Submission 141
Infosys
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 that Australia should rely on its risk-based, technology-neutral approach and existing frameworks, complemented by targeted measures for agentic AI. Proposes a four-factor definition of AI sovereignty, domain-specific model investment, workforce retraining, auditable decision records for public services, and runtime assurance through the AI Safety Institute.
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.
Infosys argues that AI sovereignty should be defined through four dimensions of control, not data location alone.
Infosys recommends that the Australian Government define AI sovereignty through four dimensions of control, recognising that data location is only one component of a broader sovereignty framework. Defining sovereignty solely in terms of data residency does not provide meaningful control over AI systems or their operation.
Infosys argues that Australia is likely to gain greater strategic value from targeted domain-specific AI investment, given frontier development's scale, capital requirements and competitive dynamics.
Given the scale, capital requirements and competitive dynamics of frontier model development, Australia is likely to derive greater strategic value from targeted investment in domain-specific AI capability.
Infosys says many of the workers most affected by AI adoption are already employed, so workforce policy should support both future talent pathways and large-scale retraining and upskilling.
While Australia's education system remains critical to long-term capability development, many of the workers most affected by AI adoption are already employed. Effective workforce policy should therefore support both future talent pathways and large-scale workforce retraining and upskilling.
Infosys recommends evidence-grade decision records for AI in federal and state service delivery, enabling affected individuals, oversight bodies and auditors to understand decisions and authority.
Infosys recommends that the Australian Government require evidence-grade decision records for AI systems used in federal and state service delivery. Affected individuals, oversight bodies and auditors should be able to determine how a decision was reached, what information informed it, and under whose authority it was made.
Infosys argues that regulation should keep accountability traceable to a human or institutional principal through agentic execution chains, in proportion to risk.
Regulatory settings should ensure that accountability, authorisation and auditability remain traceable throughout those chains. Where AI systems act on behalf of an individual or organisation, the lineage of authority should remain attributable to a human or institutional principal, regardless of the number of intermediate steps performed by the system. Governance obligations should also be proportionate to risk.
Infosys argues that AI supply chain assurance should extend beyond the model itself to the full set of components on which a system depends.
AI supply chain assurance should therefore extend beyond the model itself to encompass the full set of components on which a system depends.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Adopt a four-factor definition of AI sovereignty covering data residency, operational control, model provenance and runtime authority.
The Australian Government should adopt a four-factor definition of AI sovereignty encompassing data residency, operational control, model provenance and runtime authority. This framework would provide a more comprehensive basis for managing sovereignty, security and resilience risks than data localisation requirements alone.
Prioritise investment in domain-specific models built on Australian sectoral expertise, regulated data and institutional knowledge.
Government should prioritise investment in domain-specific models built on Australian sectoral expertise, regulated data and institutional knowledge.
Support workforce adaptation through employer-led retraining, tertiary education, vocational pathways and lifelong learning, with attention to AI oversight, validation, governance and accountability capabilities.
Government should support workforce adaptation to AI through a combination of employer-led retraining, tertiary education, vocational pathways and lifelong learning initiatives. Particular attention should be given to developing capabilities associated with AI oversight, validation, governance and accountability, recognising that many of the workforce impacts of AI arise through changes to existing roles rather than occupational displacement.
Build on the incoming Privacy Act automated decision-making transparency reforms, and require evidence-grade decision records for consequential AI-supported decisions in public service delivery.
The Australian Government should build on the incoming Privacy Act reforms regarding automated decision-making transparency obligations. The Government should require evidence-grade decision records for consequential AI-supported decisions in public service delivery.
Include in procurement requirements for AI used in public administration the provision of auditable runtime evidence, including execution records, policy enforcement data and attribution of decision authority.
Procurement requirements for AI used in public administration should include the provision of auditable runtime evidence, including execution records, policy enforcement data and clear attribution of decision authority.
For higher-impact applications, ensure agencies can identify each operational AI system, its accountable business owner, approved scope of authority, current system and policy version, risk classification and lifecycle status.
For higher-impact applications, agencies should also be able to identify each operational AI system, its accountable business owner, approved scope of authority, current system and policy version, risk classification and lifecycle status.
Ensure regulatory settings keep authorisation, responsibility and auditability traceable to a human or institutional principal throughout an AI system's execution chain, with obligations proportionate to the consequences of actions.
Regulatory settings should ensure that authorisation, responsibility and auditability remain traceable to a human or institutional principal throughout an AI system's execution chain. Accountability obligations should be proportionate to the consequences of the actions undertaken.
Subject systems capable of irreversible or high-impact actions to stronger approval, monitoring and assurance requirements than low-risk or readily reversible systems.
Systems capable of executing irreversible or high-impact actions should be subject to stronger approval, monitoring and assurance requirements than systems performing low-risk or readily reversible activities.
Make AI assurance frameworks cover the full AI component stack, including models, data sources, inference services, connectors and external tools.
AI assurance frameworks should cover the full AI component stack, including models, data sources, inference services, connectors and external tools.
Ensure critical infrastructure operators can constrain and revoke AI system authority as a resilience measure.
Critical infrastructure operators should also be able to constrain and revoke AI system authority as a resilience measure.
Recognise AI-driven actions affecting critical infrastructure or public service delivery within relevant incident reporting frameworks, with the required evidence specified in advance.
AI-driven actions affecting critical infrastructure or the delivery of public services should be recognised within relevant incident reporting frameworks. The evidence required to support such reporting should be specified in advance to ensure systems are designed to generate it.
Support the Australian AI Safety Institute to evaluate AI systems in deployment environments.
The Australian AI Safety Institute should be supported to evaluate AI systems in deployment environments.
Have government procurement and governance requirements encourage runtime evidence, audit trails and decision records that enable oversight while respecting security, privacy and commercial confidentiality.
Government procurement and governance requirements should encourage the generation of runtime evidence, audit trails and decision records that enable effective oversight while respecting security, privacy and commercial confidentiality.
Accompany domain-specific AI investment with governance, operational controls and runtime infrastructure so capability remains secure, accountable and under Australian control.
Investment in domain-specific AI capability should therefore be accompanied by investment in the governance, operational controls and runtime infrastructure required to ensure that capability remains secure, accountable and under Australian control.
Keep consequential decisions accountable to identified people, supported by auditable decision records, and develop and resource meaningful human oversight as a skill.
Infosys recommends that accountability for consequential decisions remain with identified people, supported by the auditable decision records described elsewhere in this submission, and that the capability to exercise meaningful oversight be treated as a skill to be developed and resourced in its own right.
Enable affected individuals, oversight bodies and auditors to determine how AI-supported service decisions were reached, what informed them and under whose authority.
Infosys recommends that the Australian Government require evidence-grade decision records for AI systems used in federal and state service delivery. Affected individuals, oversight bodies and auditors should be able to determine how a decision was reached, what information informed it, and under whose authority it was made.
For consequential AI-supported actions, record initiating authority, delegation, system and policy versions, material information, approvals or escalations and resulting action; make detail proportionate to impact, with privacy, security and records safeguards.
For consequential AI-supported actions, an activity log alone may not provide sufficient accountability. Records should identify the initiating authority, delegation pathway, system and policy version, material information relied upon, relevant approvals or escalation points, and the resulting action. The level of detail should be proportionate to the potential impact of the outcome and supported by appropriate privacy, security and records management safeguards.
Prioritise AI that improves decision quality alongside efficiency in public services, resource allocation, compliance, infrastructure management and risk oversight.
More broadly, governments should prioritise AI applications that improve decision quality alongside operational efficiency, particularly in areas such as public service delivery, resource allocation, compliance, infrastructure management and risk oversight.
Make assurance consider harmful outcomes, misleading advice, inappropriate recommendations, contextual reasoning failures and affected individuals' ability to recognise errors.
Assurance frameworks should also consider the potential for harmful outcomes, misleading advice, inappropriate recommendations, failures of context-specific reasoning and the ability of affected individuals to recognise when a system may be incorrect.
Provide transparency, human review and accessible contestation or verification of AI-influenced government decisions; do not leave error identification solely to affected individuals, particularly in high-impact contexts.
For government services, this reinforces the importance of transparency, human review mechanisms and accessible pathways through which citizens can contest or verify decisions influenced by AI systems. Responsibility for identifying errors should not rest solely with affected individuals, particularly in high- impact contexts.
Extend AI Safety Institute evaluation beyond benchmarks to realistic deployment, including non-expert interpretation, safeguards against inappropriate reliance and practical harms.
As the Australian Government's specialist capability for understanding, testing and monitoring advanced AI systems, the Institute's evaluation activities should extend beyond benchmark testing and model performance assessments to include the examination of how AI systems behave in realistic deployment environments. This includes understanding how non-expert users interpret AI outputs, evaluating the effectiveness of safeguards designed to prevent inappropriate reliance, and identifying the types of harms that may emerge in practical use.
Support lifecycle evaluation of interactions between models, users, operational processes and deployment environments.
Evaluation frameworks should support a lifecycle-based understanding of AI system performance, encompassing the interactions between models, users, operational processes and deployment environments.
Emphasise runtime accountability, auditable execution and evidence generation in risk-based AI governance.
Infosys recommends that Australia adopts a risk-based approach to AI governance that places greater emphasis on runtime accountability, auditable execution and evidence generation.
Require high-impact AI systems to generate runtime evidence reconstructing actions, permissions, policies, authorisations and outcomes.
AI systems deployed in high- impact contexts should be capable of generating sufficient runtime telemetry and 6 Infosys Knowledge Institute (n.d.), ‘Enterprise AI runtime architecture: The layer that turns reasoning into execution’, Frontier Tech Perspectives, available at: Enterprise AI runtime architecture evidence to reconstruct actions taken, permissions exercised, policies applied, authorisations granted and outcomes executed.
Apply stronger evaluation, monitoring and oversight to applications affecting legal rights, public services, health, safety, financial outcomes or critical infrastructure than to low-risk administration.
Applications affecting legal rights, public services, health, safety, financial outcomes or critical infrastructure warrant stronger evaluation, monitoring and oversight than low-risk administrative applications.
Make assurance assess what systems may do when interacting with untrusted content and whether permissions can be constrained or withdrawn.
Assurance frameworks should therefore consider not only the reliability of system components, but also what a system is permitted to do when interacting with untrusted content, and whether those permissions can be constrained or withdrawn.
Give people approving AI outputs time, information and authority to disagree, with assurance that disagreement is legitimate rather than a performance concern.
A person asked to approve the output of an AI system needs the time to examine it, the information required to judge it, the authority to disagree with it, and the assurance that disagreement is a legitimate outcome rather than a performance concern.
Apply scoped, time-bound permissions, separation of duties for consequential actions and rapid, verifiable revocation across affected systems to mitigate AI identity and credential risks.
AI systems acting on behalf of individuals and institutions increasingly require their own identities and credentials. Where those credentials are broad in scope and persistent over time, they can create a concentration of authority that amplifies the consequences of compromise. The controls required to mitigate this risk are well established in human identity management, but have not yet been applied consistently to machine actors. These include scoped and time-bound permissions, separation of duties for consequential actions, and the ability to revoke authority rapidly and verifiably across all affected systems.