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 537

Dr Aashish Srivastava

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.

An academic submission arguing that Australia's technical governance of agentic AI should be matched by legal accountability. It makes two targeted recommendations: require risk-based preservation of records so legally consequential agentic actions can be reconstructed, and examine how existing rules of attribution, authority and causation apply, drawing on EU, UK, Singapore and US examples.

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 Australia has substantial technical controls for agentic AI and that the next challenge is translating them into legal accountability for consequential external actions.

    Australia has developed increasingly sophisticated technical governance for agentic AI, including requirements concerning human accountability, permissions, authorised scope, monitoring and auditability. The next challenge is ensuring that these technical controls translate into effective legal accountability when AI systems move from generating outputs to taking consequential external actions.

  2. Argues that technical observability and legal reconstructability overlap but may not be identical, since a legal inquiry may require a broader factual picture than technical logs.

    Technical observability and legal reconstructability overlap, but they may not be identical. A technical log may demonstrate that an API was called, a particular agent executed a task or an external tool was invoked. A legal inquiry may require a broader factual picture.

  3. Argues that the EU AI Act model should not simply be transplanted into Australia because it addresses high-risk classifications rather than agentic external actions.

    The EU model should not simply be transplanted into Australia. Its obligations are connected to the AI Act’s classification of high-risk AI systems, whereas the distinctive concern raised by agentic AI is its capacity to undertake consequential external actions.

  4. Argues that Australian law applies to agentic AI, but that it is less clear whether existing attribution rules operate with sufficient clarity for autonomous multi-step conduct.

    The question is therefore not whether Australian law applies to agentic AI. It does. The more difficult question is whether existing rules, many of which were developed before systems could autonomously plan and execute multiple steps in pursuit of a general objective, will operate with sufficient clarity when applied to agentic conduct.

  5. States that the preservation obligation should be risk-based and should not cover every internal computational process or model chain-of-thought.

    This should be a risk-based obligation. It should not require preservation of every internal computational process, nor should it require disclosure or preservation of internal model chain- of-thought.

Recommendations

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

  1. For higher-consequence agentic AI, require deploying organisations to preserve reconstructable action records, identifying where relevant: objectives; permissions, credentials and tools; material tool/system interactions and external actions; human approvals; and material permission/boundary changes.

    For higher-consequence agentic AI, deploying organisations should be required to preserve sufficient records to reconstruct legally consequential external actions. At a minimum, where relevant to the application, those records should identify: 1. the objective or instruction under which the agent was operating; 2. the permissions, credentials and tools available to the agent at the relevant time; 3. material tool invocations and interactions with external systems; 4. material external communications, transactions or other consequential actions; 5. applicable human approval requirements and whether those approvals occurred; and 6. material changes to permissions or operational boundaries relevant to the event.

  2. Make preservation risk-based: do not require every internal computation or internal chain-of-thought; avoid disproportionate low-risk requirements; trigger obligations on capacity materially to affect legal, financial, protected-system/information, governmental or prescribed higher-consequence interests.

    This should be a risk-based obligation. It should not require preservation of every internal computational process, nor should it require disclosure or preservation of internal model chain- of-thought. Ordinary low-risk uses of AI should not attract disproportionate record-keeping requirements. The trigger should instead be the capacity of an agentic system to undertake actions capable of materially affecting legal rights or obligations, financial interests, access to protected systems or information, the exercise of governmental powers, or other prescribed higher-consequence interests.

  3. Build on existing infrastructure: possible use of government AI assurance, Corporations Act s 286 for corporations, particularly financial actions, and proportionate preservation requirements from relevant private-sector regulators; avoid a separate logging regime.

    Implementation should build upon existing regulatory infrastructure rather than create a separate technical logging regime. For Australian Government use, the requirement could be incorporated into existing AI assurance arrangements. For corporations, existing record- keeping obligations under the Corporations Act 2001 (Cth) (s 286) provide a possible legislative starting point, particularly where agentic systems undertake transactions or other financially consequential actions. In other regulated private-sector contexts, relevant regulators could specify proportionate record-preservation requirements for higher-consequence agentic applications.

  4. Set retention periods according to the activity and existing statutory record-keeping requirements, without requiring uniform periods.

    The precise retention period need not be uniform. It should reflect the nature of the activity and existing statutory record-keeping requirements.

  5. Examine attribution, authority and causation for consequential autonomous actions pursuing human objectives, complementing existing technology-neutral-law work; focus on distributed conduct across providers, deployers, users and autonomous processes without human direction of each action.

    The Australian Government should examine how existing Australian rules of attribution, authority and causation apply where an AI agent autonomously determines and undertakes consequential external actions in pursuit of a human-defined objective. This work should complement, rather than duplicate, existing government examination of the application of technology-neutral laws to AI. Its specific focus should be circumstances in which relevant conduct is distributed across system providers, deployers, users and autonomous system processes, and where the person providing the objective may not directly determine each consequential action undertaken by the system.

  6. Have that examination consider whether authority and attribution concepts cover actions taken without human direction at each step, how causation and contribution operate across actors, the significance of control by providers, deployers and users, and whether particular statutory provisions create uncertainty.

    The examination should consider: • whether existing concepts of authority and attribution adequately address actions undertaken by an AI system without specific human direction at each step; • how existing rules of causation and contribution operate where different actors participate at different stages of an agentic workflow; • the legal significance of the respective control exercised by system providers, deployers and users; and • whether particular statutory provisions create uncertainty when applied to systems that autonomously determine and execute intermediate actions.

Read the original

Open in the interactive guide