AI by Defence and the national security community.

    Accelerate the adoption of frontier AI by Defence and the national security community to ensure Australia maintains asymmetrical advantage.

  21. Pursue sovereign capability by enabling frontier model training in Australia, for assured access for defence, intelligence and critical infrastructure.

    Pursue sovereign capability by enabling Frontier model training in Australia to ensure that defence, intelligence agencies and critical infrastructure operators have more assured access arrangements to models and the people that build them.

Read the original

Open in the interactive guide