Policy in practice

Share of tracked statements whose text reports each item, month by month. Dashed rules mark the instruments’ own dates: v2.0 commencement, the policy’s six-month and twelve-month deadlines, and the AI Plan’s July 2026 date for Chief AI Officers. A statement counts from the month we first saw it, and agencies joined the tracker in waves (61 in November 2025, 24 more in June 2026), so a month-on-month move can be new entrants as much as new text; open a chart’s figures for the denominator.

A statement that promises less than it used to is easy to miss. These are the promises, self-imposed limits and public-facing safeguards that a later revision no longer makes, newest first, as read by a language model comparing the two revisions’ profiles. A promise reworded rather than withdrawn can land here, so each links to the revision, where the exact words are shown.

• 1 Aug 2026Australian Electoral CommissionCommitment dropped (will not): Staff must not use AI as part of any AEC interaction with the public
• 23 July 2026Department of Social ServicesCommitment dropped (will not): Restrict access to publicly available AI tools from our ICT environment
• 11 July 2026Net Zero Economy AuthorityCommitment dropped (will not): AI input must not include or reveal classified, sensitive or personal information
• 11 July 2026Net Zero Economy AuthorityCommitment dropped (will not): Only AI tools authorised for access and use within PM&C's ICT systems are to be deployed
• 11 July 2026Net Zero Economy AuthorityCommitment dropped (will not): Will not use AI that the public may directly interact with or be significantly impacted by, without a human intermediary
• 11 July 2026Net Zero Economy AuthorityCommitment: no public-facing AI without a human removed

Uses, domains, named tools and training claims that a later revision no longer mentions. A tool that drops out of a statement may simply have been renamed or folded into a broader description; a use that drops out may still be happening. These are changes in what the agency says, not in what it does.

• 26 Aug 2026Department of Industry, Science and ResourcesDomain dropped: compliance and fraud detection
• 25 Aug 2026Fair Work OmbudsmanChief AI Officer: in place → not mentioned
• 1 Aug 2026Australian Electoral CommissionStaff training: mandatory → not mentioned
• 9 June 2026Australian Trade and Investment CommissionUsage pattern dropped: decision making and administrative action
• 12 May 2026Australian Transport Safety BureauNamed tool dropped: Microsoft 365 Copilot
• 12 May 2026Australian Transport Safety BureauStrategic position on AI: planned → not mentioned

33 statements say a Chief AI Officer is in place and 1 say one is coming. Most statements name the post without saying which job holds it; where they do, this is who holds it (job titles only, never people). The counts below are therefore smaller than the counts above.

The Standard asks each agency to say whether the public may interact with, or be significantly affected by, its AI without a human in the loop.

Say they use no public-facing AI

65 (54%)

Use it with human review

22 (18%)

Use it without stating human review

5 (4%)

Do not address it

29 (24%)

Explicitly commit to a human intermediary

20 (17%)

The DTA’s Standard for AI transparency statements lists what every statement must contain. 58 of 121 show all eight; this is how many show each. “Date last updated” counts a date the page itself carries as well as one in the statement’s prose.

• Intentions behind AI use111 (92%)
• Use classified by DTA usage pattern or domain106 (88%)
• Public-facing use addressed95 (79%)
• Monitoring and protection measures109 (90%)
• Compliance with the policy102 (84%)
• Compliance with legislation91 (75%)
• Date last updated98 (81%)
• Public contact106 (88%)

The Standard requires statements to be “reviewed and updated” at least once a year and whenever the approach to AI changes, and to say when they were last updated. 23 statements give no date at all.