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 417

Mr Joel Leslie

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.

Joel Leslie proposes AI capability disclosures, consequence-based assurance, intelligible pricing, lifecycle-cost guidance, sector assurance, sovereign continuity and institutional skills. The personal submission and supporting cases argue for safer adoption through existing powers and sector regulation, with light-touch treatment of low-consequence uses. It takes no position on pacing or capability thresholds for frontier development. Its selected cases illustrate governance gaps rather than inherent AI unsafety.

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 governance must also cover what systems may do, how people rely on them, how behaviour changes and who is accountable, not only technology assets.

    AI governance must also govern what systems are permitted to interpret, recommend, decide or do; how people rely on them; how their behaviour changes; and who remains accountable for their effects.

  2. Argues that Australia should establish a proportionate, plain-language AI capability statement or mark for products, services, workplace systems and organisational processes where AI is materially present.

    Australia should establish a proportionate, plain-language AI capability statement or mark for products, services, workplace systems and organisational processes where AI is materially present.

  3. Argues that assurance should be proportionate to consequence, with low-consequence uses kept light-touch.

    Low-consequence use should remain light-touch. Higher-consequence use should demonstrate behaviour under expected, unexpected and adversarial conditions alongside existing privacy, cyber-security, legal, architecture and domain controls.

  4. Argues that consumers and small businesses should receive clear pricing for AI services, including before material plan changes.

    AI services should provide consumers and small businesses with clear, plain-language pricing before purchase and before any material change to a plan or charging model.

  5. Argues that strategically important public functions should not be structurally dependent on a single foreign AI provider or jurisdiction, while allied and commercial models can continue to be used.

    Australia should continue to use allied and commercial models where they provide the best capability. However, strategically important public functions should not be structurally dependent on the access, pricing, retirement, policy or model-behaviour decisions of a single foreign provider or jurisdiction.

Recommendations

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

  1. Define and operationalise material AI capability change, with proportionate disclosure of changes to models, providers, data processing, decision influence, autonomy, tool access, intended use and charging.

    Define and operationalise material AI capability change. Establish proportionate expectations for disclosure when AI materially enters or changes a product, service, workplace or decision process, including changes to model, provider, data processing, decision influence, autonomy, tool access, intended use or charging model.

  2. Establish a proportionate AI capability statement or mark wherever AI is materially present, explaining presence, role, intended use, reliance limits, accountable organisation and material changes before reliance.

    Australia should establish a proportionate, plain-language AI capability statement or mark for products, services, workplace systems and organisational processes where AI is materially present. The statement should enable a consumer, worker, business or public institution to see that AI is involved before relying on it, understand its role and intended use, know what it must not be relied upon for, identify the accountable organisation, and recognise when a material change has occurred.

  3. Require organisations to maintain visibility of AI developed, procured, inherited through software or encountered through shadow use.

    Organisations should maintain visibility of AI they develop, procure, inherit through existing software or encounter through shadow use.

  4. Require supplier disclosure of material AI presence and changes to models, providers, data processing, decision influence, autonomy, tool access, intended use and charging, including software updates; update the statement accordingly.

    Suppliers should disclose material AI presence and material changes to a system’s model or provider, data processing, decision influence, autonomy, tool access, intended use or charging model, including changes introduced through routine software updates. The statement or mark should be updated when these matters materially change.

  5. Give workers intelligible notice when AI performs work or influences decisions about them, including material changes, with consultation where consequences are meaningful.

    Where AI is used to perform work or influence decisions about workers, affected workers should receive intelligible notice of its role and material changes, with consultation where the consequence is meaningful.

  6. Where consequences are meaningful, have the relying organisation proportionately reassess material changes initiated by organisations, suppliers or users.

    Where consequence is meaningful, a material change should trigger proportionate reassessment by the organisation relying on the system, whether the change is initiated by the organisation, supplier or users in practice.

  7. Use existing powers, frameworks and sector regulation where appropriate rather than a stand-alone AI Act; assign disclosure work to PM&C’s Office of AI, Treasury, ACCC, OAIC, Finance, DTA and sector regulators as specified.

    The approach gives practical effect to themes identified by the Senate Select Committee on Adopting Artificial Intelligence in 2024, including transparency, accountability and workplace protections. It should be implemented through existing powers, frameworks and sector regulation where appropriate, rather than through a stand-alone AI Act. Proposed lead capabilities: PM&C Office of AI, Treasury and ACCC for economy-wide settings; OAIC for privacy alignment; Finance and DTA for government procurement and registers; sector regulators for high-consequence AI.

  8. Develop common behavioural-assurance methods reflecting function, consequence, autonomy, human reliance, observability, reversibility and context, with reassessment after material change.

    Develop common behavioural-assurance methods. Require assurance intensity to reflect function, consequence, autonomy, human reliance, observability, reversibility and context, with reassessment after material change.

  9. Keep low-consequence use light-touch; test higher-consequence AI under expected, unexpected and adversarial conditions alongside existing controls, repeating evaluation after material changes to models, prompts, data, integrations, permissions, use or reliance.

    Low-consequence use should remain light-touch. Higher-consequence use should demonstrate behaviour under expected, unexpected and adversarial conditions alongside existing privacy, cyber-security, legal, architecture and domain controls. Evaluation should be repeated when models, prompts, data, integrations, permissions, intended use or actual reliance materially change.

  10. Use structured behavioural evidence of intended boundaries, with testing suited to use, including retrieval fidelity, provenance, uncertainty, repeatability, prohibited behaviours, ambiguous inputs, degradation, silent transformation, over-reliance and agent interactions.

    Behavioural assurance means structured evidence that an AI system remains within intended boundaries. Depending on use, this may include retrieval fidelity, provenance, uncertainty, repeatability, prohibited behaviours, ambiguous inputs, degraded operation, silent transformation, user over-reliance and agent behaviour when using tools or interacting with other systems.

  11. Assign common assurance methods and evidence to AISI and DTA, government requirements to Finance, and domain requirements to sector regulators and assurance bodies.

    Proposed lead capabilities: AISI and DTA for common methods and evidence; Finance for government assurance requirements; sector regulators and assurance bodies for domain requirements.

  12. Require clear AI pricing before purchase and material plan or charging changes, explaining charging models, inclusions, typical use, escalation circumstances, spending controls and notifications.

    AI services should provide consumers and small businesses with clear, plain-language pricing before purchase and before any material change to a plan or charging model. This should explain the charging model, what is included, realistic examples of typical use, circumstances that may increase cost, and available spending controls or notifications.

  13. For metered AI services, give customers enough information to understand likely financial exposure, beyond token, compute, model-call or agent-action unit prices.

    Where services are metered through technical units such as tokens, compute, model calls, or agent actions, those units should not be the only price signal: customers should be given enough information to understand their likely financial exposure.

  14. Have Treasury assess visibility of AI pricing, billing and service quality, establishing ongoing ACCC monitoring where necessary.

    Treasury should assess whether existing arrangements provide sufficient visibility of AI pricing, billing, and service-quality practices and, where necessary, establish an ongoing ACCC monitoring function.

  15. Develop a voluntary AI lifecycle-cost framework supporting rather than determining business decisions; cover intended benefits and process changes, prohibited actions, accountability and integration, data, monitoring, assurance, support, training and redesign costs.

    Government should also develop a practical AI lifecycle-cost framework that organizations can voluntarily adopt when assessing AI investments. The framework should support, rather than determine, organizational business decisions. It should prompt consideration of:  the intended benefit and the process, role or decision being changed;  what the AI must not do and who remains accountable;  integration, data, monitoring, assurance, support, training and process-redesign costs;

  16. Include usage-based charges, compute, tool use, autonomous actions, model/vendor/pricing monitoring, post-deployment benefits and costs, and support, succession and decommissioning in lifecycle-cost considerations.

    usage-based charges, compute, tool use and autonomous actions;  how model, vendor and pricing changes will be monitored;  how benefits and actual costs will be reviewed after deployment; and  how the system will be supported, succeeded or decommissioned.

  17. Apply lifecycle-cost discipline to Commonwealth procurement and business cases; encourage industry and sector bodies to adapt it voluntarily.

    Commonwealth agencies should apply this discipline in government procurement and business cases. Industry and sector bodies should be encouraged to adapt it for wider voluntary use.

  18. Assign pricing and market monitoring to Treasury and ACCC, Commonwealth lifecycle-cost and value-realisation practice to Finance and DTA, and voluntary guidance to industry and sector bodies.

    Proposed lead capabilities: Treasury and ACCC for consumer pricing, market visibility and monitoring; Finance and DTA for Commonwealth lifecycle-cost and value-realisation practice; industry and sector bodies for proportionate voluntary guidance.

  19. Establish a national high-consequence AI assurance model, testing shared evidence, behavioural evaluation, human oversight and reassessment in an initial sector, with healthcare a strong candidate and later adaptation to other sectors.

    Establish a national high-consequence AI assurance model. Select an initial sector, with healthcare as a strong candidate, to develop and test shared evidence requirements, behavioural-evaluation methods, human-oversight expectations and reassessment triggers that can later be adapted for other high-consequence sectors.

  20. Make the model complement existing state, territory and sector frameworks, retaining risk assessment as a foundational input and demonstrating controls under expected, unexpected and adversarial conditions after deployment and material change.

    The model should complement, not replace, existing state, territory and sector risk-management and governance frameworks. Risk assessment should remain a foundational input, identifying potential harms, controls and accountabilities. Existing risk frameworks may identify the relevant risks and controls; the national assurance model would specify the behavioural evidence needed to demonstrate that those controls and intended boundaries hold in practice. Assurance should provide evidence that those controls and intended boundaries hold, including under expected, unexpected and adversarial conditions, and continue to hold after deployment and material change.

  21. Set proportionate national assurance expectations for intended and prohibited use, boundaries, behavioural evidence under varied conditions, relevant source fidelity, provenance, uncertainty, repeatability, consequence-equivalent prohibited behaviours and users’ ability to recognise, challenge and override unsafe outputs.

    A nationally agreed model should establish proportionate expectations for:  declared intended use, prohibited use and functional boundaries;  evidence that the system behaves within those boundaries under expected, unexpected and adversarial conditions;  behavioural evaluation, including source fidelity, provenance, uncertainty, repeatability and consequence- equivalent prohibited behaviours where relevant;  human oversight, including evidence that users can recognise, challenge and override unsafe or wrong outputs;

  22. Provide separated or independent assurance for higher-consequence uses, post-deployment monitoring, incident escalation and learning, and reassessment after material changes to models, data, prompts, integrations, permissions, autonomy, intended use or actual function.

    independent or appropriately separated assurance for higher-consequence uses;  monitoring of performance, drift, reliance and emerging failure modes after deployment;  incident, escalation and organisational-learning pathways; and  reassessment when models, data, prompts, integrations, permissions, autonomy, intended use or actual function materially change.

  23. Retain risk assessments, clinical governance, privacy, cybersecurity, therapeutic regulation and professional obligations; use shared evidence standards without duplicating approval processes.

    Existing risk assessments, clinical governance, privacy, cyber security, therapeutic regulation and professional obligations should remain part of the assurance environment. The objective is not to create a duplicate assessment or approval process. It is to establish shared evidence standards so organisations and regulators can determine whether identified controls work in practice and whether a system continues to operate within its intended boundaries.

  24. Cover medical devices and internally developed, locally configured or workflow-integrated systems outside device regulation; retain local risk management and clinical accountability within a common jurisdictional and sector assurance layer.

    The model should apply not only to regulated medical devices, but also to internally developed, locally configured and workflow-integrated systems that may fall outside medical-device regulation. It should operate as a common assurance layer across existing jurisdictional and sector governance, allowing states, territories and organisations to retain responsibility for local risk management and clinical accountability while reducing the need for every health service to develop its own behavioural evidence model.

  25. If healthcare is selected, consider Health, Disability and Ageing, ACSQHC and TGA as leads with states, territories, Ahpra and professional regulators.

    Proposed lead capabilities: If healthcare is selected as the initial sector, the Department of Health, Disability and Ageing, ACSQHC and TGA could lead the work with states and territories, Ahpra and relevant professional regulators.

  26. First commission costed sovereign-AI options covering Australian-controlled open-weight deployment, domestic adaptation, shared compute and hosting, independent evaluation, portability and substitution, security and continuity, and Australian industry and research participation.

    The Commonwealth should first commission a costed sovereign-AI options assessment covering:  Australian-controlled deployment of suitable open-weight models;  domestic model adaptation;  shared compute and hosting;  independent evaluation;  portability and provider substitution;  security and continuity requirements; and  opportunities for Australian industry and research participation.

  27. Build on GovAI with Australian-controlled civilian capability for government and nationally significant uses, including compute, hosting, secure government data, independent evaluation, portability and at least one adaptable general-purpose language-model capability under Australian authority.

    Building on GovAI, the Commonwealth should establish an Australian-controlled civilian capability for government and nationally significant uses. It should include Australian-controlled compute and hosting, secure government data use, independent evaluation, model portability and at least one general-purpose language-model capability that can be adapted, evaluated and operated under Australian authority.

  28. Continue allied and commercial models where best, while avoiding structural dependence of strategically important public functions on one foreign provider’s or jurisdiction’s access, prices, retirement, policies or model behaviour.

    Australia should continue to use allied and commercial models where they provide the best capability. However, strategically important public functions should not be structurally dependent on the access, pricing, retirement, policy or model-behaviour decisions of a single foreign provider or jurisdiction.

  29. Keep civilian AI capability aligned with separate defence, intelligence, national security and law-enforcement arrangements; consider PM&C’s Office of AI, DISR and Finance as leads, with CSIRO and AISI for research and evaluation.

    This civilian capability should remain aligned with separate arrangements for defence, intelligence, national security and law enforcement. Proposed lead capabilities: PM&C Office of AI, DISR and Finance; CSIRO and AISI for research and evaluation; aligned with Defence, ASD and national-security arrangements.

  30. Build multidisciplinary government, regulator and regulated-sector capability in ownership, engineering, data, evaluation, domain expertise, human factors, privacy, cybersecurity, law, governance, risk, change management, support and succession.

    AI literacy is necessary. Government, regulators and regulated sectors need multidisciplinary capability in:  product and service ownership;  engineering and data;  behavioural evaluation;  domain expertise;  human factors;  privacy and cyber security;  legal interpretation;  governance and risk;  change management; and  operational support and succession.

  31. Recognise AI evaluation as a professional discipline; make AISI a national methods, training and evidence centre, with pathways linking AISI, agencies, regulators, universities and industry.

    Australia should treat AI evaluation as a recognised professional discipline. AISI should serve as a national centre for methods, training and shared evaluation evidence, with pathways between AISI, agencies, regulators, universities and industry.

  32. Build regulators’ ability to evaluate what they accept and organisations’ ability to question vendors, recognise capability change and test boundaries.

    Regulators must develop enough capability to evaluate what they are asked to accept, while organisations must be able to question vendor claims, recognise capability change and test boundaries in practice.

  33. Assign APS capability to APSC and DTA; evaluation methods and training to AISI; regulatory capability to eSafety, OAIC, ACCC and sector regulators; and workforce measures to DISR.

    Proposed lead capabilities: APSC and DTA for APS capability; AISI for evaluation methods and training; eSafety, OAIC, ACCC and sector regulators for regulatory capability; DISR for workforce measures.

  34. Include training on when not to use AI, and give professionals standing to decline premature adoption.

    Training must therefore include not only how to use AI but when not to use it, and organisations must provide the professional standing to decline premature adoption.

  35. Prepare a consultation draft and pilot material-change disclosures and a common AI capability statement across government and selected sectors.

    1. Material AI change Define material AI capability change Consultation draft and pilot across Material changes are identified, made visible and draft a common AI capability government and selected sectors. disclosed and linked to reassessment (d), (g), (h), (i), (k), (n) statement. where warranted.

  36. Develop common evaluation vocabulary and evidence templates for assistive, advisory and agentic AI, adaptable by sector.

    2. Consequence-based Develop a common evaluation Evaluation methods and evidence Organisations can demonstrate behavioural assurance vocabulary and evidence template for templates adaptable by sector. relevant behavioural properties rather (h), (i), (m) assistive, advisory and agentic use. than rely only on vendor claims.

  37. Develop consumer AI cost-summary options, assess Part VIIA price monitoring and issue Commonwealth lifecycle-cost guidance, with pricing consultation, monitoring advice and revised procurement guidance.

    3. AI pricing and Develop consumer AI cost-summary Consumer-pricing consultation, Consumers can understand likely cost lifecycle-cost discipline options; assess the case for Part VIIA market-monitoring advice and exposure; government AI business (a), (d), (i), (k) price monitoring; and issue revised Commonwealth cases identify lifecycle cost and Commonwealth lifecycle-cost guidance procurement guidance. value-realisation assumptions.

  38. If healthcare is selected, establish a Commonwealth–state and territory working group on shared behavioural evidence, evaluation and reassessment, leading to a national assurance model and healthcare pilot.

    4. High-consequence If healthcare is selected, establish a National high-consequence AI Shared evidence methods are used sector assurance Commonwealth–state and territory assurance model, with a healthcare across participating services, with (c), (i), (n) working group to define shared pilot. clear escalation, reassessment and behavioural-evidence requirements, monitoring pathways. evaluation methods and reassessment triggers for high-consequence AI.

  39. Identify strategically important public functions and develop a sovereign capability roadmap, funding options and continuity criteria, providing an independently operable pathway for identified functions.

    5. Sovereign continuity Commission a costed options Capability roadmap, funding At least one independently operable for strategic functions assessment and identify strategically options and continuity criteria. pathway exists for identified strategic (a), (e), (f), (l), (m) important public functions.

  40. Establish an AI evaluation and operations capability program, including a skills framework, training pathways, shared test environments and a regulator capability plan.

    6. National AI Establish an AI evaluation and Skills framework, training Government and regulators can evaluation and operations capability program. pathways, shared test independently assess, operate and institutional capability environments and regulator challenge AI systems. (g), (h), (k), (m) capability plan.

  41. Stage implementation proportionately, keeping low-consequence AI light-touch and triggering reassessment when capability, authority, reliance, data use, cost or dependency materially changes.

    Implementation should be staged and proportionate. The framework should not make every AI feature a high-risk system. Instead, it should create a clear trigger when capability, authority, reliance, data use, cost or dependency materially changes.

Read the original

Open in the interactive guide