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 560

Ms Viveka Mohan Das

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.

Supports national data-centre minimums and targeted AI training and answer-engine duties, including notification from awareness. Proposes proportionality tiers and independent researcher access. Discloses commercial interests in transparency and a potential lower-tier service; says the author does not speak for First Nations peoples and seeks direct engagement with rightsholders.

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. Says a single megawatt figure is not sufficient but is the right primary screen, aligning with the AEMC draft access-standard tiers of 30 MW and 100 MW.

    A single MW figure is not sufficient, but it is the right primary screen. It aligns with the AEMC draft access-standard tiers of 30 MW and 100 MW.

  2. Says new renewable capacity should be defined by commissioning date and exclude already-contracted supply.

    “New” should mean capacity commissioned after a stated reference date, and not capacity already contracted or supported under RET, REGO or equivalent state schemes.

  3. Argues that a four-week gap between awareness and notification of an incident would not have been caught by an annual risk-assessment cycle, but would have been caught by a duty that runs from awareness.

    That four-week awareness-to-notification gap would not have been caught by an annual risk-assessment cycle. It would have been caught by a duty that runs from awareness.

  4. Recommends that the Digital Duty of Care Bill expressly cover AI answer engines, whether or not their outputs are designed to be shared.

    Amend section 25A(1)(j), or add a category, so that a service is in scope if it generates answers, summaries or recommendations in response to user queries, whether or not those outputs are designed to be shared.

  5. Says the Digital Duty of Care Bill has no definition of an agentic system, no pre-deployment assessment of tool-use, no audit-log duty and no notice duty when an agent affects a third-party system or an Australian user.

    The Bill has no definition of an agentic system, no pre-deployment assessment of tool-use, no audit-log duty, and no notice duty when an agent affects a third-party system or an Australian user.

  6. Says traditional owners of each land where data centre infrastructure is sited, and of catchments from which water is drawn, are the appropriate parties to define the terms of their own engagement and should be treated as rightsholders, not stakeholders.

    The traditional owners of each land on which data centre infrastructure is sited — and the traditional owners of catchments from which water is drawn — are the appropriate parties to define the terms of their own engagement. That engagement should treat them as rightsholders, not as stakeholders.

Recommendations

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

  1. Support mandatory national minimums for large data centres, with states and territories keeping planning and able to set higher rules.

    Support mandatory national minimums for large data centres. States and territories keep planning and may set higher rules.

  2. Keep megawatts as the primary screen and add triggers for annual MWh, stressed-catchment water, cumulative campus/portfolio capacity and frontier training.

    Keep MW as the primary screen. Add triggers: annual MWh; water in stressed catchments; cumulative campus or portfolio capacity; frontier-training activity.

  3. Define 'sovereign' and 'onshore' as five attributes (developed, owned, hosted, operated, accountable) rather than collapsing them.

    Define “sovereign” and “onshore” as five attributes: developed, owned, hosted, operated, accountable. Do not collapse them.

  4. Not attach private metrics or third-party tools to the standards, and use templates and a common schema.

    Do not attach private metrics or third-party tools to the standards; use templates and a common schema.

  5. Recognise AI answer-engine discoverability as a public-interest obligation.

    Recognise AI answer-engine discoverability as a public-interest obligation.

  6. Use the AEMC draft tiers (below 30 MW out; 30-100 MW baseline; 100 MW or 150 MW portfolio equivalent for the heavy tier) and define campus and portfolio aggregation in the statute by the same connection point, same water system, or common control.

    Use the AEMC draft tiers (below 30 MW out; 30–100 MW baseline; 100 MW or 150 MW portfolio equivalent for the heavy tier) and define campus and portfolio aggregation in the statute — same connection point, same water system, or common control.

  7. Choose one statutory capacity basis from design capacity, ultimate network connection capacity or IT load; report contracted, connected and actual load separately.

    The paper uses three bases — design capacity, ultimate network connection capacity, and IT load. Legislation should pick one and require separate reporting of contracted, connected and actual load.

  8. Apply the standards at commencement to projects not yet connected and to material expansions, require a compliance plan before connection, start with reporting and ramp performance obligations.

    Apply the standards to projects that have not yet connected, and to material expansions, at commencement. Require a compliance plan before connection. Start with reporting; ramp performance obligations.

  9. Have a site that goes live before new generation exists operate inside an agreed import limit that steps up as supply arrives.

    A site that goes live before new generation exists should operate inside an agreed import limit that steps up as supply arrives.

  10. Assign facility energy, water, siting and community duties to facility owners, and training conditions to model-training organisations.

    Facility owners carry facility duties (energy, water, siting, community). Organisations that train models carry training conditions.

  11. Bind tenants contractually to supply the facility owner's reporting data.

    Tenants should be contractually bound to supply the owner’s reporting data.

  12. Define new renewables as capacity commissioned after a reference date, excluding already contracted supply or RET/REGO/equivalent state-scheme support.

    “New” should mean capacity commissioned after a stated reference date, and not capacity already contracted or supported under RET, REGO or equivalent state schemes.

  13. Use Ireland’s structure: at least 80% of annual demand from additional local-network renewables, excluding existing scheme generation, with a six-year glide-path.

    Ireland’s decision CRU/2025236 (12 December 2025) requires at least 80 per cent of annual demand from additional renewables that feed the Irish network, excludes generation under REFIT/RESS/ORESS, and allows a six-year glide-path from energisation. [Verified] Use that structure.

  14. Report the share of supply met from existing and from new capacity, with the new share rising on a published schedule, prohibit double counting, and use REGO time-of-generation data for defined peak periods for the largest tier only.

    Report the share met from existing and from new capacity, with the new share rising on a published schedule. Prohibit double counting. For the largest tier only, use REGO time-of-generation data for defined peak periods.

  15. Link earlier or larger staged connection to firm flexibility, including curtailment in system-stress events; restrict non-renewable backup to emergencies and report run-hours.

    Link earlier or larger staged connection to firm flexibility (storage, on-site generation, agreed curtailment in system-stress events). Restrict non-renewable backup to emergencies and report run-hours.

  16. Publish criteria for the AER modification route for state-owned supply: grid-stability benefit, price benefit for other customers, and no substitution of already-planned state generation.

    The AER modification route for state-owned supply needs published criteria: grid-stability benefit, price benefit for other customers, and no substitution of already-planned state generation for the additionality test.

  17. Use operational PUE for compliance (design PUE only at approval), always report PUE with WUE, and do not copy a single national PUE before testing against Australian climate zones. Use NABERS for comparison, not as a substitute for additionality.

    Use operational PUE for compliance; design PUE only at approval. Always report PUE with WUE. Do not copy a single national PUE — including the paper’s <1.25 / <1.34 examples —before testing against Australian climate zones. NABERS is useful for comparison, not a substitute for additionality.

  18. Require monthly operational data to the regulator and public reporting at least annually, quarterly for the top tier, with NGER facility-level reform as the vehicle.

    Monthly operational data to the regulator; public reporting at least annually, quarterly for the top tier. Pending NGER facility-level reform is the natural vehicle.

  19. Use local water-performance tests rather than national averages; report withdrawal and consumption separately by source, with peak/seasonal demand and a regional stress factor.

    Yes to performance metrics. WUE alone is not enough. Report withdrawal and consumption separately, by source, with peak and seasonal demand, and a regional water-stress factor. Do not use a national average as the test.

  20. Prioritise efficient cooling, then non-potable/recycled water, potable last; allow same-catchment offsets/desalination only if verifiable and additional, with a transition timeframe and drought trigger limiting operation.

    Support the hierarchy: efficient cooling first, then non-potable and recycled water, potable last. Same-catchment offsets and desalination only if verifiable and additional, with a transition timeframe and a drought trigger that limits operation.

  21. Have water plans cover existing users, environmental water, Traditional Owner interests, and cumulative load from nearby facilities.

    Plans should cover existing users, environmental water, Traditional Owner interests, and cumulative load from nearby facilities.

  22. Test no-net-costs with a published counterfactual, direct/project/shared/system cost allocation, uncertainty range and review/dispute process; also track land competition, housing pressure and skilled-labour displacement.

    The gaps most likely to land on consumers are network augmentation recovered across all users; mixed-benefit shared infrastructure; and water delivery and servicing. Require a published counterfactual, an allocation method that separates direct, project-triggered, shared and system-wide costs, a stated range where attribution is uncertain, and a review and dispute process. Also track land competition, housing pressure and skilled-labour displacement.

  23. Engage before irreversible site selection, with a named contact, response time, commitments register, public change report, complaint/incident path, and reporting through operation and expansion.

    Engage before site selection is irreversible. Named contact and response time. Commitments register. Public report of what changed. Complaints and incident path. Report against commitments through operation and expansion.

  24. Publish plans and engagement records as a Commonwealth transparency floor; leave setbacks to states/councils based on noise, traffic and construction impacts, without a national metre rule.

    Commonwealth role: the transparency floor. Publish plans and engagement records. Leave setback distances to states and councils, set by reference to noise, traffic and construction impact, not a national metre number.

  25. Require an alternatives assessment where land is suited to housing or irrigated agriculture.

    Require an alternatives assessment where land is suited to housing or irrigated agriculture.

  26. Require engagement with relevant traditional-owner Polities before site lock-in, treating land and catchment traditional owners as rightsholders who define their engagement terms.

    The traditional owners of each land on which data centre infrastructure is sited — and the traditional owners of catchments from which water is drawn — are the appropriate parties to define the terms of their own engagement. That engagement should treat them as rightsholders, not as stakeholders. The standards should require engagement with the relevant Polities before site lock-in;

  27. Require water plans to include cultural water and Traditional Owner interests.

    water plans that include cultural water and Traditional Owner interests;

  28. Establish a rule that Indigenous data, language, art and biocultural knowledge are not training inputs without authority from the relevant rights-holders, including the right to refuse.

    a rule that Indigenous data, language, art and biocultural knowledge are not training inputs without authority from the relevant rights-holders, including the right to refuse.

  29. Make community-controlled bodies eligible for any public-interest compute reservation.

    Community-controlled bodies should be eligible for any public-interest compute reservation.

  30. If the Commonwealth resources an Aboriginal- and Torres Strait Islander-led AI centre, have those peoples design and lead it; government may fund infrastructure and support.

    If the Commonwealth later resources an Aboriginal- and Torres Strait Islander–led AI centre, that body should be designed and led by those peoples, not specified by a submission from a non-Indigenous author. Government may fund infrastructure and support.

  31. Offer community fibre uplift as an option in the commitments register, not a hidden condition.

    Community fibre uplift should be an offered option in the commitments register, not a hidden condition that makes every site a telecommunications project.

  32. Require paid construction/operation apprenticeships, traineeships and cadetships; report shortages, net new places, completions, retention, wages and displacement from housing, health, energy or defence, by gender/other cohorts where data quality allows.

    Paid apprenticeships, traineeships and cadetships in construction and operation. Report additionality: baseline shortages, net new places, completions, retention, wages, and any evidence of displacement from housing, health, energy or defence. Report by gender and other cohorts where data quality allows.

  33. Link skills conditions to existing minimum-share tools such as the Australian Skills Guarantee rather than inventing a parallel scheme.

    Link to existing minimum-share tools (Australian Skills Guarantee; Victoria’s Major Projects Skills Guarantee) rather than inventing a parallel scheme.

  34. Commission Australian early-career employment data by AI exposure using ABS or ATO payroll before setting AI-specific labour conditions.

    Commission an Australian series on early-career employment by AI exposure (ABS or ATO payroll) before setting AI-specific labour conditions.

  35. Layer defined-incident reporting, safety-testing information on request to AISI/relevant agencies, and specified published summaries; protect commercial confidentiality through purpose limitation and secure handling.

    Layer information sharing: proactive reporting of defined reportable incidents; on-request safety-testing information to the AI Safety Institute and relevant agencies; published summaries in a specified form. Safeguards: confidentiality for commercially sensitive detail, purpose limitation, secure handling.

  36. Legislate outcomes and reporting duties, put technical detail in instruments that can be updated, and use the SOCI Act and FATA rather than a parallel security code where they apply.

    Legislate outcomes and reporting duties. Put technical detail in instruments that can be updated. Use the SOCI Act and FATA rather than a parallel security code where they already apply.

  37. Prefer a compute-reservation scheme with published eligibility, pricing and reporting over in-kind deals, with unused reserved hours lapsing back to the public pool or being reallocated.

    Prefer a compute-reservation scheme with published eligibility, pricing and reporting over in-kind deals that cannot be compared. Unused reserved hours should lapse back to the public pool or be reallocated.

  38. Measure outcomes, not announcements, using indicators such as local research and engineering staff, partnerships, compute hours reserved and used by user type, onshore-use share, research outputs and VET and TAFE places.

    Indicators: local research and engineering staff; university and public-research partnerships; compute hours reserved and used, by user type; an onshore-use share (workloads whose data is stored and processed in Australia); research outputs and commercialisation; VET and TAFE places. Measure outcomes, not announcements.

  39. Keep hosting location separate from ownership and control, with conditions any provider can meet and no nationality test.

    Keep hosting location separate from ownership and control. Conditions any provider can meet; no nationality test.

  40. Do not stretch the Hosting Certification Framework silently over every commercial training run.

    The Hosting Certification Framework already looks at ownership, control and location for government-classified hosting and is under reform. [Verified] Do not stretch it silently over every commercial training run.

  41. Have organisations running frontier training jobs report defined incidents to the AI Safety Institute from awareness (not detection or external notification, nor annually), with the duty attaching to the organisation running the training job.

    The infrastructure standards being developed by PM&C should require that organisations running frontier training jobs report defined incidents to the AI Safety Institute from awareness — not from detection, not from external notification, not on an annual schedule.

  42. Require model-training conditions for answer systems to disclose broad training-content categories, origin jurisdictions and licensing status, without publishing training data or model weights.

    [Proposed] Training conditions for facilities used to train models that power AI answer systems should include a source-category disclosure requirement: a published account of the broad categories of content used in training (news media, government documents, scientific literature, social media, licensed data, synthetic data), the jurisdictions from which content was drawn, and whether content was used with or without licence. This is not a requirement to publish training data or model weights.

  43. Keep source-category disclosure separate from licensing questions handled by CAIRG; do not delay disclosure for that copyright process.

    The Copyright and AI Reference Group (CAIRG) is the right vehicle for licensing questions. This disclosure obligation is separate and should not wait for the copyright process.

  44. Ask proponents who state productivity or public-value gains for the success criteria and success rate of the goal, with facility energy, water and emissions reporting remaining the compliance core.

    Where a proponent claims productivity or public-value gains, ask for the success criteria and success rate of the claimed goal. Facility energy, water and emissions reporting remains the compliance core.

  45. If the Office tests outcome reporting, run a small independent pilot with an evaluator who is not the submitter, with no mandatory use.

    If the Office later tests outcome reporting, it should be a small independent pilot with an evaluator who is not the author. No mandatory use.

  46. Pair each reporting duty with a template and a common schema.

    Pair each reporting duty with a template and a common schema.

  47. Require disclosure of owner, hosting, retention, deletion and independence where third-party tools are used, and evaluate tools independently before any government mention of a product.

    If third-party tools are used by proponents or councils, require disclosure of owner, hosting, retention, deletion and independence. Evaluate tools independently before any government mention of a product.

  48. Publish an interface map showing which body decides which question.

    Publish an interface map: which body decides which question.

  49. Pair the standards with faster connection processes for genuine proposals, a common schema shared with states, and a published decision report after the first reporting period.

    Pair the standards with faster connection processes for genuine proposals; a common schema shared with states; and a published decision report after the first reporting period.

  50. Calibrate the reporting schema during implementation and do not defer the standards for a research pilot.

    Calibrate the reporting schema during implementation. Do not defer the standards for a research pilot.

  51. Form mandatory data-centre standards and AI training conditions into a single legislative package rather than separate instruments.

    The Committee should recommend that mandatory data-centre standards and AI training conditions form a single, coherent legislative package rather than separate instruments, to avoid a gap in accountability between facility-level and model-level obligations.

  52. Align the Commonwealth water floor with the water ministers' agreed minimums (minimise use, climate resilience, efficiency standards with transparent reporting, paying a share of water infrastructure) rather than writing a second code.

    water ministers agreed, with Queensland dissenting that day and the Northern Territory not agreeing the principles that day, minimums of minimise use, climate resilience, efficiency standards with transparent reporting, and paying a share of water infrastructure. Those principles go to National Cabinet. [Verified] Align the Commonwealth water floor with that text rather than writing a second code.

  53. Include in scope for training conditions platforms that serve AI-generated answers to Australian users using compute at facilities covered by the standards.

    Add: platforms that serve AI-generated answers to Australian users using compute at facilities covered by these standards.

  54. Have training duties follow the organisation that runs the training job and source-category disclosure duties follow the organisation that publishes an AI answer service.

    Training duties follow the organisation that runs the training job. Source-category disclosure duties follow the organisation that publishes an AI answer service.

  55. Design Commonwealth data-centre and training standards to inform, not pre-empt or overlap with, state-level AI governance, with a common reporting schema shared with the Royal Commission's secretariat.

    Commonwealth data-centre and training standards should be designed to inform, not pre-empt or overlap with, state-level AI governance processes. A common reporting schema shared with the Royal Commission’s secretariat would reduce duplication and improve data quality at the national level.

  56. Require reserved-compute resellers to disclose owner identity/country, default and Australian hosting including plan tier, sub-processors, training use of customer content, and export/deletion arrangements including confirmation.

    This annex is the disclosure list for platforms that resell reserved compute. Provenance record — minimum fields • Legal name and country of the platform owner • Default hosting region and any Australian region, including plan tier required • Sub-processors that touch customer content • Whether customer content is used to train models • How a customer exports and deletes data, and how deletion is confirmed

  57. Apply training-corpus source-category, origin-jurisdiction and licence-status disclosure to reserved-compute resellers and AI answer-service operators; avoid nationality tests and universal facility-owner duties.

    [New] Broad source categories used in training corpus, jurisdiction of origin, and whether each category was licensed, royalty-free, or used without licence This is a disclosure duty on platforms that resell reserved compute to Australian builders and on operators of AI answer services. It is not a nationality test and not a condition on every data- centre owner.

  58. Cost the reporting schema during implementation.

    It does not offer a Regulatory Impact Statement cost figure. Cost the reporting schema in implementation.

  59. If answer engines are already covered, clarify this in the Bill or explanatory material; otherwise amend scope to cover AI answers to queries regardless of sharing.

    If the Department’s view is that these services are already inside designated internet services or search engines, the Bill or the explanatory material should say that. If not, the definition should be amended so that AI-mediated answers to user queries are in scope whether or not the product is built for sharing.

  60. Have any AI governance framework expressly address AI answer engines as a class of service with distinct accountability obligations: source attribution, factual accuracy for high-stakes topics, and scope under online safety and consumer protection law.

    The Committee should recommend that any AI governance framework — whether through the DDC Bill, the forthcoming AI mandatory standards legislation, or a separate Act — expressly addresses AI answer engines as a class of service with distinct accountability obligations: source attribution, factual accuracy obligations for high-stakes topics, and scope under online safety and consumer protection law.

  61. Require preliminary notice to the eSafety Commissioner within 72 hours of awareness of a material online safety incident, with a fuller report within 30 days.

    Require a preliminary notice to the eSafety Commissioner within 72 hours of the responsible person becoming aware of a material online safety incident, and a fuller report within 30 days.

  62. Define incident materiality by foreseeable serious harm to Australian users/non-users from service operation, design or AI output; initial notices must identify events, affected systems/cohorts and containment steps.

    Materiality should be severity, not complaint volume: a reasonably foreseeable serious harm to an Australian user or non-user from the service’s operation, design, or AI output. The initial notice should state what happened, the systems or cohorts affected, and containment steps.

  63. Make delay, and knowingly incomplete incident notice, separate civil-penalty grounds.

    Delay, and knowingly incomplete notice, should be separate civil-penalty grounds.

  64. Where the same event is an eligible data breach or critical-infrastructure cyber incident, let the SOCI Act, ASD and OAIC clocks prevail with eSafety notice cross-referring rather than duplicating, and let the Commissioner publish a de-identified summary.

    Where the same event is an eligible data breach or a critical-infrastructure cyber incident, the SOCI Act, ASD and OAIC clocks prevail, and the eSafety notice should cross-refer rather than duplicate. The Commissioner should be able to publish a de-identified summary.

  65. State expressly that search-integrated generative features are covered with the search service.

    State expressly that search-integrated generative features are covered with the search service.

  66. Not use the answer-engine category to require disclosure of model weights.

    Do not use this category to require disclosure of model weights.

  67. Require, in transparency reports under Part 12A, a service-level account of how sources are selected, ranked and attributed in answers on health, financial, legal and emergency topics.

    Do require, in transparency reports under Part 12A, a service-level account of how sources are selected, ranked and attributed in answers on health, financial, legal and emergency topics.

  68. Define an agentic AI system as one that can take material actions on external systems or data without a discrete human instruction at each step.

    Define an agentic AI system as one that can take material actions on external systems or data without a discrete human instruction at each step.

  69. For agents offered to Australians or used in a way that can affect Australian systems, require pre-deployment tool-risk assessments, material-action logs and 72-hour notice from awareness of unintended actions affecting Australian users, personal information or third-party systems.

    Where such a system is offered to Australian users, or is used in a way that can affect Australian systems, require: a pre-deployment assessment of tool-use risks; an audit log of material actions; and notice within 72 hours of awareness that the system has taken an unintended action affecting an Australian user, personal information, or a third-party system.

  70. Notify unauthorised access to a government or critical-infrastructure system first through the cyber channel, with eSafety notified where online safety is also engaged.

    Unauthorised access to a government or critical-infrastructure system should be notified first through the cyber channel. eSafety should receive that notice where online safety is also engaged.

  71. Publish a proportionality test for complaint duties so micro-services are not prescribed by default.

    Complaint duties need a published proportionality test so micro-services are not prescribed by default.

  72. Do not freeze the eSafety architecture against the Rickard review's longer institutional options; review the duty within three years of commencement.

    Do not freeze the eSafety architecture against the Rickard review’s longer institutional options. Review the duty within three years of commencement.

  73. Allow biennial Commissioner-template declarations for invite-only services without public user content or reasonably likely seriously harmful material; retain six years, produce within 30 days and re-tier on service changes.

    Publish a harm-profile tier. Invite-only services with no public user-generated content and no reasonable likelihood of seriously harmful material should meet the assessment duty by a biennial declaration on a Commissioner template, retained for six years and producible within 30 days. Re-tier when the service changes.

  74. Require risk assessments to record, for each material risk, a baseline, intended effect and whether the indicator moved, with a failed mitigation at a higher tier triggering earlier reassessment.

    For each material risk, record a baseline, the intended effect, and — at the next assessment — whether the indicator moved. A failed mitigation at a higher tier should trigger an earlier reassessment.

  75. Keep university access and add accredited networks meeting ethics, conflict-register, Act-related research and Commissioner-proposal requirements, without commercial affiliation to the target platform.

    Keep the university path. Add an accredited-network path: published ethics standard, public conflict register, research related to the Act, a proposal lodged with the Commissioner, and no commercial affiliation with the platform whose data is sought.

  76. Decide accredited-researcher applications within 60 days, renew every three years, and confer no entitlement to model weights.

    Decide applications within 60 days. Renew every three years. Do not treat accreditation as a right to model weights.

  77. Add, for services that generate answers to user queries, a duty to publish design principles for selecting, ranking and attributing sources, at a level an independent reviewer can use to assess systematic error on high-stakes topics.

    Add, for services that generate answers to user queries, a duty to publish the design principles for selecting, ranking and attributing sources, at a level an independent reviewer can use to assess systematic error on high-stakes topics.

  78. Recommend direct engagement with traditional owners and First Nations peak bodies, who define their governance needs, rather than adopting prescriptions from the non-Indigenous submitter.

    The traditional owners of each relevant Country, and First Nations peak bodies, are the appropriate parties to define their own governance needs. The author asks the Committee to recommend direct engagement with those parties rather than adopting proposals specified by a non-Indigenous submitter.

  79. Before ministerial harm-category or child-harm determinations, require First Nations peak-body co-design and, where harm concerns a specific Country/community, engagement with its traditional owners as rightsholders.

    Before the Minister adds a category under s.25C(2) or makes a child-harm determination under s.25D(2), require co-design with First Nations peak bodies and, where the harm relates to a specific Country or community, engagement with the traditional owners of that Country as rightsholders — not only Commissioner advice.

  80. Require accuracy and attribution for AI content about First Nations culture, language or heritage.

    (b) AI systems that generate content about First Nations culture, language or heritage be subject to an accuracy and attribution obligation;

  81. Recommend that AI governance instruments, not only online safety law, expressly address First Nations data sovereignty, with any such instrument led by or developed in partnership with First Nations peoples and communities.

    The JSCAI may wish to recommend that AI governance instruments — not only online safety law — expressly address First Nations data sovereignty, and that the design of any such instrument be led by, or developed in partnership with, First Nations peoples and communities rather than specified by government or industry alone.

  82. Allow eighteen months after Royal Assent before the s.26B penalty is enforceable against services below a published harm threshold, publishing the tier guide, template assessment and declaration path in that window.

    Eighteen months after Royal Assent before s.26B is enforceable against services below a published harm threshold. In that window, publish the tier guide, the template assessment, and the declaration path.

  83. Treat children's best interests as a primary consideration in the s.26A assessment for services reasonably likely to be accessed by children, recording why if another interest overrides them.

    For a service reasonably likely to be accessed by children, the assessment must treat children’s best interests as a primary consideration. If another interest overrides them, record why.

  84. Include frontier developers, hyperscalers and neoclouds selling or reserving relevant compute within training conditions.

    Who is in scope for training conditions Frontier developers, plus hyperscalers and neoclouds that sell or reserve the relevant compute.

Read the original

Open in the interactive guide