Reading the data

This site turns a stack of government web pages into summaries, profiles, charts and colours: a one-line reading of every change, a report card for every statement, adoption curves for the policy’s obligations, originality figures, a reuse heat-map, and a record of which statement we saw a passage in first. This page explains what each one means, and what it does not claim.

When the record starts

Continuous tracking opens on 11 Nov 2025, the week the Australian Public Service launched its AI Plan. Ben Swift began the tracker from the audience, opening a laptop to archive the agencies’ statements then and there. That is why the record starts exactly when it does.

The Policy for the responsible use of AI in government has required these statements since February 2025. Their first nine months, the period when agencies drew their wording from one another, fell before the tracker was watching. Everything that already existed enters the record together on day one, marked “first tracked” rather than “published”. We cannot know what those statements looked like before we first saw them. This is the single biggest limit on everything below.

Shared passages and the reuse heat-map

Agencies reuse one another’s wording heavily. On each statement page, passages carry two different tints. An amber tint means the same passage appears in other agencies’ statements; the deeper the amber, the more widely it is shared:

uncoloured: unique to that statement;
a few: shared with two or three other agencies;
several: shared with four to nine;
many: shared with ten to twenty-four;
most: shared with twenty-five or more.

An amber passage also carries a small count — open it to see exactly which other agencies use the same words, each linked to its statement.

A separate cool grey tint marks template wording: phrasing lifted straight from the Policy or the Digital Transformation Agency’s template, the scaffolding most agencies start from. It gets its own colour, not a faint amber, because it is boilerplate rather than a passage we can trace to particular agencies — so it has no count to open.

First observed in our corpus

For every shared passage we also record which tracked statement showed it earliest, and the order in which the others picked it up. Each passage falls into one of three tiers, by how much weight that ordering can bear:

added during tracking (57): we watched the first agency add the passage well after the record opened, with others following later. The strongest signal.
present at the start (92): the first agency already had the passage when tracking began, and others adopted it later. We can order the latecomers, but not that first agency’s own source.
tied (45): several agencies share the earliest date, usually because they were all present on day one, so we make no claim about order.

Read this as first observed by us, never as proof of who wrote it first. A passage two agencies both carried on day one cannot be ordered at all, and even a clean ordering only says who we saw with it first. The real source could be an agency we don’t track, an internal draft circulated off the web, or the DTA template itself.

Originality scores

A statement’s originality score is the share of its text, by length, that is neither shared with other agencies nor drawn from the template. A statement at 80% is mostly in its own words; one at 20% is mostly shared or templated language. The leaderboard and the agency grid colour each statement on this scale, from borrowed to bespoke.

The DTA scores low by design: it publishes the template, so almost everything it says is, by definition, shared. We label it the template source rather than reading its low score as a lack of effort. Two other things push a score down without saying anything about effort. Shared text includes wording the Standard itself calls for (the document’s title, its section headings, the OECD definition of an AI system that the policy adopts), so an agency that follows the Standard to the letter shares more than one that ignores it. And a statement two bodies publish jointly (the AASB and AUASB, for instance) is shared with itself. Read the score as “how much of this text appears elsewhere”, not as a grade.

How changes are read

The scrape records every time a statement’s text changes, and most of those changes are not the agency saying anything new: a rotating sidebar, a “last reviewed 3 days ago” counter ticking over, a page rebuilt with the same words. So every revision is classified from its diff. Mechanical rules settle the obvious cases (formatting only, links only, page chrome, a date stamp, passages reordered). Everything else is read by a language model (Claude) that is shown the diff and asked to sort it into one of five kinds and to write a one-sentence summary in plain English:

scrape noise: nothing the agency wrote changed;
cosmetic: wording, typos, formatting, link targets, a date;
expanded: the same substance at more length, such as a bullet grown into a paragraph or an example added to an existing claim;
restructured: reorganised or reworded, claims and commitments substantially the same;
substantive: a claim, commitment, fact or disclosure added, removed or altered.

Only the last three count as changes on the timeline and in the story on each statement page. The model is told to report removals and additions alike, that an expansion which keeps the same substance is never substantive, and that text which vanished because the capture was cut short is scrape noise, not an edit. Every summary names the model that wrote it and is shown next to the raw diff, so you can check its reading against the words.

One further guard sits outside the model. A capture we know to be broken (a page whose body is rendered by a script the scraper cannot run, a sentence our own cleanup once ate) is listed in the repository and dropped before anything reads it, and a fresh capture that lost more than half its text is held off the site until it is confirmed or the scraper is fixed. Dropping a capture can shift the date we observed a genuine change by a few weeks; we would rather that than a false change.

Statement profiles

Each revision of each statement is also read into a profile: a fixed set of questions with fixed vocabularies. Which DTA usage patterns and domains does it say are in use? Does it address public-facing AI, and does it commit to a human intermediary? Is an accountable official named? A Chief AI Officer, in place, planned, or not mentioned? A use-case register? Mandatory training? How often does it promise to be reviewed, and when does it say it was last updated? What does it explicitly promise not to do?

Where the questions come from

The questions are the instruments’ own wherever an instrument asks one, and ours where none does. The usage patterns and domains are the DTA’s classification system verbatim. Nothing here is a requirement unless its source says so: the Chief AI Officer, in particular, comes from the APS AI Plan, not the policy.

DTA Standard for AI transparency statements v2.0 (minimum content): Intentions behind AI use
Usage patterns and domains in use (Attachment A classification)
Whether the public interacts with or is affected by AI without human review
Measures to monitor effectiveness and protect the public
Compliance with the policy
Compliance with legislation
When the statement was last updated
A public contact
Policy for the responsible use of AI in government v2.0 (mandatory requirements): Review cadence (annually, or sooner on a significant change)
Accountable official designated
Strategic position on AI (due within 6 months of v2.0)
Internal AI use-case register (due within 12 months)
Mandatory staff training (due within 12 months)
AI Plan for the Australian Public Service 2025 (Department of Finance): Chief AI Officer (due July 2026)
This tracker's own reading, not required by any instrument: An explicit commitment to a human intermediary for public-facing AI
Named safeguards, named tools, explicit commitments
Which policy version the statement refers to; a stated first-published date

Because the vocabularies are closed, two agencies’ profiles compare directly, and two revisions of one agency’s profile can be diffed: that is where the “commitment dropped” and “Chief AI Officer: planned → in place” markers come from. A bullet point expanded into a paragraph about the same thing produces no profile change at all. The profile records only what the statement itself says, never what is typical, and it records job titles, never people’s names.

The reading is a model’s, and it can be wrong: a commitment paraphrased two ways across revisions may read as dropped and re-added; a role described vaguely may land in the wrong box. To limit the drift, each revision is read with the previous revision’s profile as its baseline, and changes are reported only on revisions whose diff was judged a change of substance. Every profile sits beside the statement it came from, and every change beside its diff, so both can be checked.

Adoption over time

The policy in practice page counts, for each month since tracking began, how many tracked statements’ profiles report each item (a Chief AI Officer, a register, mandatory training, a reference to version 2.0) and shows that as a share of the statements tracked that month. Three cautions. Early months have small denominators: only 121 statements are tracked now, and fewer were on day one. Agencies joined in waves, so a month-on-month move can be new entrants rather than new text. And a statement that mentions a register is not proof the register exists; the charts show what agencies say, which is what a transparency statement is for.

A statement’s own date is the one its page carries (“last updated 25 February 2026”), read by the scraper before that line is set aside so it cannot churn the diffs, or failing that a date the model finds in the prose. “Updated since policy 2.0” is true when either that date or a change we observed falls on or after 15 December 2025, and false when we have been watching since before then and have seen neither. For a statement that entered the tracker after that date and gives no date of its own we cannot tell, and say so. “Overdue” uses only the statement’s own date: the Standard requires review at least yearly, and a self-declared date more than a year old is the agency’s own admission. A statement with no date is neither. Both verdicts are dated, because they move with the calendar.

The fine print

The About page covers how all of this is computed: how spurious scrape churn is collapsed out, the exact passage-matching rules, and which model does the reading. It also sets out the limits of the scraping itself.

One limit is worth keeping in mind. Because we read each statement straight from a live web page, what we capture is sometimes the page rather than the statement, and the classifier’s job is to tell the two apart. It is good at it, not perfect. Follow the source link on any statement to check the original.