Submission 130
Noizend
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.
Noizend says it builds physics-informed AI for infrastructure noise and proposes using such tools in planning and government assessment. It seeks sovereign domain-specific AI funding, compute access, reconsideration of R&D supporting-activity exclusions and a ten-year refundability cap, published infrastructure verification, and engineering assurance standards. It defers fraud, deepfake and vulnerable-person harms to others.
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.
It argues Australia can lead in AI for bounded engineering problems in sectors where it already holds data, expertise and physical assets.
We can be the best in the world at AI that solves bounded engineering problems in sectors where we already hold data, expertise and physical assets: energy systems, resources, agriculture, health, defence.
It argues the Bureau of Statistics 12 per cent business adoption figure is the official statistic to plan against, rather than the 40 per cent figure.
Policy calibrated to 40 per cent will conclude the diffusion problem is largely solved. Policy calibrated to 12 per cent, with small and micro businesses at 11 per cent, will conclude it has barely started. The Bureau’s is the official statistic and the one to plan against.
It says the binding constraint for a deep tech SME is capital and compute, not awareness or skills.
The binding constraint for a deep tech SME is not awareness and it is not skills. It is capital and compute, so support should address both inputs directly.
It argues ten years is roughly the point at which a deep tech firm running physical prototyping, field validation and certification becomes commercially credible, while for a software firm it is generous.
For a software firm, ten years is generous. For a deep tech firm running physical prototyping, field validation and certification, ten years is roughly the point at which the technology becomes commercially credible.
It argues planning authorities may rely on technical assessments that acknowledge limitations and client-supplied information, while affected third parties are told not to rely on them.
Planning authorities may rely on technical assessments that expressly acknowledge limitations in scope and reliance on client supplied information, while affected third parties are told not to rely on those same reports.
It says it has found no published Australian regulatory assurance framework for AI-generated engineering assessments and that regulators will need methods to assess reliability.
We have been unable to identify a published Australian regulatory assurance framework specifically addressing AI-generated engineering assessments. As these tools enter professional practice, regulators will need methods for determining whether their technical outputs are reliable.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Prioritise domain-specific, Australian-trained AI capability.
The Committee should therefore prioritise domain specific, Australian trained AI capability in its recommendations.
Use tools that rapidly test infrastructure siting options in planning and design workflows where repeated iterations cause the biggest delays.
The gain arrives when the same physics is queried thousands of times, as in infrastructure design and operation. A noise assessment for a large battery system or data centre currently takes weeks of consultant time per iteration. If a developer can test forty siting options in an afternoon rather than two over a month, the design reaching the planning authority is materially better. That is productivity growth in the sense that matters: fewer rejected applications, less retrofit, fewer disputes, faster connection. The action is to use these tools in planning and design workflows where repeated iterations create the biggest delay.
Anchor AI capability in the regions where infrastructure is built, rather than running another awareness campaign.
The answer is not another awareness campaign. It is to put the capability where the infrastructure is. If a data centre or a large battery is required to instrument itself and publish verified operating data, regional engineering firms and university campuses can do the analytical work of interpreting that data locally, using domain models trained on that asset class. That is a permanent regional industry rather than a construction phase employment spike, so the policy response has to anchor capability in the regions.
Back firms commercialising acoustic, thermal and compliance solutions for grid-scale storage and data centres abroad.
The export case is simple. Every jurisdiction building grid-scale storage and data centre capacity faces the same acoustic, thermal and compliance problems and Australia’s strict rural night time limits and fast renewable build out force Australian firms to solve the hard version first. That is an export advantage, in the same way Australian mining technology exports followed from operating in difficult conditions. The practical step is to back firms that can commercialise those solutions abroad.
Target funding at the transition of sovereign AI from research to deployment.
What is missing is the capital to cross from research to deployment. Business expenditure on R&D sits at 0.9 per cent of GDP, below the OECD average of 1.99 per cent and total R&D intensity has fallen from 2.24 per cent of GDP in 2008-09 to 1.69 per cent in 2023-24. 9 Business R&D on AI specifically reached $668.3 million in 2023-24, up 142 per cent in two years and the fastest growing area of business R&D in the country. 10 Sovereign ownership and development are the missing piece of that puzzle, so funding should target the transition from research to deployment.
Confirm that the Australian Bureau of Statistics 12 per cent figure is the official business AI adoption statistic to plan against.
Policy calibrated to 40 per cent will conclude the diffusion problem is largely solved. Policy calibrated to 12 per cent, with small and micro businesses at 11 per cent, will conclude it has barely started. The Bureau’s is the official statistic and the one to plan against. We ask the Committee to confirm this.
If sovereign AI capability is a policy objective, strengthen the public instruments intended to fill the funding gap.
If sovereign AI capability is a policy objective, the market will not deliver it on current settings and the public instruments meant to fill the gap are oversubscribed and slow, so they need strengthening.
Recommend settings that bridge the Series B funding gap for sovereign AI and deep tech firms.
The Committee should recommend settings which bridge the Series B gap for sovereign AI and similar deep tech firms.
Expand compute support and make it more accessible to sovereign AI projects.
The Committee should recommend that compute support be materially expanded and made more accessible to sovereign AI projects.
Legislate favourable compute access for Australian start-ups, innovative small businesses, researchers and not-for-profits as a condition in the planned early-2027 AI standards, with published allocation volumes.
Expectation 5 of the Expectations for Data Centres and AI Infrastructure Developers asks operators to enable “access to compute for Australian start-ups, innovative small businesses, researchers and not-for-profits on favourable terms”. 17 That expectation is non-binding and, as far as we can determine, unmeasured. It should be legislated as a condition, with published allocation volumes, in the AI standards planned for early 2027.
Reconsider both R&D Budget measures before commencement: the ten-year refundability cap and removal of supporting activities.
Under the ten year cap, our refundability ends in 2033, at or just before the point where the underlying research finally becomes an exportable product. Removing supporting activities compounds this. In physical AI research, the experimental work and the activities supporting it are not separable in the way the drafting assumes. Building a weather rated microphone array is not core research; it is the only way to obtain the field data the core research requires. The Strategic Examination of R&D recommended simplifying supporting activities and extending eligibility to translation and commercialisation. 20 The Budget instead removed them outright. We would ask the Committee to note that divergence and to recommend both measures be reconsidered before commencement,
When reconsidering both R&D Budget measures, provide at least a deep-tech carve-out gated on assets rather than age.
Removing supporting activities compounds this. In physical AI research, the experimental work and the activities supporting it are not separable in the way the drafting assumes. Building a weather rated microphone array is not core research; it is the only way to obtain the field data the core research requires. The Strategic Examination of R&D recommended simplifying supporting activities and extending eligibility to translation and commercialisation. 20 The Budget instead removed them outright. We would ask the Committee to note that divergence and to recommend both measures be reconsidered before commencement, at minimum with a deep tech carve out gated on assets rather than age.
Mandate post-construction verification and publish the results.
Mandating post-construction verification and publishing the results would close a real regulatory gap, create a sovereign training dataset under public control and give Australian AI companies validation data they cannot currently obtain at any price.
Expand the National Industry PhD (Industry Researcher Stream) subsidy.
We host a National Industry PhD candidate under the Industry Researcher Stream, for which the Commonwealth pays the industry partner a subsidy of $47,572 per year at the 2026 rate. That is how a company our size accesses doctoral level AI research capability. We support its expansion.
Support government AI where tasks are bounded and outputs checkable, including domain models that flag environmental assessments departing from physical expectations without requiring the model to decide.
We support AI in government where the task is bounded and the output is checkable, which is exactly the case for technical assessment. A regulator reviewing an environmental impact statement is reviewing a document prepared by a proponent’s consultant, on assumptions the regulator has limited capacity to test independently. A domain model that flags where a submitted assessment departs from physical expectation is a genuine improvement in regulatory capacity and it does not require the model to make the decision.
For AI noise assessments, require independent-validation standards, a computable output confidence measure, and disclosure of the physical model and its validation; regulate for assurance rather than against the method.
AI can make assessment cheaper and faster, but regulation must determine the level of independent validation required. AI can make that evidence cheaper and faster, but only regulation can make it accountable. If AI is used in this space and it will be, the framework should require a computable confidence measure on the output and disclosure of the physical model and its validation. Regulate for assurance, not against the method.
Prioritise continual monitoring and enforcement of noise near data centres.
The most important step is continual monitoring and enforcement. Regarding noise near data centres, residents need honest and accountable action when noise affects their lives.
Allow patent, IP and copyright costs as eligible R&D expenditure for tax refunds, to protect IP ownership of AI models built in Australia.
IP ownership of AI models built in Australia needs protection. One suggestion would be to allow patent, IP and copyright costs as eligible R&D expenditure for tax refunds.
Allow a commercially validated capability to be assessed for defence application without restarting the funding process.
Letting a commercially validated capability be assessed for defence application without restarting the funding process would strengthen sovereign capability under both headings.
If the AI Safety Institute's remit is to support regulators, develop assurance methods for engineering AI.
We have been unable to identify a published Australian regulatory assurance framework specifically addressing AI-generated engineering assessments. As these tools enter professional practice, regulators will need methods for determining whether their technical outputs are reliable. If the Institute’s remit is to support regulators, developing assurance methods for engineering AI is the highest value work available.