About the Bureau
We read the world's government-AI reporting each weekday and publish the few items that matter. Our focus is Australia, about a quarter of what we file.
GovAI Bureau is a daily record of what governments are doing with artificial intelligence: building it, buying it, governing it, and getting it wrong.
It is published by Salsa Digital, and it exists because there was no single place to follow this subject. Government AI is reported in fragments across trade press, policy media, agency announcements and parliamentary records in a dozen countries. The Bureau reads those sources every weekday and keeps the record in one place.
It is a wire service, not a commentator. It reports what others have reported and links to the original. It does not make recommendations, does not run opinion, and does not publish its own investigations. If you want to know what happened, this is the index. If you want to know what it means for you, that judgement is yours.
Public servants and agency leaders who need to know what peer governments are doing. Policy and digital teams tracking obligations as they form. Anyone who has to brief someone else on this subject on short notice.
Reading it costs nothing and requires no account. There is no newsletter to join and nothing is sold here.
The beat is government's own use of AI. An entry qualifies when a public body is doing something with AI:
- Building, buying, procuring, deploying or retiring an AI or algorithmic system
- Accountability, audit, review or redress for a government algorithm or automated decision
- Assurance, impact assessment, transparency obligations and governance frameworks applying to public bodies
- Public servants' own use of AI, capability and skills in the public service, chief AI officers
- Vendors selling AI to government, and the procurement and sovereignty questions that raises
- A government AI plan, policy or strategy governing how that government uses AI
Out of scope, even where the words look right: AI regulation aimed at private industry with no public body's own use at issue; AI and the labour market; complaints about a government service that merely mention AI; academic, classroom or corporate AI policies; general model and capability news with no public-sector actor.
Bureau summaries are machine-generated. A large language model reads the day's candidate articles and writes a summary of roughly forty words for each item selected. No individual journalist writes them, and no individual reviews each one before it is filed.
The model works to fixed editorial rules: report rather than advise, name the actor, attribute contested claims to the source rather than asserting them, and never use the first person. Summaries are the Bureau's own words, not extracts from the original article, and are deliberately set in a lighter tone than headlines so that Bureau writing is never mistaken for a publisher's reporting.
The Bureau publishes this because the beat is government transparency about AI. A publication covering that subject while concealing its own use of it would not deserve to be read.
The limits are real. A machine-written summary can misread emphasis, flatten a qualification, or state something more confidently than the source did. The original is always linked and always governs. Where the two disagree, the original is right.
Around 88 sources are read every weekday: government and agency publications, trade press, policy media, and keyword alerts. Roughly 120 candidate items surface daily.
Items reaching the archive index have passed relevance filtering only. They are unedited, and their presence is not an endorsement of the reporting or a judgement about its importance.
Items reaching briefings have additionally been selected as among the three to five most significant of that day and carry a Bureau summary. Selection weighs policy and regulatory significance, notable public-sector adoption, and substantive commentary. Vendor marketing, generic hype and near-duplicates are excluded. Fewer than three items on a given day is a normal outcome.
Topic tags on the historic archive were assigned by keyword matching against headlines and summaries. This is approximate. Around a third of entries match no topic strongly and are tagged General, and an entry may carry a tag that a human reader would not have chosen.
Tags are a navigation aid, not a classification scheme, and should not be relied on as a complete index of a topic. Tagging is being moved into the daily selection process, where the same model that writes the summary will assign the topic.
Jurisdiction codes follow ISO 3166-1 alpha-2. Entries without a determinable jurisdiction show an em dash rather than a guess.
The Bureau covers government AI failures the same way it covers government AI successes: by reporting what has been reported elsewhere and linking to it. A public body is named when the source names it.
The Bureau does not conduct its own investigations, does not allege wrongdoing, and does not editorialise about the bodies it covers. Where reporting is contested, the contest is reported rather than resolved.
No agency is excluded from coverage because of a commercial relationship with the publisher, and no agency receives advance notice of an entry. An index that quietly omitted certain bodies would not be a record.
Agencies that consider an entry inaccurate can raise it using the contact below. Corrections are handled under the policy above and are published as corrections.
Every entry names its publisher and links to the original. The Bureau does not republish article text and does not reproduce more of a source's substance than a short summary requires. Readers are expected to follow the link; the Bureau is an index to reporting, not a substitute for it.
Headlines and the underlying reporting remain the property of the publishers named.
Bureau summaries — the short passages written by the Bureau, not the headlines or the source articles — are published under Creative Commons Attribution 4.0 ↗. You may copy, republish, translate or build on them, including commercially, provided you credit GovAI Bureau and link back to the entry.
Suggested credit: Summary by GovAI Bureau, CC BY 4.0.
Any publisher who would prefer their material not to be indexed here can say so using the contact below, and it will be removed.
Corrections are published as corrections. When a Bureau summary is found to be wrong, the entry is amended, the change is dated, and what it previously said is recorded. Entries are not silently edited and are not deleted.
Links to original reporting are checked each weekday for the items selected that day, and the result is recorded against the entry. That record is not currently shown on the site. Older entries have not been checked, and link rot across an archive of this age is expected.
The most frequently indexed publishers, by volume. The full source list runs to more than 1,300 outlets.
The Mandarin
Government Technology
FedScoop
Australian Financial Review
StateScoop
Global Government Forum
OpenGov Asia
Civil Service World
Federal News Network
ABC News
The Guardian
Nextgov/FCW
iTnews
Reuters
The Register
Canberra Times
Mirage News
The Australian
Route Fifty
TechCrunch
Politico
Public Sector News
The Bureau aims to meet WCAG 2.2 Level AA. The site is plain HTML, works without JavaScript for reading, uses semantic headings and landmarks, and does not rely on colour alone to convey meaning.
It has not yet been independently audited, and this statement will be updated when it has. If you encounter a barrier, report it using the contact below and it will be treated as a defect.
Corrections, takedown requests, accessibility problems and questions about method all go to the same address. The Bureau does not have a named editor; correspondence is handled by Salsa Digital on the Bureau's behalf.
1.0 03 Sep 2026 Public launch. Front page rebuilt, accessibility audited to WCAG 2.2 AA, site made indexable.
0.5 03 Sep 2026 Editorial position on covering named agencies published. Link checking described accurately.
0.2 31 Aug 2026 Methodology published. Attribution corrected to publisher rather than byline. Topic tags added to the archive.
0.1 31 Aug 2026 First prototype. Archive of 4,440 entries from 25 June 2026.