Submission 99
Chameleon Confidential Solutions Pty Ltd
Submitter group: Companies.
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 for strengthening existing AI governance through functional accountability, lifecycle assurance, regulatory coordination, context-sensitive proportionality and strategic dependency assessment. Recommends testing existing mechanisms before creating new structures. A supplement proposes examining just-in-time decision authority for high-consequence uses, with human authority retained and no continuous reauthorisation of routine decisions.
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 accountability is functional only where responsible actors have the capability, information and authority to act.
Accountability, however, is effective only where the responsible actor has sufficient information, authority and access to relevant expertise to understand material risks and act upon them.3
Argues that assurance should combine a trusted baseline with dynamic detection of material change, followed by proportionate reassessment.
Lifecycle assurance should consequently combine a trusted baseline with dynamic detection of relevant change. Detection should not automatically require complete reassessment. Rather, identified change should be assessed for materiality, with proportionate reassessment undertaken where the evidence supporting continued reliance may no longer remain valid.
Argues that assurance of individual components should not automatically be treated as assurance of an assembled AI system.
This submission therefore proposes that component assurance should not automatically be treated as establishing system assurance.
Argues that governance proportionality should reflect consequence and context rather than organisation size.
Organisational size is an imperfect proxy for technological consequence. A large organisation may use AI for a relatively low-consequence administrative function, while a small health provider, agricultural business or specialist service may deploy technology capable of affecting safety, privacy, livelihoods or access to essential services.
Argues that sovereign AI capability cannot be assessed solely by where technology or infrastructure is located, or by whether a supplier is Australian or foreign owned.
Australia’s sovereign AI capability cannot be assessed solely by where technology or infrastructure is physically located, or by whether a supplier is Australian or foreign owned.
Proposes just-in-time decision authority that responds proportionately to changes affecting authority, without renewed approval for every decision.
JIT-DA should not require renewed approval for every decision. It should operate proportionately where an authority-relevant threshold or material change is capable of affecting the validity, scope or level of decision authority required.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Recommendation 1: examine proportionate evidence of continuing competence, information, authority and escalation access for high-consequence AI accountability; certification/qualifications cannot replace demonstrated capability, and distributed capability must retain clearly allocated accountability.
Build on existing accountability frameworks by examining whether high-consequence AI uses require clearer, proportionate evidence that accountable roles have continuing access to the competence, information, authority and escalation pathways necessary to identify and act upon material risk. Certification or qualification may contribute to that evidence but should not substitute for demonstrated operational capability. Capability may be distributed, but accountability should remain clearly allocated.
Recommendation 2: establish proportionate high-consequence AI material-change triggers, potentially including technology, functionality, purpose, data, autonomy, regulation, suppliers or dependencies; assess materiality first and reassess proportionately to effects on evidence supporting continued reliance.
Further develop lifecycle-assurance principles by establishing consistent and proportionate material-change triggers for high-consequence AI. Relevant changes could include technology, functionality, purpose, data use, autonomy, regulation, suppliers or critical dependencies. Detection of change should lead first to an assessment of materiality, with reassessment proportionate to its potential effect on the evidence supporting continued reliance.
Recommendation 4: examine existing Australian arrangements for independent, regulator-aligned assurance of defined high-consequence AI, considering assessor competence, independence, methodology and continuing quality as technology, dependencies and threats change.
Examine whether existing Australian accreditation, conformity-assessment and professional assurance arrangements can support independent, regulator-aligned assurance for defined high-consequence AI uses. This should include consideration of assessor competence, independence, methodology and continuing assurance quality as technologies, dependencies and threat environments change.
Recommendation 4: complement lifecycle monitoring with independent assurance while preserving organisational accountability and regulators’ inspection, challenge and enforcement powers.
Independent assurance should complement lifecycle monitoring while preserving organisational accountability and regulatory powers to inspect, challenge and enforce.
Recommendation 3: test and strengthen significant-AI-risk pathways for signal authentication, interpretation, lawful sharing and escalation to actors with decision and intervention authority.
Test and strengthen existing cross-regulatory pathways for significant AI risks to determine whether emerging signals can be authenticated, interpreted, lawfully shared, escalated and transferred to actors with clear decision and intervention authority.
Recommendation 3: exercise compromised-supplier, credential and shared-service scenarios for containment, continuity and provenance; consider new coordinating structures only where demonstrated gaps cannot reasonably be addressed through existing mechanisms.
Cross-agency exercises should include scenarios involving compromised suppliers, credentials or shared services and test whether containment, continuity and provenance can be maintained. New coordinating structures should be considered only where demonstrated gaps cannot reasonably be addressed through existing mechanisms.
Recommendation 5: develop proportionate AI governance using factors such as function, autonomy, control, consequence, dependency, affected people and communities, cumulative effects and operating environment.
Further develop proportionate AI governance using objective factors that could include intended function, autonomy, control, consequence, dependency, people and communities affected, cumulative effects and operating environment.
Recommendation 5: enable shared, federated or externally accessed capability where it improves access to specialist expertise without reducing protection, accountability, information security or resilience; organisational size alone should not determine the level of protection.
Shared, federated or externally accessed capability should be enabled where it improves access to specialist expertise without reducing protection, accountability, information security or resilience. Organisational size alone should not determine the level of protection appropriate to an AI use.
In regional AI governance, recognise connectivity, infrastructure constraints, specialist access and physical consequences of technology failure.
In regional environments, governance arrangements should also recognise practical conditions such as connectivity, infrastructure constraints, access to specialists and the physical consequences of technology failure.11
Recommendation 6: build on existing frameworks to assess strategic AI dependencies across government and critical sectors: control, jurisdiction, concentration, privileged access, substitutability, continuity, domestic capability and denial, compromise or supplier-change consequences.
Build on existing technology-security and supplier-assurance frameworks by establishing a consistent, risk-based approach to strategically significant AI dependencies across government and critical sectors. Assessment should consider effective control, jurisdiction, concentration, privileged access, substitutability, continuity, domestic capability and consequences of denial, compromise or material supplier change.
Recommendation 6: consider stronger sovereign capability or assured domestic access where external dependency could materially impair critical national functions.
Stronger sovereign capability or assured domestic access should be considered where external dependency could materially impair critical national functions.
In strategic AI planning, consider supporting energy, water, telecommunications, land, supply chains and communities alongside digital and cyber requirements.
Sovereign capability also has a physical dimension. AI and digital infrastructure depend upon energy, water, telecommunications, land, supply chains and communities. Strategic planning should therefore consider the resilience and consequences of these supporting systems alongside digital and cybersecurity requirements.15
Supplement: examine JIT-DA for defined high-consequence AI, testing human authority’s continuing validity where predefined conditions or authority-relevant material change indicate an existing delegation or authority may no longer suffice.
Examine whether Just-in-Time Decision Authority (JIT-DA) could strengthen existing governance arrangements for defined high-consequence AI uses by enabling the continuing validity of human decision authority to be tested where predefined conditions or authority- relevant material change indicate that an existing delegation or authority may no longer be sufficient.
Supplement: examine building on existing identity, access-management, delegation, assurance and human-oversight mechanisms, including clear, proportionate and, where appropriate, machine-readable authority conditions.
Any examination should consider how JIT-DA could build upon existing identity, access- management, delegation, assurance and human-oversight mechanisms rather than duplicate them, including whether decision authority can be represented through sufficiently clear, proportionate and, where appropriate, machine-readable conditions.
Supplement: test when human authority should be confirmed, constrained, elevated, suspended or escalated before consequential action proceeds.
Testing should consider circumstances in which authority may need to be confirmed, constrained, elevated, suspended or escalated before consequential action proceeds, together with safeguards for human accountability, explainability, override, emergency authority, auditability, resilience and the prevention of accountability laundering.
Supplementary recommendation: focus initial evaluation of JIT-DA on defined high-consequence use cases where AI capability, evidence, operating conditions or dependencies may change materially before authority is exercised.
Initial evaluation should focus on defined high-consequence use cases where AI capability, supporting evidence, operating conditions or dependencies may change materially between the establishment of decision authority and the point at which that authority is exercised.
Supplement: evaluate minimum sufficient friction where material change threatens valid human authority, without adding approval layers to routine decisions.
The objective should not be to introduce additional approval layers into routine decision- making. It should be to determine whether minimum sufficient friction can preserve valid and accountable human decision authority where material change would otherwise create a gap between formal authority and the conditions under which consequential decisions are made.
In examining JIT-DA safeguards, retain human organisational governance over authority, delegation and escalation; automated infrastructure may apply rules but must not independently create, expand or permanently reallocate human decision authority.
Human authority should remain the source of legitimate authority. Organisational governance should determine who holds authority, what may be delegated and when escalation is required. AI or automated infrastructure may recognise and apply those rules but should not independently create, expand or permanently reallocate human decision authority.
Consider defined, proportionate JIT-DA triggers for authority-relevant conditions, with materiality thresholds reflecting consequence, context and effects on existing authority.
Triggers should be defined and proportionate. JIT-DA should respond to authority- relevant conditions rather than every operational variation. Materiality thresholds should reflect consequence, context and the potential effect upon existing authority.
Consider explainable, reviewable authority decisions, recording the authority, delegation, material trigger, intervention and subsequent human decision where appropriate.
Decision pathways should remain explainable and attributable. Where authority is confirmed, constrained, elevated, suspended or escalated, the basis for that outcome should be capable of being understood and reviewed. Records should identify the relevant authority, applicable delegation, material trigger, resulting intervention and subsequent human decision where appropriate.6
Consider authority-, consequence- and purpose-bounded human override, recorded reasons and subsequent review for higher-consequence decisions; end emergency authority when its justification expires.
Human override should exist within defined boundaries. Exceptional circumstances may require authorised actors to proceed despite an automated constraint or escalation. Override should therefore be bounded by authority, consequence and purpose, with reasons recorded and subsequent review available for higher-consequence decisions. Emergency authority should expire when the conditions justifying it no longer apply.
Consider authority-validation resilience, avoiding critical dependency and addressing availability and fallback arrangements.
The authority-validation mechanism should itself be resilient. JIT-DA should not create a new critical dependency capable of unnecessarily preventing legitimate action. Organisations should consider system availability, fallback arrangements,
Consider compromised or erroneous authority validation and validation information that may be incomplete, manipulated or unavailable.
compromise, erroneous triggers and the possibility that information used to validate authority may itself be incomplete, manipulated or unavailable.7
Prevent JIT-DA validation from establishing that a decision is correct, reasonable or lawful, or transferring consequential-decision responsibility to the validation system.
JIT-DA should not become a mechanism for accountability laundering. Automated validation of authority would provide evidence that defined governance conditions were satisfied at a particular point in time. It would not establish that the underlying decision was correct, reasonable or lawful, nor should it allow organisations to transfer responsibility for consequential decisions to the system performing the validation.