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 451

Mr Stuart Bond

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.

The founder of an AI adoption service provider argues for business control over data, know-how and operations. Proposes service-chain disclosure, durable customer choices, specialised Australian capability, practical adoption support, worker learning and tested safeguards. Describes these as policy and procurement proposals, without controlled impact evidence, quantified productivity gains or claims that every safeguard is deployed.

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 businesses need to understand what each party in the AI service chain retains, derives and reuses.

    Businesses need to understand what each party retains, derives and reuses, and what happens when suppliers, owners or terms change.

  2. Argues that labelling information as metadata does not settle its sensitivity or permitted use.

    Metadata and logs may contain commercially valuable information about demand, recurring problems, processes and decision patterns. Calling information metadata does not settle its sensitivity, commercial value or permitted use.

  3. Argues that acquiring an AI model ordinarily means deployable weights and licence rights, not necessarily owning the foundation model's intellectual property.

    Acquiring a model ordinarily means obtaining suitable deployable weights and licence rights; it does not necessarily mean owning the underlying foundation model's intellectual property.

  4. Argues that small and family businesses need help integrating AI into existing systems, beyond software access.

    Small and family businesses need help connecting AI to existing systems and everyday routines. A software subscription or demonstration is only part of an operational service. Adoption support should cover workflow discovery, data preparation, integration, staff training and controlled pilots with a clear owner.

  5. Argues that a model benchmark does not establish that an integrated business service is fit for purpose.

    A model benchmark does not establish that an integrated business service is fit for purpose.

Recommendations

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

  1. Require clear, versioned disclosure of material AI processing across the whole service chain, including metadata, logs, feedback and derived information.

    1. Require clear, versioned disclosure of material AI processing across the whole service chain, including metadata, logs, feedback and derived information.

  2. Separate processing necessary to deliver and secure a service from optional reuse, and provide durable choices when purposes, providers or ownership change.

    2. Separate processing necessary to deliver and secure a service from optional reuse, and provide durable choices when purposes, providers or ownership change.

  3. Support Australian businesses to acquire and operate specialised AI capability, including appropriate licence rights, domain data, evaluations and portable configurations.

    3. Support Australian businesses to acquire and operate specialised AI capability, including appropriate licence rights, domain data, evaluations and portable configurations.

  4. Fund workflow discovery, integration, data preparation, staff training and supported pilots alongside access to software.

    4. Fund workflow discovery, integration, data preparation, staff training and supported pilots alongside access to software.

  5. Make disclosure, usable human escalation, proportionate evaluation and a practical ability to stop or roll back a service normal adoption requirements.

    5. Make disclosure, usable human escalation, proportionate evaluation and a practical ability to stop or roll back a service normal adoption requirements.

  6. Build capability in workers and managers, with employee participation and evaluation of service quality, workload and job quality.

    6. Build capability in workers and managers, with employee participation and evaluation of service quality, workload and job quality.

  7. Standardise disclosure of accountable suppliers, routing, subprocessors, information categories, locations, purposes, access, retention, security, necessary/optional uses, controls/defaults, export/deletion, irreversibility and change notification.

    I recommend a standard disclosure record covering:  The accountable application supplier, material model providers, routing arrangements and subprocessors.  The categories of information processed, including inputs, outputs, attachments, logs, feedback, embeddings and other derivatives where relevant.  Processing and storage locations, purposes, access, retention periods and security arrangements.  Which uses are necessary for the contracted service and which are optional, with the applicable controls and defaults.  What can be exported or deleted, what cannot be reversed, and how material changes are notified.

  8. Provide businesses an understandable concise explanation before adoption, backed by detailed assurance/procurement records; identify material processing changes without requiring customers to reconstruct them from multiple websites.

    Businesses should receive a concise explanation they can understand before adopting the service, supported by a more detailed record for assurance and procurement. Suppliers should identify material processing changes without requiring customers to reconstruct them from multiple websites.

  9. Apply processing-change information requirements proportionately to embedded AI and firms handling customers’ information; give small operators usable responsibilities and supply-chain information without expecting forensic investigation of every model call.

    Suppliers should identify material processing changes without requiring customers to reconstruct them from multiple websites. This should apply proportionately to embedded AI and to businesses using AI on their customers' information. A small operator needs usable information about its responsibilities and supplier chain; it should not be expected to conduct a forensic investigation of every model call.

  10. Require agreements to distinguish necessary operation from optional reuse, record choices, and give advance notice of material data-use, provider or ownership changes, with appropriate renewed consent for materially different optional uses.

    I recommend that service agreements distinguish necessary operation from optional reuse, record the customer's choices, and provide advance notice of material changes to data use, providers or ownership. Appropriate renewed consent should accompany materially different optional uses.

  11. Make export, termination and transition arrangements practically usable, with clear timeframes and costs.

    Export, termination and transition arrangements should be usable in practice, with clear timeframes and costs.

  12. Specify deletion coverage for source records, operational copies, backups, indexes, derived datasets and trained adaptations where applicable; explain limitations and distinguish stopping future processing from removing historical model effects.

    Deletion commitments should identify the objects they cover: source records, operational copies, backups, indexes, derived datasets and trained adaptations where applicable. Suppliers should explain limitations honestly. Removing a source record is not evidence that its influence has been removed from a model already trained on it. Customers should be able to distinguish removal from future processing from removal of historical effects.

  13. Include domain-data and retrieval control, adaptations, evaluations, configuration, documentation and maintained agreed versions in capability packages; a dedicated endpoint or branded private service alone does not establish those rights.

    An effective capability package should include control of domain data and retrieval, adaptations, evaluations, configuration, documentation and the ability to maintain and operate agreed versions. A dedicated endpoint or a branded private service alone does not establish those rights.

  14. Base capability choices on domain evaluation and failure testing; compare compute, energy, security, staffing, maintenance, utilisation, licensing, integration and exit costs, recognising local operation is not automatically cheaper or safer.

    Decisions should follow domain evaluation and failure testing. The comparison should include compute, energy, security, staffing, maintenance, utilisation, licensing, integration and exit costs. Local operation is not automatically cheaper or safer.

  15. Include controlled pilots with clear owners in adoption support.

    Adoption support should cover workflow discovery, data preparation, integration, staff training and controlled pilots with a clear owner.

  16. Provide staged adoption assistance starting from a defined business problem and baseline, with pilots that set what success means, what stays under human control, cost and what would justify stopping.

    I recommend staged assistance that starts with a defined business problem and a baseline. A pilot should identify what success means, what must remain under human control, what it will cost and what would justify stopping it.

  17. Include spending limits, warnings and understandable billing in evaluations of usage-based services.

    Spending limits, warnings and understandable billing should be included in evaluation of usage-based services.

  18. Expand only on sustained task-completion, rework, human-time, service-quality and operating-cost results; programme evaluations should count remaining human work and integrated-service maintenance costs.

    Expansion should depend on sustained results, including completed tasks, rework, human time, service quality and operating costs. Counting licences, demonstrations or generated outputs is insufficient. Programme evaluation should include the work that remains with people and the cost of maintaining the integrated service.

  19. Give regional and suburban businesses access to adoption support through local delivery capability and practical remote assistance.

    Regional and suburban businesses should be able to access this support through local delivery capability and practical remote assistance.

  20. Make support usable by firms without internal AI or procurement specialists.

    Support should be usable by firms without an internal AI or procurement specialist.

  21. Include employees in application selection and work design; train them to judge reliance, question and correct outputs, escalate problems and stop unsafe or failing processes.

    Employees should participate in choosing useful applications and designing how the work changes. Training should explain when to rely on a system, question its output, correct it, escalate a problem and stop an unsafe or failing process.

  22. Support workplace learning alongside technical implementation, with opportunities for staff feedback during pilots.

    I recommend support for workplace learning alongside technical implementation, with opportunities for staff feedback during pilots.

  23. Evaluate workload, rework, job quality and service quality alongside business cost and speed.

    Evaluate effects on workload, rework, job quality and service quality as well as business cost and speed.

  24. Make skills for selecting, integrating, evaluating and maintaining systems in local businesses and delivery partners a direct programme objective.

    Australian capability includes the people who can select, integrate, evaluate and maintain these systems. Building those skills in local businesses and delivery partners should be a direct programme objective.

  25. Inform people when they interact with AI, provide a straightforward human route and test it in real operating conditions, including task failure.

    People should know when they are interacting with AI and have a straightforward route to a person. In our AI receptionist work, the approach we have described is to identify the assistant, offer human transfer and take a message if nobody is available. The human pathway needs testing in real operating conditions, including when the system cannot complete the task.

  26. Issue proportionate disclosure, escalation, fallback and accessibility guidance; require stronger review, a named person accountable for exceptions, and a clear concern-raising route for higher-consequence uses.

    I recommend proportionate sector guidance that treats disclosure, escalation, fallback and accessibility as normal service-design requirements. Higher-consequence uses require stronger review and a named person accountable for exceptions, with a clear route for affected people to raise a concern.

  27. Connect use cases to purpose, users, ownership, authority limits, baselines, representative evaluation and human review, supplier/data dependencies, access controls, limited rollout, incident records, periodic review, and tested fallback, stopping, correction or rollback.

    Practical assurance should connect each use case to:  Its purpose, users, accountable owner and limits of authority.  A baseline, representative evaluation cases and human review of results.  The data and supplier chain, access controls and operating dependencies.  A limited initial rollout, incident records and periodic review.  A tested fallback and the ability to stop, correct or roll back the service.

  28. Limit action-taking systems’ permitted actions and access to their tasks; evaluate misuse, misleading instructions, connected-service failure and escalation when requests exceed authority.

    Where an AI system can take actions, its permitted actions and access should be limited to the task. A system that drafts a message and one that sends it create different consequences. Evaluation should include misuse, misleading instructions, failure of connected services and escalation when the requested action exceeds its authority.

  29. In Commonwealth procurement and service delivery, assess actual services, identify accountability and require supplier dependency explanations; scale templates with consequences to keep low-risk adoption practical and scrutinise sensitive uses more strongly.

    The same principles can inform Commonwealth procurement and service delivery: assess the actual service, identify who is accountable and require suppliers to explain their dependencies. Templates should scale with consequences so that low-risk adoption remains practical while more sensitive uses receive stronger scrutiny.

  30. Connect adoption assistance, procurement and supplier transparency in the committee's response.

    I recommend that the committee's response connect adoption assistance, procurement and supplier transparency.

  31. As a first step, develop a common disclosure template and proportionate assurance guidance with small businesses, workers, Australian delivery providers and relevant regulators.

    A practical first step is a common disclosure template and proportionate assurance guidance, developed with small businesses, workers, Australian delivery providers and relevant regulators.

  32. Run pilot programmes that fund integration and training, establish baselines and publish appropriately anonymised evaluation methods and results.

    Pilot programmes should fund integration and training, establish baselines and publish appropriately anonymised evaluation methods and results.

  33. Assess support for specialised Australian capability by domain value, lifecycle cost, portability and continuing operating rights.

    Support for specialised Australian capability should assess domain value, lifecycle cost, portability and continuing operating rights.

  34. Examine where existing legal, contractual and procurement mechanisms already support these outcomes and where gaps justify targeted reform.

    The committee should examine where existing legal, contractual and procurement mechanisms already support these outcomes and where gaps justify targeted reform.

Read the original

Open in the interactive guide