Submission 230
Cordexa
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.
Cordexa, a small pre-revenue Brisbane company that builds a governed AI runtime, argues accountability should attach to a readable evidence record rather than the model. It recommends a common serious-incident definition aligned with the EU AI Act, procurement conditions, a reasonable-deployer expectation for SMEs, developer/deployer-differentiated standards, and a machine-readable evidence format.
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.
Recommends making the evidence record the unit of accountability for AI outputs that inform consequential decisions.
Make the evidence record the unit of accountability. Any AI output that informs a consequential decision should carry, in a form a person or regulator can read, what sources it relied on, what checks it passed, what it refused to answer, and who approved it.
Argues that obligations attached to models rather than deployments will be out of date before they commence.
Any standard that attaches obligations to the model rather than to the deployment will be out of date before it commences.
Says the main adoption barrier it meets among SMEs is uncertainty about what happens when AI is wrong.
(d) Barriers to adoption for small and medium businesses The barrier we meet most often is not cost or skills. It is that a business owner cannot answer the question “what happens when it is wrong?” and so declines to start.
Recommends Commonwealth procurement set a floor requiring an evidence record and a human approval gate.
Use Commonwealth procurement to set the floor: require the evidence record and a human approval gate in every AI procurement that affects a person or a public decision.
Says Australia has no definition of a serious AI incident or reporting path and recommends one interoperable with the EU AI Act.
There is no Australian definition of a serious AI incident and no reporting path. We recommend the Committee propose one, drafted to interoperate with the EU AI Act Article 73 definition
Argues that standards should distinguish developers from deployers, with the heaviest obligations on developers and proportionate, technology-neutral deployer obligations.
The Australian Standards for AI should distinguish developers from deployers, attach the heaviest obligations to the former, and keep the deployer obligations proportionate and technology-neutral.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Make the evidence record the unit of accountability: AI output informing a consequential decision should carry, in readable form, its sources, checks passed, refusals and approver.
Make the evidence record the unit of accountability. Any AI output that informs a consequential decision should carry, in a form a person or regulator can read, what sources it relied on, what checks it passed, what it refused to answer, and who approved it.
Define a serious AI incident once, interoperable with the EU AI Act Article 73 definition.
Define “serious AI incident” once, in terms interoperable with the EU AI Act Article 73 definition, so Australian vendors and deployers build one reporting path, not several.
Require an evidence record and human approval gate in Commonwealth AI procurement affecting people or public decisions.
Use Commonwealth procurement to set the floor: require the evidence record and a human approval gate in every AI procurement that affects a person or a public decision.
Clarify that deployers act reasonably when they keep complete evidence records, apply human approval gates to consequential outputs and can produce both on request.
A clear public expectation that a deployer who keeps a complete evidence record, applies a human approval gate to consequential outputs and can produce both on request has acted reasonably
Keep the Australian Standards for AI technology-neutral and deployer-focused so small Australian vendors can comply without licensing a frontier model.
Keep the Australian Standards for AI technology-neutral and deployer-focused so that small Australian vendors can comply without licensing a frontier model.
Give regulators and AISI a machine-readable evidence-record format enabling decision reconstruction without vendor cooperation.
Give regulators and the Australian AI Safety Institute a machine-readable evidence record format so a decision can be reconstructed after the fact without the vendor’s cooperation.
Ask AISI to publish a minimum machine-readable record consequential-AI deployers must produce: workflow, inputs and sources, checks and results, withheld items and reasons, human approvals with identity/time, model and version.
We recommend the Committee ask the Australian AI Safety Institute to publish a minimum evidence record format, machine-readable, that deployers of consequential AI must be able to produce. The fields are not exotic: workflow identifier, inputs and their sources, checks applied and their results, items withheld and why, human approvals with identity and time, model and version used.
Consider whether the design of closed procurement panels is keeping Australian firms out of early government AI work.
If the Committee wants sovereign capability (term (e)) it should look at whether panel design is keeping Australian firms out of the early AI work.
Have the Australian Standards for AI distinguish developers from deployers, with the heaviest obligations on developers and proportionate, technology-neutral deployer obligations.
The Australian Standards for AI should distinguish developers from deployers, attach the heaviest obligations to the former, and keep the deployer obligations proportionate and technology-neutral.
For AI informing decisions about people or public outcomes, require accessible source references, recorded named human approval before consequential outputs take effect, and readable decision records on request.
We suggest three conditions for any AI system that informs a decision about a person or a public outcome: • every output carries source references that a reviewer can open; • consequential outputs pass through a named human approval step before they take effect, and the approval is recorded; • the deployer can produce, on request, a readable record of a given decision: sources, checks, exclusions, approver, time.