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 152

Dr Tony Carden

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.

Open in the interactive guide

Summary

Read the original submission for the full text.

Argues for a dynamic, integrated AI safety regulator built on responsive, smart and agile regulatory principles and informed by sociotechnical systems thinking. Describes three tiers of AI risk (current, emerging and prospective, including AGI and ASI) and proposes legislation setting powers and principles, with detailed rules left to adaptable codes and standards.

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. Argues that AI risks fall into three tiers (current, emerging and prospective) and that an effective regulatory system must address all three simultaneously.

    Comprehensive: an effective AI regulatory system must perpetually and simultaneously address all three tiers of AI risk – current, emerging and prospective.

  2. Argues that a reactive approach, relied on for tier 3 risks, will almost certainly result in catastrophic outcomes.

    A reactive approach, as seen with tier 1 and 2 risks, will almost certainly result in catastrophic outcomes if relied upon to mitigate tier 3 risks.

  3. Says a different regulatory approach is required, and that regulation used previously cannot simply be adopted.

    This requires a different regulatory approach and critically means we cannot just adopt regulation that has been used previously.

  4. Says empowering legislation should establish powers, limits and principles for regulation but refrain from detailed rule-making.

    Empowering legislation should, therefore, establish powers, limits and principles for regulation but refrain from any detailed rule-making.

  5. Says an AI regulator can leverage, educate and support existing regulators to cover AI safety within their domains.

    An AI regulator can leverage, educate and support existing regulators to encompass AI safety within their domains.

  6. Says the barrier to entry for AI experimentation, in knowledge and finances, is very low compared with many other risky technologies, and that this in itself increases risk.

    The bar for entry to this kind of experimentation in terms of both knowledge and finances is very low compared with many other risky technologies. This low bar in itself increases risk.

Recommendations

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

  1. Pay close attention to precursors of AGI and ASI to mitigate their high-end risks as far as reasonably practicable.

    close attention must be paid to precursors of AGI and ASI to mitigate in advance their high-end risks as far as reasonably practicable.

  2. Adapt existing regulatory regimes to account for the introduction of AI into the domains they regulate.

    In fact, existing regulatory regimes will need to adapt to account for the introduction of AI into the domains they seek to regulate.

  3. Use a less prescriptive, more flexible, principles-based and agile approach to managing and regulating emerging (tier 2) risks.

    The consequences and likelihood of tier 2 risks are inherently less known as they are at any given time, currently emerging. Therefore, a less prescriptive, more flexible, principles-based and agile approach to both the management and regulation of these risks is required.

  4. Urgently develop and implement mechanisms for managing tier 3 (AGI and ASI) risks before such systems are realised, rather than relying on a reactive approach.

    This necessitates the urgent development and implementation of mechanisms for managing these risks before AGI and ASI are realised.

  5. Guide decisions on managing risks whose probability is unknowable by the expected severity of consequence.

    Decisions about how to manage and regulate risks for which the probability of manifestation is unknowable should be guided by the expected severity of consequence.

  6. Apply complexity theory and sociotechnical systems principles in the design and operation of AI regulatory systems.

    Effective regulation of AI must recognise and account for the formal complexity of the sociotechnical systems of which AI is and will increasingly be a part. This will require the application of complexity theory and sociotechnical systems principles in the design and operation of AI regulatory systems.

  7. Use multiple risk-control layers with intentional redundancy when risks are high in complex systems.

    When risks are high in complex systems, robust risk mitigation requires multiple layers of risk control with intentional redundancy, to cover the inevitable occasional failure of some controls.

  8. Have the regulator encourage and support innovation in AI risk control.

    Given the quickly evolving state of AI and the uncertainty about its effects, the regulator should encourage and support innovation in AI risk control.

  9. Establish an integrated AI risk regulator able to promote, monitor and assure controls across all three tiers of AI risk.

    An integrated AI risk regulator must be able to promote, monitor and assure the best available controls for all three tiers of risk.

  10. Legislate the authority, resourcing and principles of the regulatory system, leaving rule-making detail to compliance codes and standards that can be rapidly adapted.

    the authority, resourcing and principles of the regulatory system should be legislated, the detail of rulemaking should occur further down the system and be articulated through compliance codes and standards that can be rapidly adapted as conditions change,

  11. Have AI safety regulation promote, monitor and assure a range of risk controls across the system, engaging developers, suppliers, modifiers and users.

    Similarly, AI safety requires risk controls to be applied across the system in which AI operates, engaging multiple classes of duty holder such as developers, suppliers, modifiers, and users.

  12. Make the design and operation of each AI regulatory system responsive, smart and agile, and cover all three tiers of AI risk: current, emerging and prospective.

    what should not vary is that the design and operation of each AI regulatory system should be responsive, smart and agile, and should cover all three tiers of AI risk: 1. Known risks associated with established AI systems (current risks) 2. Likely risks associated with emerging AI systems (emerging risks) 3. Possible risks from future AI systems (prospective risks)

  13. Have empowering legislation establish powers, limits and principles for regulation but refrain from any detailed rule-making.

    Empowering legislation should, therefore, establish powers, limits and principles for regulation but refrain from any detailed rule-making.

  14. Build continuous, broad consultative processes into rule-making and standard-setting.

    For rule-making and standard-setting, this means continuous, broad consultative processes must be built into the foundations of the regulatory structure.

  15. Include AI, regulatory and legal experts and wider academia and society, including philosophers, educators, doctors and community interest groups, in regulator governance and rule-making.

    Both governance and rule-making functions of the regulator must include experts from the fields of AI, regulation and law, but also from across academia and society including philosophers, educators, doctors, and community interest groups.

  16. For jurisdictions adopting this approach, support operational functions across all three risk tiers and 15 focus areas; establish compliance processes preferring duty-holders, then third parties, then regulator action.

    A jurisdiction adopting this approach to AI regulation will need to outline how it will support each of its operational functions in relation to current, emerging and prospective AI risks for each of the 15 focus areas in the smart regulatory pyramid. This will require specifying, establishing and maintaining processes and actions that can enable, motivate, influence, improve and enforce compliance through the agency of (in this order of preference) duty-holders, third parties, and the regulator itself.

  17. Give any central AI regulatory agency integrated governance, management and operational functions to monitor and assess existing and emerging risk continuously, choose the optimal tool from the responsive and smart pyramids and apply them with agility.

    Any central AI regulatory agency should include governance, management and operational functions that integrate to continuously monitor and assess existing and emerging risk, dynamically choose the optimal tool from the responsive/smart pyramids and apply them with agility.

Read the original

Open in the interactive guide