Submission 497
Mr Zac Broeren
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.
A personal submission by a machine learning student and AI safety worker argues catastrophic AI risks differ in kind from familiar harms and that behavioural testing is weakening as models advance. It makes three recommendations on the AI Safety Institute, DNA synthesis import screening and critical infrastructure coverage.
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 catastrophic AI outcomes are a different kind of policy problem because there is no person to penalise and no single decision to reverse.
There is no person to penalise in that story, and no single decision to reverse. That is what makes it a different kind of policy problem rather than a larger version of the familiar one.
Argues that situational awareness defeats behavioural testing and says a report finds this property is emerging.
Situational awareness is precisely the property that defeats behavioural testing, and it is one of the properties the Report says is now emerging.
Says a report finds general-purpose AI systems can now give expert-level guidance on building biological weapons.
The same Report found that general-purpose AI systems can now give expert-level guidance on building biological weapons, including detailed laboratory procedures.
Argues the Biosecurity Act already permits import conditions through BICON, so closing the identified gap needs no new legislation or agency.
The Biosecurity Act 2015 already lets the Director of Biosecurity impose conditions on import permits through BICON, so closing this gap needs no new legislation, no new agency, and no burden on legitimate researchers.
Argues AI data centres sit awkwardly outside the Security of Critical Infrastructure Act because of how that Act defines its sectors.
The data centres training and running general-purpose models sit awkwardly outside the Security of Critical Infrastructure Act 2018, because of how that Act defines its sectors.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Fund and mandate the AI Safety Institute to work on frontier and catastrophic risks, including loss of control, cyber misuse and dual-use biology, with pre-deployment model access.
Fund and mandate the Australian AI Safety Institute to work on frontier and catastrophic risks, including loss of control, cyber misuse and dual-use biology, with the pre-deployment model access that genuine evaluation requires.
Require synthetic nucleic acids to be imported from providers screening orders for sequences of concern and verifying customer identity, as a BICON import condition.
Require, as a BICON import condition, that synthetic nucleic acids be sourced from providers that screen orders for sequences of concern and verify customer identity.
Extend the Security of Critical Infrastructure Act to infrastructure training and operating general-purpose AI models, with ownership transparency, incident reporting and risk management obligations.
Extend the Security of Critical Infrastructure Act 2018 to cover the infrastructure training and operating general-purpose AI models, with the ownership transparency, incident reporting and risk management obligations that critical infrastructure status carries.
Weight accordingly the small number of decisions that cannot be revisited in five years if wrong.
Almost everything this Committee recommends can be revisited in five years if it turns out to be wrong. A small number of these decisions cannot be, and I would ask that they be weighted accordingly.