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 200

Jeff Bleich Centre for Democracy and Disruptive Technologies, Flinders University

Submitter group: Universities.

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 sovereignty needs capacity and autonomy and proposes a national governance harness the Centre offers to build. It represents participating researchers, not necessarily all Centre members or Flinders University. It describes limited advisory trials, including adverse results, and describes Anthropic and Microsoft affiliations; it says Anthropic officers had no study role. It seeks independent evaluation and Robodebt reforms, while stating the harness does not deliver provider independence or substitute for Indigenous governance.

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 sovereignty is a status resting on capacity and autonomy, and that Australia is funding only capacity.

    Sovereignty is a status, generated by two dimensions — capacity and autonomy — of which Australia is systematically funding only one.

  2. Argues that a capacity-only account of sovereignty cannot distinguish owning an AI stack from governing it.

    a capacity-only account of AI sovereignty cannot distinguish between an actor that owns its AI stack and governs it on its own terms, and an actor that owns an impressive stack whose rules of use, model behaviour, content policies and update cycles are in fact determined elsewhere.

  3. Describes a governance harness as a thin, deterministic, auditable software layer between users and frontier models, enforcing named functions.

    We define a governance harness as a thin, deterministic, auditable software layer that sits between users and one or more frontier models and enforces a small number of named functions

  4. Argues that safety-calibrated behaviour on regulated work is a property of the architecture the model is embedded in, not of the model's parameter count.

    The strategic implication for a middle power is direct: safety-calibrated behaviour on regulated work is a property of the architecture the model is embedded in, not of the parameter count of the model.

  5. States that the Centre's Advisory Board chair is Anthropic's International Special Envoy and that Anthropic models were evaluated in both studies.

    First, the Chair of the Centre’s Advisory Board, the Hon. Jeffrey Bleich, is International Special Envoy at Anthropic; Anthropic models were among those evaluated in both studies.

  6. States that both papers' lead author is a Centre Research Fellow employed as a Microsoft security researcher, and two Microsoft models were evaluated in the second study.

    Second, the lead author of both papers is a Research Fellow of the Centre who is also employed as a security researcher at Microsoft, and two Microsoft models were among the ten evaluated in the second study.

Recommendations

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

  1. Adopt, and recommend the Commonwealth adopt, a two-dimensional definition of AI sovereignty comprising capacity and autonomy, as set out in Part 2 of the submission.

    That the Committee adopt, and recommend the Commonwealth adopt, a two- dimensional definition of AI sovereignty comprising capacity and autonomy, in the terms set out at Part 2.

  2. Define sovereignty through AI access, building, deployment, operation and use across economic, administrative, social, scientific and security domains, plus independent purposes, behaviour and domestic accountability free from foreign or domestic coercion or imposition.

    AI sovereignty is the status an actor holds when it possesses both the capacity to access, build, deploy and operate artificial intelligence and to use it to achieve goals across the economic, administrative, social, scientific and security realms; and the autonomy to determine independently what it wants AI to do and how it should behave, and to govern its use under domestic norms and accountability mechanisms, free from coercion or imposition by divergent interests, whether foreign or domestic. Both dimensions are required. Neither is sufficient. Sovereignty is never delivered by capacity alone.

  3. In the adopted definition, require resilience of essential AI functions under external-stack disruption through diversified, substitutable supply and domestic fallback, eliminating external single points of control without requiring domestic production of the entire stack.

    Autonomy in this sense has two parts. The first is non-dependence: the ability to keep essential AI- enabled functions running when the external stack is withdrawn or disrupted — by export controls, sanctions, price shocks, a provider’s unilateral change of terms, or geopolitical rupture. The relevant test is not everyday self-reliance but resilience under disruption. This does not imply autarky. A fully autonomous actor is not one that builds everything itself, but one that can rely on diversified, substitutable sources of supply and fall back on domestic solutions when it must. The goal is not to reproduce the entire stack domestically; it is to eliminate single points of external control over functions that cannot be allowed to fail.

  4. Define governance autonomy as adding constraints above provider outputs while retaining provider safety policies, adding accountability and never subtracting it.

    Critically — and this is the point the Committee may find most useful — this rule-setting is exercised by adding governance constraints on top of a provider’s outputs, not by relaxing, overriding or circumventing the safety policies the provider itself enforces. Those policies remain in force regardless of how a jurisdiction configures its governance layer. Autonomy so defined is conservative: it can only ever add accountability, never subtract it.

  5. Classify every Commonwealth investment presented as advancing sovereign AI as capacity-building or autonomy-building, and report the balance annually.

    That every Commonwealth investment presented as advancing sovereign AI be classified on the record as capacity-building or autonomy-building, and that the balance be reported annually.

  6. Make commissioning a national reference governance harness the primary autonomy-building investment of this Parliament.

    That the Commonwealth commission a national reference governance harness — an open, auditable reference implementation with a published conformance benchmark — as the primary autonomy-building investment of this Parliament.

  7. Require an open, auditable reference harness held and versioned by a Commonwealth body, with reference code, a curated Commonwealth legal and administrative authority registry, and a published conformance benchmark for testing deployments.

    The Commonwealth should commission an open, auditable reference implementation of the harness pattern, held and versioned by a Commonwealth body, bound to a curated registry of Commonwealth legal and administrative authorities, and released with a published conformance benchmark against which any deployment — Commonwealth, State, or vendor-supplied — can be tested. The deliverable is three things: the reference code, the source registry, and the benchmark. All three are artefacts government owns, can version, can redline, and can re-run on every model substitution.

  8. Fund each State and Territory to instantiate the reference harness against its own corpus of legal authorities, so governance capability is distributed across the federation.

    That the Commonwealth fund each State and Territory to instantiate that reference harness against its own corpus of legal authorities, so that governance capability is distributed across the federation rather than concentrated in Canberra.

  9. Complete Robodebt Royal Commission recommendations 17.1 and 17.2, giving the monitoring body under 17.2 an architectural rather than model-based audit surface.

    That the Commonwealth complete recommendations 17.1 and 17.2 of the Royal Commission into the Robodebt Scheme, and that the monitoring body contemplated by 17.2 be given an audit surface that is architectural rather than model-based.

  10. Specify the governance architecture rather than the model in procurement for regulated advisory workloads, so models stay substitutable without re-certification.

    That Commonwealth procurement for regulated advisory workloads specify the governance architecture rather than the model, so that models remain substitutable between procurement cycles without re-certification.

  11. Fund evaluation work to extend the evidence base from advisory to decisional systems, conducted independently of the vendors whose products are assessed.

    That the Commonwealth fund the evaluation work required to extend this evidence base from advisory systems to decisional systems, and that the evaluation be conducted independently of the vendors whose products it assesses.

  12. Build and evaluate the Commonwealth instance rather than assuming transfer from the two-state evidence.

    Recommendations 3 and 4 are therefore a policy inference from a promising result, not a finding the papers report. That is precisely why the Commonwealth instance should be built and evaluated rather than assumed.

  13. If applying the architecture to Indigenous governance, use CARE Principles and Indigenous-led design from the outset; communities must lead corpus, authority-list and escalation-rule design.

    The CARE Principles are the appropriate framework for that domain and require Indigenous-led design from the outset. The architecture could in principle support Indigenous-led governance — its policy router, authority gate and escalation logic are general — but only if Indigenous communities lead the design of the corpus, the authority list and the escalation rules.

  14. Formally protect independence in any Commonwealth engagement for the proposed evaluation.

    The Centre considers these disclosures appropriate given that Recommendation 7 proposes evaluation independent of vendors, and would expect any Commonwealth engagement to be structured with that independence formally protected.

Read the original

Open in the interactive guide