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 561

Mr Jason Bennett

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.

Recommends independently evaluating whether existing Australian AI assurance practices demonstrate continuing validity after systems or conditions change, with targeted improvements only where justified by findings. Says existing frameworks already address ongoing oversight. Sets out evaluation methods and criteria, seven recommendations, and supplementary guidance covering traceable change, evidence-based reassessment, accountable decisions and practical testing.

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. Asks how Australia can demonstrate that AI assurance conclusions remain justified when their supporting conditions change.

    How can Australia demonstrate that AI assurance conclusions remain justified when the conditions supporting those conclusions change?

  2. States that existing frameworks already address risk management, testing, accountability and ongoing oversight, and that the submission does not presume they lack these provisions.

    These frameworks already address risk management, testing, accountability and ongoing oversight. This submission does not presume that they lack these provisions.

  3. Argues that Australia should evaluate not only whether AI assurance has been performed but whether its conclusions remain justified when conditions change.

    Australia should be able to evaluate not only whether AI assurance has been performed, but whether its conclusions remain justified when conditions change.

  4. Says agencies should explain AI's role in consequential decisions, identify responsible decision-makers and substantiate conclusions.

    Where AI contributes to consequential decisions, agencies should be able to explain its role, identify responsible decision-makers and substantiate conclusions.

  5. Argues that scrutiny should reflect the potential consequences of incorrect conclusions, with lighter controls for low-risk tools and deeper review for essential services.

    Scrutiny should reflect the potential consequences of incorrect conclusions, including cascading impacts. For example, a low-risk internal productivity tool may require only change logging and assumption traceability. A system supporting essential services may warrant independent reconstruction of the reasoning following a material change.

  6. States that a potential concern does not by itself establish a regulatory gap, and that the evaluation should distinguish established requirements from their effectiveness in practice.

    The evaluation should distinguish requirements already established in policy from their effectiveness in practice. A potential concern does not, by itself, establish a regulatory gap.

Recommendations

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

  1. Recommend an independent, proportionate evaluation of existing Australian AI assurance practices after relevant changes, building on existing frameworks and recommending improvements only where demonstrable weaknesses justify them.

    The Committee should recommend an independent, proportionate evaluation of existing Australian AI assurance practices to determine whether they adequately demonstrate the continuing validity of material assurance conclusions following relevant changes. The evaluation should build on existing frameworks, identify demonstrable weaknesses and recommend improvements only where justified by findings.

  2. Evaluate whether initial conclusions were adequately supported.

    Original assessment: Was the initial conclusion adequately supported?

  3. Evaluate identification of material, cumulative and assumption-invalidating changes, including those generating no obvious alert.

    Change identification: Can material, cumulative and assumption-invalidating changes be identified, including changes that do not generate an obvious alert?

  4. Evaluate whether conclusions can be traced to supporting evidence, criteria and assumptions.

    Evidence traceability: Can conclusions be traced to supporting evidence, criteria and assumptions?

  5. Evaluate which earlier findings remain supported, require revision or cannot be resolved.

    Impact assessment: Which earlier findings remain supported, require revision or cannot be resolved?

  6. Evaluate whether alternative interpretations, conflicting evidence and uncertainty were adequately considered.

    Whether alternative interpretations, conflicting evidence and uncertainty were adequately considered.

  7. Evaluate whether reviewers distinguish established findings from estimates, assumptions and unresolved questions.

    Uncertainty: Can reviewers distinguish established findings from estimates, assumptions and unresolved questions?

  8. Evaluate whether AI-system changes can be distinguished from environmental, data, human-intervention or assessment-method changes.

    Attribution: Can changes attributable to the AI system be distinguished from those arising from the environment, data, human interventions or assessment method?

  9. Evaluate whether reassessment responsibility is clearly assigned.

    Accountability: Is responsibility for reassessment clearly assigned?

  10. Evaluate whether resulting decisions were attributable to appropriately authorised decision-makers.

    Whether resulting decisions were attributable to appropriately authorised decision-makers.

  11. Evaluate whether a qualified independent reviewer can reconstruct and challenge a decision without relying solely on its original decision-maker.

    Independent review: Can a qualified reviewer reconstruct and challenge the decision without relying solely on the original decision-maker?

  12. Evaluate whether corrective action is appropriate and effective when a conclusion is no longer justified.

    Corrective action: Where a conclusion is no longer justified, is the response appropriate and effective?

  13. Test assurance conclusions against credible counterexamples.

    The evaluation should test conclusions against credible counterexamples

  14. Evaluate whether new evidence connects to earlier findings, uncertainty is represented appropriately and interventions reduce risk rather than merely being completed.

    The evaluation should examine whether new evidence is connected to earlier findings, whether uncertainty is appropriately represented and whether interventions reduce risk — rather than merely confirming that an action was completed.

  15. Make scrutiny proportionate to potential consequences, including cascading impacts.

    Scrutiny should reflect the potential consequences of incorrect conclusions, including cascading impacts.

  16. Consider small-business implementation costs and avoid unnecessary duplication of requirements in the evaluation.

    The evaluation should consider implementation costs for small businesses and avoid unnecessary duplication of existing requirements.

  17. Focus on evidence and appropriate review access rather than public disclosure of sensitive methods or commercially confidential information.

    The focus should be on whether adequate evidence and appropriate review access exist, rather than requiring public disclosure of sensitive methods or commercially confidential information.

  18. Evaluate representative government and other relevant applications using controlled scenarios: model/component changes, changed use/conditions/populations, new behaviour/error/harm evidence, changed data/dependencies, challenged assumptions and changed assessment methods/decision criteria.

    The evaluation should examine representative applications across government and other relevant settings. Controlled scenarios should include: ● A model update or significant component change. ● A change in intended use, operating conditions or affected population. ● New evidence of unexpected behaviour, error or harm. ● A change in data source or external dependency. ● Evidence that challenges an underlying assumption. ● A change in assessment method or decision criteria.

  19. Test cumulative changes and interactions, including individually insignificant changes whose combined effect may invalidate an earlier conclusion.

    Testing should also consider cumulative changes and interactions between multiple changes, including circumstances in which no single change appears significant but their combined effect may invalidate an earlier conclusion.

  20. Set evaluation criteria for whether conclusions remain logically defensible as evidence and conditions change.

    Reasoning integrity: Whether conclusions remain logically defensible as evidence and conditions change.

  21. Set evaluation criteria distinguishing system history from assessment-method history.

    Historical integrity: Whether system history can be distinguished from the history of the assessment method.

  22. Set evaluation criteria for whether assurance findings remain appropriately supported when transferred to different contexts.

    Transferability: Whether assurance findings remain appropriately supported when applied in different contexts.

  23. Define evidence requirements and decision rules for satisfactory, unsatisfactory and inconclusive outcomes before testing.

    Evidence requirements and decision rules for satisfactory, unsatisfactory and inconclusive outcomes should be defined before testing begins.

  24. Keep evaluation independent of exclusive reliance on organisations' own accounts; protect sensitive information and commercial confidentiality while scrutinising outcomes, without prescribing one methodology or unnecessarily duplicating frameworks.

    The evaluation should be independent and should not rely exclusively on an organisation’s own account of its assurance performance. It should protect sensitive information and commercial confidentiality while enabling meaningful scrutiny of observable outcomes. It should evaluate effectiveness without prescribing a single methodology or unnecessarily duplicating existing frameworks.

  25. Have AI assurance arrangements assess whether material assurance conclusions remain justified following changes to systems, evidence, assumptions or operating conditions.

    Examine continuing validity: Assess whether material assurance conclusions remain justified following relevant changes to systems, evidence, assumptions or operating conditions.

  26. Have assurance arrangements distinguish policy design from practical capability, implementation and access to supporting evidence.

    Assess implementation effectiveness: Distinguish policy design from practical capability, implementation and access to supporting evidence.

  27. Have assurance arrangements evaluate whether conclusions are justified, rather than relying solely on evidence that prescribed procedures were completed.

    Test substantive conclusions: Evaluate whether conclusions are justified, rather than relying solely on evidence that prescribed procedures were completed.

  28. Have assurance arrangements determine whether system changes, environmental changes and changes to assessment methods can be distinguished and evaluated appropriately.

    Examine attribution and methodological integrity: Determine whether system changes, environmental changes and changes to assessment methods can be distinguished and evaluated appropriately.

  29. Recommend targeted improvements where demonstrable deficiencies are identified.

    Strengthen proportionate accountability: Recommend targeted improvements where demonstrable deficiencies are identified.

  30. Build on existing frameworks, avoid unnecessary duplication and protect legitimate commercial and security interests.

    Preserve innovation and confidentiality: Build on existing frameworks, avoid unnecessary duplication and protect legitimate commercial and security interests.

  31. Examine periodic or change-triggered reviews for higher-consequence applications to establish whether assurance conclusions remain justified over time.

    Consider ongoing evaluation: Examine periodic or change-triggered reviews for higher-consequence applications to establish whether assurance conclusions remain justified over time.

  32. Develop and evaluate practical ongoing-assurance guidance, building on existing policy and technical standards.

    Accordingly, the Committee should recommend that the Australian Government develop and evaluate practical guidance for demonstrating ongoing AI assurance, building on existing policy and technical standards.

  33. Have ongoing-assurance guidance address recording material changes and emerging evidence so their implications for existing assessments, safeguards and decisions can be identified.

    Traceable change: Material changes and emerging evidence should be recorded in a way that enables their implications for existing assessments, safeguards and decisions to be identified.

  34. Have ongoing-assurance guidance address organisations' ability to explain, with supporting evidence, whether previous conclusions remain valid, need qualification or must be revised.

    Evidence-based reassessment: Organisations should be able to explain, with supporting evidence, whether previous conclusions remain valid, require qualification or must be revised.

  35. Have ongoing-assurance guidance address documented reassessment decisions, identified decision-makers, appropriate corrective action and proportionate scrutiny where potential consequences warrant it.

    Accountable outcomes: Reassessment should lead to a documented decision, with an identified decision-maker, appropriate corrective action and proportionate scrutiny where the potential consequences warrant it.

  36. Test guidance effectiveness on selected AI use cases, including material changes in behaviour, evidence or operating conditions; examine reassessment adequacy and subsequent decision bases rather than simply whether review occurred.

    The effectiveness of this approach should be tested against selected AI use cases, including scenarios involving material changes in system behaviour, evidence or operating conditions. Evaluation should examine whether organisations can demonstrate the adequacy of their reassessments and the basis for subsequent decisions, rather than simply confirm that a review occurred.

  37. Use findings to determine whether additional guidance, standards or targeted oversight are needed, without unnecessarily duplicating existing requirements.

    The findings should inform whether additional guidance, standards or targeted oversight are needed, without unnecessarily duplicating existing requirements.

Read the original

Open in the interactive guide