The ATO’s use of AI has allowed it to analyse more work-related expense claims over several years, but questions have been raised about the design and explainability of who the models flag for audit.
The tax office has four AI models either deployed or in production to assess work-related expense claims.
The models are all designed and developed by the ATO’s data science function, which sit within its smarter data division, and are used to support human teams and decision-making.
One of these, a ‘substantiation risk model’ first deployed in August 2021, aims to identify the likelihood of whether work-related expenses submitted by an individual may be non-compliant.
The work-related expense models, along with a number of AI models, are the subject of a detailed analysis by the federal auditor-general.
Specifically on the work-related expense models, the audit [pdf] questions design decisions that were made, and the extent to which the ATO can be assured the models are free from biases, meet modern data ethics standards, and produce easily explainable results.
Both the auditor – and ATO – note that the AI landscape has changed significantly since some of these models were created.
The ATO has recently assessed whether the work-related expenses models comply with data ethical standards, noting there was no mandate for such an assessment when the models were first developed.
The tax office cleared the expenses models of two potential biases: that more men were flagged, and that self-prepared tax returns were more likely to attract attention than tax agent submissions.
Across the board, the ATO was urged to incorporate “ethical and legal considerations into its design and development of AI models.”
The tax body was also found to have not established a framework of “policies and procedures for the development of AI models”.
As of now, the ATO is developing an enterprise-wide AI policy and AI risk management guidance.
It is also developing guidance for reviewing models prior to deployment and “creating arrangements to support performance monitoring and review of models”.
Additionally, it intends to introduce an enterprise-wide approach to monitoring the performance of its AI models by December 2026.