What an AI Register Should Actually Record

Key points

• Australia’s minimum register fields cover owners, risk ratings and lifecycle stage, but none asks what an AI use case was measured against or what it got wrong.
• The Netherlands Court of Audit found in May 2026 that police did not systematically record what hits from one algorithm led to, so its effect could not be established.
• A register that adds a baseline, a measured error rate, a named sign-off and an outcome log turns an inventory into evidence a minister can act on.

Eighteen Australian Commonwealth departments and agencies have now handed over their internal registers of artificial-intelligence use under freedom-of-information requests, and the picture they paint is of a public service that summarises everything. The Canberra Times reported on 7 October 2026 that every responding agency uses AI to summarise documents, from meeting minutes and legislation to grant applications and community feedback. The Department of Health said it uses AI to analyse Australian National Audit Office reports into its own conduct; according to the paper, the department spent A$114m this year on a licensing agreement for Microsoft’s product suite, including Copilot. The Department of Climate Change, Energy, the Environment and Water used AI to draft synthesis reports for the next State of the Environment report. The National Health and Medical Research Council uses an app that certifies whether academic documents are wholly human-written; the Australian Securities and Investments Commission redacted entries, mostly on law enforcement and fraud detection. What the registers, as reported, do not show is whether any of it worked.

What Canberra’s Policy Requires by December

The obligation comes from the Digital Transformation Agency’s (DTA) Policy for the responsible use of AI in government. In a January 2026 update notice, the DTA said the revised policy took effect on 15 December 2025, that its first new mandatory requirement began on 15 June 2026, and that the remaining requirements take effect in December 2026. It applies to non-corporate Commonwealth entities, with the defence portfolio and national intelligence community excluded. Under the policy’s strategy and oversight requirements, agencies must create an internal register of in-scope use cases, and name an accountable owner for each, within 12 months of the policy taking effect, and share the register with the DTA every six months thereafter. The Canberra Times, which calls the agency the Digital Transformation Authority, reported that the policy came into force on 15 June, with registers finally due in December; the DTA’s own notice dates the revised policy from 15 December 2025 and puts the first new mandatory requirement on 15 June 2026.

The Standard for accountability lists 13 minimum fields: use case name, agency reference number, description, technology type, lifecycle stage, use of the government’s technical standard for AI, domain, usage pattern, accountable owner, the criteria that brought the use case into scope, inherent and residual risk ratings, and the date the impact assessment was last updated. High-risk use cases must also carry last and next review dates. That is a competent inventory of who owns a tool and how risky it was judged. It does not record what the tool was compared with, how often it was wrong, or what decisions it fed.

The public layer is thinner still. The DTA published a central register of transparency statements on 31 March 2026, with 94 agencies meeting their obligations and 20 publishing voluntarily. Health’s AI transparency statement says the department does not use AI to automate decisions and claims benefits such as efficiency, but cites no measure of them.

How Australia’s Fields Compare With London’s and The Hague’s

Two older schemes show what else can be recorded. Britain’s Algorithmic Transparency Recording Standard (ATRS) is, according to the government’s guidance for public-sector bodies, mandatory for all government departments and for arm’s-length bodies that deliver public services or deal directly with the public. The Netherlands runs a public Algorithm Register, which listed more than 1,500 algorithm descriptions when this desk opened it on 11 October 2026. The table below, built from the three governments’ own documents and, for the Dutch case, from a published entry such as the Tax and Customs Administration’s Dynamisch Monitoren description (in Dutch), compares what each actually asks for.

Field	Australia (DTA minimum register)	UK (ATRS)	Netherlands (Algorithm Register, publication standard 1.0)
Accountable owner	Named owner, with email	Senior responsible owner (role)	Organisation and contact details
Purpose and description	Yes	Yes, with rationale	Yes (“Doel en impact”)
Risk rating or assessment	Inherent and residual rating	Risks, mitigations, links to impact assessments	Risk management; impact assessments
Human role in the decision	Usage pattern only	Degree of automation; appeals and review	Dedicated human-intervention field
Process it replaced	No	Yes, plus alternatives considered	Weighing of benefits
Measured performance	No	Model performance (e.g. precision, recall)	No dedicated field
Outcomes of use	No	No	No
Review schedule	High-risk only	Update on substantive change	Last-modified date
Public?	No; shared with DTA	Yes	Yes

Only the British template asks for a performance number, and its scope stops short of most of Canberra’s examples: the guidance says it targets tools that significantly influence operational decisions with public effect, not analytical models supporting broad policymaking. None of the three asks what happened after the system produced its output.

Why Outcome Data Is the Missing Field

That gap is not hypothetical. On 20 May 2026 the Netherlands Court of Audit (Algemene Rekenkamer) published an audit of two security-domain algorithms (in Dutch; translation by AI & GPP). It found that the police managed most risks of their iTrechter sensor-analysis system well or partly, but did not systematically track what its hits led to, so the auditors could not establish how many were followed up or produced a result. Customs could not show that the risk profiles in its strategic-goods screening algorithm were effective, and had not followed its own evaluation procedure. Both systems could have filled in every field above and still failed that test.

The Ada Lovelace Institute reached a parallel conclusion from the other end of the pipeline. Its 8 September 2026 report with the National Institute for Health and Care Excellence (NICE), which examined a protocol for using a large language model to screen research abstracts, found “no standardised or organisation-specific approaches to ethically adopt AI”. It called for evaluation that runs end to end and repeats as models change, for a clear statement of “how success is measured and against what baseline”, and for an organisation to set its risk appetite before testing and piloting begin. Those are precisely the fields an inventory leaves blank.

A register of tools tells a minister what the department owns. A register of decisions tells her whether it was worth owning.

In this desk’s reading, the distinction matters most for summarisation, the one use every Canberra agency reported: a summary is rarely the decision itself, yet it shapes what an official reads before deciding. Evidence arriving too late to inform choices is the subject of the long read The Decision Gap; the gulf between what governments disclose and what citizens believe they are told is examined in The Transparency Gap.

A Minimum Schema That Records Decisions

The schema below is this desk’s illustration: it keeps the DTA’s fields and adds those that make a register an evidence base. Example values are invented and describe no real agency.

Field	What it records	Why it matters	Example entry (illustrative)
Use case	Task in plain words	Scope of the record	Summarise public consultation submissions
Decision supported	The decision the output informs, and who takes it	Links the tool to accountability	Branch head’s advice to minister on draft regulation
Accountable owner and sign-off	Named owner; who approved go-live, and when	Someone answers for it	Director, Policy Branch; signed 3 March
Data sources	Inputs and their provenance	Errors start in inputs	412 submissions received via consultation portal
Baseline	Performance of the previous method, measured before AI	No baseline, no claim of improvement	Manual summaries: 96 staff-hours; 4% of points missed in audit sample
Success metric and risk appetite	Target, and tolerable error set before the pilot	Prevents moving the goalposts	Under 5% of material points missed
Measured error	Result of a documented check against source	The number a minister needs	50-summary sample: 3 material omissions (6%)
Evaluation dates	Last and next check	Models and prompts change	Checked 14 April; next due 14 October
Incidents	Errors that reached a decision, and response	Learning, and notification	One misattributed submission, corrected before advice sent
Outcome tracking	What happened to the decisions the output informed	The missing field above	Advice accepted; two summaries challenged by submitters
Retirement criteria	Conditions to pause or retire	Ends drift	Error above 8% in two consecutive checks

The same record, in the structured form an agency could share every six months:

{
"use_case": "Summarise consultation submissions",
"decision_supported": "Advice to minister on draft regulation",
"owner": "Director, Policy Branch",
"baseline": {"method": "manual", "missed_points_pct": 4},
"risk_appetite": {"max_missed_points_pct": 5},
"measured_error": {"sample": 50, "material_omissions": 3},
"next_evaluation": "2026-10-14",
"outcome": "Advice accepted; 2 summaries challenged",
"retire_if": "missed_points_pct > 8 in 2 consecutive checks"
}

(Illustrative figures; not a real agency’s entry.) Here the tool missed 6 per cent of material points against a target of 5 and a manual baseline of 4: an inventory would show the use case as approved and operating, while this record shows it is not yet meeting its own bar.

What Accountable Officials Should Do Before December

Agencies finishing registers for December need not wait for the DTA: the standard’s fields are stated as a minimum. Three additions carry most of the value: a baseline measured before the tool is switched on, a documented error check with a date, and an outcome field that follows the decision rather than the tool. Incident fields will also matter as governments write notification duties for AI failures, the subject of this issue’s cover story on incident notification rules.

Watch whether the DTA’s next revision of the accountability standard adds evaluation fields, whether agencies publish register extracts, and whether audit offices, as in The Hague, test registers against outcomes rather than taking entries on trust.

Questions this answers

When must Australian Commonwealth agencies complete their internal AI use case registers?
Within 12 months of the revised policy taking effect on 15 December 2025, with the DTA saying the remaining requirements take effect in December 2026. Registers must then be shared with the DTA every six months, according to the policy’s strategy and oversight requirements.

Do government AI registers record whether the AI is accurate?
Only the UK’s ATRS asks for performance metrics, and none of the three schemes records what happened after the system’s output, the gap the Netherlands Court of Audit found in police use of iTrechter.

Sources

1. Rise of the machines: bureaucrats turn to AI to write and decode policy — The Canberra Times — 7 October 2026; (paywall)
2. AI policy update: strengthening responsible use across government — Digital Transformation Agency — 12 January 2026
3. Strategy and oversight (Policy for the responsible use of AI in government) — Digital Transformation Agency, digital.gov.au — undated web page, accessed 11 October 2026
4. Standard for accountability — Digital Transformation Agency, digital.gov.au — undated web page, accessed 11 October 2026
5. New central register for AI transparency statements for Commonwealth entities — Digital Transformation Agency — 31 March 2026
6. Artificial intelligence (AI) transparency statement — Australian Government Department of Health — date not shown, accessed 11 October 2026
7. Algorithmic Transparency Recording Standard: guidance for public sector bodies — GOV.UK — updated 8 May 2025
8. Het Algoritmeregister van de Nederlandse overheid — Government of the Netherlands — accessed 11 October 2026; (in Dutch)
9. Dynamisch Monitoren — Algoritmeregister, Belastingdienst — last modified 9 October 2026; (in Dutch)
10. Inzet algoritmes niet altijd effectief en veilig — Algemene Rekenkamer — 20 May 2026; (in Dutch; translation by AI & GPP)
11. Care and consideration: Lessons for AI adoption in the public sector from a research collaboration with NICE — Ada Lovelace Institute — 8 September 2026

Discussion

No comments yet. Start the conversation.

Discussion is open to members.

Join free Sign in