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 294

Ms Tahsin Alamgir Kheya

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.

Argues the regulatory gap is the practical difficulty of testing whether AI systems produce biased or unfair outcomes, rather than simply an absence of legal obligations. Makes five narrow recommendations on disaggregated reporting, fairness-criteria transparency, evidence retention, independent government measurement capability, and transparency and choice in recommendation systems.

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 the regulatory gap concerns testing AI outcomes against legal obligations, rather than simply an absence of obligations.

    The regulatory gap addressed by this submission is therefore not simply an absence of legal obligations. Rather, it concerns the practical difficulty of determining whether an AI system is producing outcomes that are inconsistent with those obligations.

  2. Argues that frameworks relying on developer self-assessment against broad principles may yield documentation regulators cannot independently verify or compare.

    A regulatory framework that relies primarily on developers assessing their own systems against broad, high-level principles risks producing documentation that is difficult for regulators to independently verify, difficult to compare across systems, and potentially incapable of detecting the forms of harm that are most consequential in practice.

  3. Argues that an obligation that cannot be tested cannot be effectively enforced.

    The underlying point is straightforward: an obligation that cannot be tested cannot be effectively enforced.

  4. Argues that aggregate measures can conceal different outcomes across demographic groups and content categories.

    AI systems can produce materially different outcomes across demographic groups and content categories even when their overall performance appears acceptable. In recommendation systems, for example, aggregate measures can conceal differences in what particular groups are shown and how those patterns develop over time.

  5. Argues that recommendation systems can progressively narrow the content and opportunities shown to users through feedback loops.

    Large-scale recommendation systems can progressively narrow the content, information, and opportunities presented to users because models learn from previous engagement and then use those patterns to determine subsequent recommendations. This can create feedback loops in which existing behavioural patterns and biases are reinforced over time.

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 appropriately disaggregated AI performance, safety or fairness reporting by affected group and content category; aggregate measures alone are insufficient where they may conceal materially different outcomes.

    Any regulatory obligation to report on the performance, safety, or fairness of an AI system should require results to be reported at an appropriately disaggregated level, including by affected demographic or user group and relevant content category. Aggregate, system-level measures alone should not be considered sufficient where they may conceal materially different outcomes across groups or categories.

  2. Require fairness assessments to identify the criteria applied, explain why they were chosen, and disclose alternative criteria the system was assessed against but did not satisfy.

    Where an AI system is subject to a fairness assessment, the assessment should clearly identify the fairness criterion or criteria applied, explain the rationale for selecting those criteria, and disclose any relevant alternative criteria against which the system was assessed but did not satisfy.

  3. Require developers and deployers in regulated contexts to retain evaluation artifacts supporting compliance claims, including, where appropriate, datasets, scoring and evaluation code, methodological documentation and raw outputs.

    Developers and deployers of AI systems operating in regulated contexts should be required to retain the underlying evaluation artifacts supporting any claim of compliance. These should include, where appropriate, the evaluation or test datasets, scoring and evaluation code, methodological documentation, and raw model outputs.

  4. Make retained evaluation materials available to the relevant regulator on request, to enable independent verification of compliance claims.

    Such materials should be made available to the relevant regulator upon request to enable independent verification of compliance claims.

  5. Invest in independent government technical capability to measure and evaluate AI performance, fairness and risks; says the AI Safety Institute is well placed, subject to resourcing, expertise, independence and regulatory access.

    The Australian Government should invest in and maintain independent technical capability for measuring and evaluating AI system performance, fairness, and associated risks, rather than relying predominantly on developer-led self-assessment. The Australian AI Safety Institute is well placed to provide such capability, subject to appropriate resourcing, expertise, independence, and regulatory access.

  6. Require platforms with large-scale algorithmic recommendation systems to report on the diversity of content to which their systems expose users, including measures of exposure over time.

    Platforms operating large-scale algorithmic recommendation systems should be required to report on the diversity of content to which their systems expose users, including appropriate measures of content exposure over time.

  7. Provide users with a meaningful alternative to algorithmically personalised ranking, such as non-personalised or independently controlled ordering.

    Users should also be provided with a meaningful alternative to algorithmically personalised ranking, enabling them to access content through a non-personalised or otherwise independently controlled ordering mechanism.

Read the original

Open in the interactive guide