Submission 453
Mrs Michelle McAllister
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.
Argues AI governance should keep model-level regulation but expand to cover lineage between systems, populations of interacting agents, dependency on foreign frontier providers, and standing institutional memory of capability developments. Makes six recommendations including lineage provenance, population-level monitoring, a capability-signal register, pre-registration for Commonwealth deployments, and a bioethics framework.
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 governance organised around models, developers, training data and outputs remains necessary but is no longer sufficient.
Australia's artificial intelligence (AI) governance has largely been organised around a visible object: a model, the company that built it, the data it was trained on, and the outputs it produces. That remains necessary. It is no longer sufficient.
Acknowledges narrower alternative explanations of research findings but argues that apparently innocuous model outputs can preserve traces of their source system that alter another compatible model.
Subsequent work has proposed narrower explanations, including rare divergence tokens, fine-tuning artefacts and steering-vector distillation.2,3,4 I include those alternative readings deliberately. They make the phenomenon less mysterious without removing the governance issue: apparently innocuous model outputs can preserve traces of the system that produced them, and those traces can alter another sufficiently compatible system.
Argues that population behaviour of interacting agents is a different object of analysis from individual-model behaviour, without assuming such populations are inherently dangerous.
These are not reasons to assume autonomous-agent populations are inherently dangerous. They are reasons to recognise that population behaviour is a different object of analysis from individual- model behaviour.
Argues, without taking a view on which party was right in a dispute, that the episode demonstrated access to a frontier-model provider can become entangled in executive action inside an allied country.
The Australian significance does not depend on deciding which party was right. The episode demonstrated that access to a frontier-model provider can become entangled in executive action inside an allied country.
Argues that Australia needs a standing function recording emerging capability developments and competing explanations.
Australia needs a standing capability-signal function that records emerging developments, competing explanations and policy significance without requiring premature certainty.
Argues that Commonwealth agencies should evaluate the interaction between AI systems and users, not just task completion, without assuming such effects are inherently harmful.
The policy implication does not require a claim that these effects are inherently harmful. It requires agencies to evaluate the interaction between system and user, rather than task completion alone.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Incorporate lineage provenance into AI governance, Commonwealth procurement/assurance and any future mandatory standards.
Incorporate lineage provenance into Australia's AI governance framework, Commonwealth procurement and assurance requirements, and any future mandatory standards.
For deployed systems, record source models generating synthetic training/fine-tuning data and relevant architectural relationships; for agentic deployments, record memory files, skills, system instructions and identity artefacts passed between instances.
For deployed systems, lineage records should include which models generated synthetic training or fine-tuning data; relevant architectural relationships between source and recipient systems; and, for agentic deployments, which memory files, skills, system instructions and identity artefacts pass between instances.
Do not equate machine accessibility with preservation of source material.
Policy should not treat machine accessibility as equivalent to preservation of source material.
Have AISI monitor autonomous-agent ecosystems as populations and develop methods identifying emergent coordination, cascading error, collusion, manipulation, persistence and influence beyond individual-model evaluation.
AISI should establish ongoing capability to monitor autonomous agent ecosystems as populations and develop methods for identifying emergent coordination, cascading error, collusion, manipulation, persistence and influence behaviours that are not observable through model-level evaluation alone.
Treat frontier-model access as a contested supply chain and plan explicitly for interruption.
Treat frontier-model access as a contested supply chain and plan explicitly for interruption while preserving competition and investing in domestic capability.
Invest in domestic narrow and domain-specific capability while preserving competition and contestability.
Treat frontier-model access as a contested supply chain, while investing in domestic narrow and domain-specific capability and preserving competition and contestability.
For frontier-model dependency, provide independent evaluation, contestability, distributed expertise, public-interest research capacity, transparent intervention thresholds, competition safeguards, and procurement settings that do not make public capability dependent on one provider.
For Australia, that means independent evaluation, contestability, distributed expertise, public- interest research capacity, transparent intervention thresholds, competition safeguards, and procurement settings that do not make public capability dependent on one provider.
Have AISI maintain a standing public capability-signal register with dates, primary sources, its assessment, strongest credible alternative explanations and relevance to Australian safety, procurement, capability or national resilience.
AISI should maintain a standing public register recording significant capability signals, their date and primary source, the institute's assessment, the strongest credible alternative explanation, and identified relevance to Australian safety, procurement, capability or national resilience.
Before Commonwealth service-delivery deployment, publish outcomes, evaluation design, applicable populations/conditions, alternative explanations and human/behavioural measures; commit to publishing a miss log of material failures and groups poorly served.
Before deployment in Commonwealth service delivery, agencies should publish: • the intended outcome and evaluation design; • the population and conditions to which conclusions will apply; • alternative explanations for any measured improvement; • human factors or behavioural effects that will be monitored; and • a commitment to publish a miss log, including material failure modes and groups for whom the system does not perform as intended.
Develop a bioethics framework for biological and hybrid computing before neural-tissue systems scale further commercially.
Develop a bioethics framework for biological and hybrid computing before neural-tissue systems scale further commercially.
Establish governance arrangements through which distributed expertise can enter, be tested, become accountable and change course with evidence.
The governance task is therefore also a commissioning task: establish the architecture through which distributed expertise can enter, be tested, become accountable and change course when the evidence changes.