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 66

NableU Community Systems Incorporated

Submitter group: Civil society.

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 that AI policy should separate cognition, data access, authority and execution, with safeguards increasing as AI affects rights and services. Recommends risk frameworks, contestability, consent separation, cognitive accessibility, public-sector safeguards and system-level evaluation. The submitter is testing a governed disability-support AI companion as practical evidence, explicitly not a national blueprint.

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 proportionate governance can preserve powerful tools in low-risk settings while stronger safeguards apply where AI affects rights, essential services, safety or real-world action.

    That distinction provides a practical basis for proportionate governance: powerful tools can remain useful in low-risk settings while stronger safeguards attach where AI affects rights, essential services, safety or real-world action.

  2. Argues downstream risk cannot be judged from model class alone, since a narrow classifier can exercise more consequential authority than a powerful general-purpose model.

    But downstream risk cannot be understood from model class alone. A powerful general-purpose model helping someone rewrite a letter may exercise almost no institutional authority; a narrow classifier rejecting a welfare claim, ranking a job applicant or triggering an investigation can materially affect a person’s life.

  3. States that human oversight is only meaningful if the reviewer has competence, authority, time and responsibility.

    A nominal 'human in the loop' is not enough if the human routinely accepts the output, lacks time or competence to interrogate it, or has no real authority to change the outcome. Meaningful human oversight requires competence, authority, time and responsibility.

  4. Argues that storage, sharing, AI context, memory, training and execution permissions should not be bundled together.

    Separate privacy and authority permissions. Storage, human sharing, AI context, persistent memory, model training or evaluation, and automated execution should not be bundled together.

  5. Argues that AI can serve as cognitive-accessibility infrastructure when it remains optional, controllable and designed to return agency.

    Treat AI as cognitive-accessibility infrastructure as well as economic infrastructure. It can reduce executive- function, communication and administrative barriers when it remains optional, controllable and designed to return agency.

    A program corrected obvious text-recognition errors in this quote for display.

  6. Argues that sovereign AI investment should extend beyond frontier models to domain-specific, local and provider-independent systems.

    Sovereign-AI investment should include domain-specific models, Australian-controlled evaluation, private/local inference capability, open interoperability and provider-independent governance layers - not only frontier-scale model development.

Recommendations

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

  1. Develop a national risk framework considering capability, context, authority and consequence, keeping upstream obligations for powerful models while high-impact deployer obligations follow the authority and consequences of the real use case.

    Recommendation 1. Develop a national risk framework that considers capability + context + authority + consequence. Maintain appropriate upstream obligations for powerful models, while making high-impact deployer obligations follow the authority and consequences of the real use case.

  2. Distinguish advisory cognition, human-approved preparation, bounded delegated execution with scope, limits, provenance and revocation, and consequential decision/control; increase obligations with authority and execution, particularly for involuntarily affected people.

    For deployed AI, government should distinguish four levels: advisory cognition (explain, summarise, translate or recommend); preparation (draft or prepare, with deliberate human approval before execution); delegated execution (bounded action under explicit scope, limits, provenance and revocation); and consequential decision or control (material effects on rights, eligibility, access, restrictions or safety). Obligations should increase as AI moves from cognition into authority and execution, particularly where the affected person did not choose the system.

  3. Establish minimum rights to notice, intelligible explanation, correction, human review, contestability and accountable ownership for high-impact AI-supported decisions.

    Recommendation 2. Establish minimum rights to notice, intelligible explanation, correction, human review, contestability and accountable ownership for high-impact AI-supported decisions.

  4. For high-impact AI decisions, enable people to identify AI's involvement and authority, material information, legal accountability, correction processes, and review and appeal rights.

    When AI materially contributes to a high-impact decision, the affected person should be able to determine that AI was involved; whether it advised, prepared, executed or decided; what information materially affected the outcome; who is legally accountable; how relevant information can be corrected; and how review and appeal rights can be exercised.

  5. Make human oversight meaningful through competence, authority, time and responsibility.

    Meaningful human oversight requires competence, authority, time and responsibility.

  6. Require understandable, purpose-specific and revocable permission where consent is the relevant basis, separating storage, disclosure, AI context, memory, training/evaluation and execution authority.

    Recommendation 3. Require clear separation of storage, disclosure, AI context, AI memory, training/evaluation and execution authority, with understandable, purpose-specific and revocable permission where consent is the relevant basis.

  7. Where technically feasible, make consent withdrawal stop new access and actions promptly; have high-risk systems fail closed when authority is absent or invalid.

    Revocation also needs operational meaning. Where technically feasible, withdrawing permission should stop new access and new actions promptly. High-risk systems should fail closed when an authority state is absent or invalid rather than silently expanding access for convenience.

  8. Treat cognitive accessibility as a core responsible-AI and innovation objective: fund disability- and community-led co-design, evaluation and adoption, and test whether AI reduces or increases cognitive burden.

    Recommendation 4. Treat cognitive accessibility as a core responsible-AI and innovation objective. Fund disability- and community-led co-design, evaluation and adoption, and explicitly test whether AI reduces or increases cognitive burden.

    A program corrected obvious text-recognition errors in this quote for display.

  9. Use public-sector AI to help people navigate government before using it to govern them, keeping high-impact administrative power lawful, attributable and reviewable.

    Use public-sector AI to help people navigate government before using it to govern people. AI can explain, organise and prepare; high-impact administrative power should remain lawful, attributable and reviewable.

  10. Require Commonwealth AI deployments to name an accountable owner, document authority, disclose AI accessibly where it materially affects people, provide correction/review and proportionate provenance, and evaluate public burden/accessibility alongside productivity.

    Recommendation 5. Require Commonwealth AI deployments to have a named accountable owner, a documented authority level, accessible disclosure where AI materially affects a person, correction and review pathways, proportionate provenance, and evaluation of public burden and accessibility - not only agency productivity.

  11. Include domain-specific models, Australian-controlled evaluation, private/local inference, open interoperability and provider-independent governance layers in sovereign-AI investment, not only frontier-scale models.

    Recommendation 6. Sovereign-AI investment should include domain-specific models, Australian-controlled evaluation, private/local inference capability, open interoperability and provider-independent governance layers - not only frontier-scale model development.

  12. Have the AISI evaluate whole AI systems for fact/inference distinctions, authority expansion, revocation, dependency, excessive agreement, honest uncertainty, cognitive accessibility, reconstructible consequential actions and proportionate safety controls.

    The Australian AI Safety Institute is tasked with analysing and testing AI models and applications and supporting regulators to respond to emerging risks. Its remit is well suited to system-level evaluation. [5] Technical model benchmarks are necessary but insufficient. A model can perform well in isolation and still become unsafe through its context, memory, tools, interface or authority. Evaluation should test whether the system distinguishes fact from inference; whether increased capability expands access or authority; whether permissions can be revoked; whether the system creates dependency or excessive agreement; whether uncertainty is communicated honestly; whether it works across different communication and cognitive needs; whether consequential actions can be reconstructed; and whether safety controls remain proportionate.

  13. Have the AISI and National AI Centre jointly maintain reusable responsible-AI implementation patterns for SMEs, not-for-profits and community organisations, including system-level evaluation and cognitive-accessibility guidance.

    Recommendation 7. The AISI and National AI Centre should jointly maintain practical, reusable responsible-AI implementation patterns for SMEs, not-for-profits and community organisations, including system-level evaluation and cognitive-accessibility guidance.

  14. Extend SME and not-for-profit adoption support into reusable impact assessments, authority/consent models, provider inventories, procurement clauses, provenance schemas, evaluation packs and architectures for external, private and local processing.

    The National AI Plan already provides practical adoption support for SMEs and not-for- profits. That work should extend into reusable technical governance patterns: scaled impact assessments, authority and consent models, provider inventories, procurement clauses, provenance schemas, evaluation packs and reference architectures for external, private and local processing.

  15. Allow useful AI cognition while limiting data access and authority to what the task requires.

    Give AI as much cognitive capability as is useful, but only as much data and authority as the task actually requires.

Read the original

In some quotes a program corrected obvious text-recognition errors for display.

Open in the interactive guide