Almost every agency says when it last updated its AI statement. No two say it the same way, and that is the part that matters

Our companion piece on the same index found that the AI policy binds 95 of the Commonwealth’s 106 core entities, and that the index reconciles exactly against the register of Commonwealth entities. That answered who has to publish. This answers what happened when they did.

What the Standard asks for

The Standard sets seven minimum contents. Six are about how an agency uses AI. The seventh is administrative and is the only one a reader can verify in a second: when the statement was most recently updated. It also sets a cadence, requiring review at least once a year.

It is worth being precise about why that field exists. The Standard’s stated purpose is that Clear and consistent transparency statements build public trust and make it easier to understand and compare how government agencies adopt AI. A date is what turns a pile of pages into a set you can sort.

The good news first

Of the 116 links on the index, 7 refused our automated reader entirely, including through a real browser, so we can say nothing about them and have excluded them throughout. That leaves 109 read in full.

Almost all of them state a date. Agencies are not ignoring this. The Australian Institute of Criminology records that its statement was first published to this website on 28 February 2025 and was last reviewed and updated on 26 May 2026. The Australian Office of Financial Management says The most recent update was on 7 August 2026. The National Health and Medical Research Council publishes a three-line update history. The Productivity Commission publishes a version history with a note against each entry.

Several go further than the Standard requires. The Northern Australia Infrastructure Facility names the edition it wrote to: its statement is prepared in accordance with the Standard for AI transparency statements (v1.1) published by the Digital Transformation Agency. It is the only one of the 109 that cites a version number, and it is the single most useful sentence in the entire set for anyone trying to compare like with like.

Now the part that does not work

Here is the same fact, in some of the ways it is written across the set.

Each row is a form we found in at least one statement. This is a floor, not a census: it counts the shapes our reading identified, and the prose forms in particular vary further than a table can show.

Some of those are meaningfully different claims. Page last updated is a fact about a web page and can move when a typo is fixed. Date of effect is a fact about a document. An authorisation line naming an officer and a date dates a signature. A version history tells you what changed; a bare date does not. A reader comparing two agencies is not comparing the same thing even when both appear to have answered.

The evidence that it cannot be checked at scale is that we could not check it

This is the part we would rather not report, and it is the strongest thing in the piece.

We wrote two automated passes over the same 109 documents. The first looked for update wording near a date and found 93 with one. The second required the date to sit immediately after a recognised label and found 78. Two careful scans of identical inputs, disagreeing by 15.

So we hand-read the residue, and both were wrong in the same direction: they had undercounted. The Office of the Director of Public Prosecutions says This statement was last updated on 25 February 2026 and will be updated as our approach to AI changes, and at least every 12 months. The Inspector-General of Intelligence and Security, the Australian Electoral Commission, the Department of the Senate, the Australian Prudential Regulation Authority, Austrade, the National Capital Authority and a dozen others all say it plainly, in sentences. Not one of them is a field.

After hand-reading, the number of statements where we could not find a date anywhere on the page falls to a handful. We say “could not find” rather than “does not have” deliberately, because several of those are long pages where the statement may be a linked document we did not open. The Defence and Veterans’ Service Commission is the clearest of them: its page carries the commitment, This statement will be reviewed annually, when we make a significant change to our approach to AI, or when any new factor impacts this statement, and offers the statement itself as a download.

Why this matters more than a compliance score would

If the finding were “twelve agencies are late”, it would be a list and a news cycle. This is worse and quieter. The requirement is being met and cannot be audited.

The Standard exists so that these statements can be compared. Comparison at 116 documents is a machine job. A field can be compared; a sentence in the middle of a paragraph cannot, not reliably, as our own two attempts demonstrate. That applies to a journalist, to a researcher, to an oversight body, and to the Digital Transformation Agency itself, which maintains the index and says on it that it does not validate or endorse individual statements.

The fix is small and it is the department’s to make, not the agencies’. The Standard already prescribes the content. What it does not prescribe is the shape: a named field, in a stated format, in a stated place. Ninety-odd agencies have already done the hard part, which is the honesty. What is missing is the two lines of drafting that would let anyone add it up.

What we are not saying

We are not saying any agency is out of date. We did not test the dates against the annual review requirement, because doing that across 109 documents needs the very comparability this article is about, and we are not going to publish a number we have just finished explaining we could not produce reliably. That check becomes possible the day the field becomes a field.