Submission 381
Mr Dorian Thom
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 regulation is necessary but insufficient for advanced-AI safety and proposes researching intrinsic, measurable human-protective properties. Says whether these goals can be fully encoded remains an open technical question. Recommends capability-proportionate safeguards, defence in depth, AI Safety Institute research and international cooperation, recognising open-weight modification and the limits of unilateral requirements.
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.
Proposes that the safety of advanced AI should not depend solely upon laws regulating its use or instructions requiring it to behave safely.
This submission proposes a simple principle: the safety of advanced artificial intelligence should not depend solely upon laws regulating its use or instructions requiring it to behave safely.
Proposes benevolence, wisdom, humility and human control as governance objectives for advanced AI, rather than human emotions or personalities.
that the development of advanced AI should therefore pursue four linked objectives: benevolence, wisdom, humility and human control. These terms describe governance objectives, not human emotions or personalities.
A program corrected obvious text-recognition errors in this quote for display.
Argues that autonomous authority should fall, and verification and human oversight rise, as potential harm, irreversibility and uncertainty grow.
As potential harm, irreversibility and uncertainty increase, autonomous authority should decrease and the requirement for independent verification and meaningful human oversight should increase.
Argues that layered safeguards are needed so that failure of one does not cause catastrophic failure, and that the whole system should be tested.
Safety assurance should therefore test the system as a whole, including interactions between layers, rather than assuming that several individually effective safeguards are automatically independent. The important principle is that failure of one safeguard should not automatically permit catastrophic failure of the entire system.
Argues that safety evaluation should use diverse, adaptive and independent testing, because systems may perform well on known tests without predicting behaviour in novel conditions.
A system may perform satisfactorily against known tests without those results reliably predicting its behaviour under novel or out-of-distribution conditions. Safety evaluation should therefore use diverse, adaptive and independent testing rather than relying on fixed benchmarks alone.
Says rigid unilateral requirements could shift development elsewhere and open-weight models can be modified beyond developer control, so intrinsic safety cannot replace enforcement and coordination.
Excessively rigid unilateral requirements could shift development to jurisdictions with weaker safeguards, while open-weight models can be copied, modified and redistributed beyond the control of their original developer. Intrinsic safety properties should therefore be researched for robustness against fine-tuning and modification, but they cannot replace cybersecurity, access controls, monitoring, enforcement or international coordination.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Advocate internationally for research, testing and standards aimed at making human-protective behaviour increasingly intrinsic, measurable and resistant to circumvention as AI capability grows.
Rather, Australia should advocate internationally for research, testing and standards aimed at making human-protective behaviour increasingly intrinsic, measurable and resistant to circumvention as AI capability grows.
A program corrected obvious text-recognition errors in this quote for display.
Support international cooperation towards adaptive and measurable safety standards for frontier AI.
Australia should support international cooperation towards adaptive and measurable safety standards for frontier AI.
A program corrected obvious text-recognition errors in this quote for display.
Consider AISI contributions to research and international coordination on human-protective behaviour, corrigibility, override, uncertainty, autonomy, deception, manipulation, unauthorised power/resources, modification robustness, adaptive evaluation, independent pre-deployment evaluation above specified capability thresholds, and increasingly rigorous international safety standards.
The AISI could contribute to research and international coordination concerning: technical methods for embedding and evaluating human-protective behaviour in advanced AI; corrigibility and reliable human override; AI behaviour under uncertainty and preservation of meaningful human autonomy; resistance to deception, manipulation and unauthorised acquisition of power or resources; robustness of safety properties under fine-tuning, modification and open-weight release; adaptive evaluation designed to reduce benchmark gaming and unintended interpretations of safety objectives; independent evaluation before deployment of systems exceeding specified capability thresholds; and international standards under which increasingly powerful systems face increasingly rigorous safety requirements.
A program corrected obvious text-recognition errors in this quote for display.
Adopt capability-proportionate AI safety: as autonomy, capability, irreversibility and potential consequences increase, increase developers' safety-demonstration burden, testing, independent evaluation, human oversight and technical safeguards.
Adopt a capability-proportionate safety principle. As autonomy, capability, irreversibility and potential consequences increase, developers seeking greater autonomous authority for AI systems should carry a correspondingly greater burden of demonstrating safety. Requirements for testing, independent evaluation, human oversight and technical safeguards should increase accordingly.
A program corrected obvious text-recognition errors in this quote for display.
Have AISI validate intrinsic advanced-AI safety properties with international counterparts, avoiding duplicated foundational research: calibrated uncertainty, corrigibility, restraint, resistance to deception and power-seeking, human autonomy, and robustness under modification or fine-tuning.
Ask the Australian AI Safety Institute to investigate and validate intrinsic and technically measurable safety properties for advanced AI, working collaboratively with international counterpart institutions rather than duplicating foundational research already undertaken elsewhere. This should include calibrated uncertainty, corrigibility, restraint under uncertainty, resistance to deception and power-seeking, respect for human autonomy, and robustness of those properties under modification or fine-tuning.
A program corrected obvious text-recognition errors in this quote for display.
Support international research into advanced-AI controllability and adaptive evaluation, working with international AI safety institutions and standards bodies on diverse tests for loss of control, deception, circumvention of safeguards, unauthorised power-seeking, resistance to correction and benchmark gaming.
Support international research into advanced-AI controllability and adaptive evaluation. Australia should work with international AI safety institutions and standards bodies on diverse tests for loss of control, deception, circumvention of safeguards, unauthorised power-seeking, resistance to correction and benchmark gaming.
A program corrected obvious text-recognition errors in this quote for display.
Require defence in depth for highly consequential AI systems, with whole-system testing of interactions and common failure modes; for high-consequence applications, independent testing of the combined safety architecture.
Require defence in depth for highly consequential AI systems, together with whole-system testing of interactions and common failure modes between safeguards. No single alignment technique, behavioural instruction, human supervisor or shutdown mechanism should be regarded as sufficient protection against catastrophic failure. For high-consequence applications, independent testing should assess the combined safety architecture rather than individual safeguards only in isolation.
A program corrected obvious text-recognition errors in this quote for display.
Ensure Australia's approach explicitly addresses open-weight and cross-border AI, complementing domestic safeguards with international standards, secure deployment practices and research into safety properties that are difficult to remove through modification.
Ensure Australia’s approach explicitly addresses open-weight and cross-border AI. Domestic safeguards should be complemented by international standards, secure deployment practices and research into safety properties that are difficult to remove through subsequent modification.
A program corrected obvious text-recognition errors in this quote for display.
Consider advocating that advanced AI benefit humanity and, as far as technically achievable, preserve life, wellbeing, dignity and autonomy, remain corrigible under meaningful human control, and increase restraint with uncertainty and potential consequences.
suggest that Australia consider advocating a guiding principle for advanced artificial intelligence along the following lines: Advanced artificial intelligence should be developed and deployed for the benefit of humanity and should, as far as technically achievable, preserve human life and wellbeing, respect human dignity and autonomy, remain corrigible and subject to meaningful human control, and exercise increasing restraint where uncertainty and the potential consequences of error are greatest.
Read the original
In some quotes a program corrected obvious text-recognition errors for display.