Publishing every algorithm government uses on the public has been mandatory since February 2024. The department running the register still cannot say how many are missing from it.
Estimated reading time: 6 minutes
.jpg){kind=link}
In short. The Algorithmic Transparency Recording Standard requires a public record before government uses an algorithm or AI system that affects the public. It has been mandatory across every government department since 6 February 2024. The policy document that created the mandate contains no section on enforcement, penalties or review. The department running it, DSIT, told Parliament's Public Accounts Committee in March 2026 that its own central register was not complete, and no document anywhere publishes how many algorithmic systems are actually meant to be on it. Two documented mistakes: four stars.
On 6 February 2024, the Department for Science, Innovation and Technology told every government department it now had to do something no department had previously been compelled to do: publish, in a public register, a plain-English record before using an algorithm or AI system that has a significant influence on a decision affecting someone, or that interacts directly with the public [1]. The standard itself, the Algorithmic Transparency Recording Standard, had existed since 2021 and had been a government-endorsed but voluntary approach since September 2022 [2]. What changed on that date was that it stopped being optional. Two years and seven months later, the register it produced is the standard's own test of whether the mandate actually works.
The claims, tested
| The claim | What the department's own documents say | Verdict |
|---|---|---|
| "The ATRS is mandatory for all government departments, and for arm's-length bodies (ALBs) which deliver public or frontline services, or directly interact with the general public" [3] | The policy document that defines the mandate's scope and exemptions contains no section addressing enforcement, penalties for non-compliance, monitoring, or review of departments that do not publish [1] | Mandatory in name, with nothing written down for what happens if a department ignores it |
| The government's own commitment to Parliament: DSIT was "aiming to publish records for all currently identified in-scope algorithmic tools (as of March 2025) in government departments (not including ALBs) by the end of 2025" [4] | The Public Accounts Committee found that "as at February 2026 the hub... recorded 110 being used in central government," two months after that deadline, and that DSIT itself "acknowledged that the ATRS hub was not a complete record" [5] | The end of 2025 target passed unmet, on the department's own evidence to Parliament |
| DSIT told the Committee it "was confident it had recorded the majority of cases" [5] | No gov.uk page, including the hub, the scope and exemptions policy, or the live register itself, publishes a total figure for how many algorithmic systems are actually within the mandate's scope [3] [6] | "Majority" cannot be checked against a number that has never been published |
| Before the mandate, an NAO survey found only 8 of 32 government bodies with deployed AI were "always or usually" compliant with the standard, and 38 percent were only "rarely" compliant [2] | Two years after the standard became compulsory, the live register shows genuine growth, 110 records in February 2026 to 152 today [6] | CLEARED on direction of travel: the count is rising, even if nobody can say against what target |
The mistakes, counted
A mandate with nothing written down for what happens if you miss it (1). The document that actually defines who has to publish and when, the ATRS Mandatory Scope and Exemptions Policy, sets out which departments and arm's-length bodies are covered and which systems are exempt, in detail [1]. It does not contain a single sentence on what happens to a department that simply does not publish. There is no penalty, no escalation route, and no named review point at which non-compliance gets checked. A mandate with a scope but no consequence is a strong instruction with no way to know if it was followed, which is exactly the shape of the problem the Public Accounts Committee found in practice.
A deadline the department quietly missed, and nobody outside government can measure by how much (2). In June 2025, responding formally to the Committee, the government committed to publishing records for every algorithmic tool identified as in scope by March 2025, in central departments, by the end of that year [4]. That deadline passed. By the Committee's next report, in March 2026, DSIT had told MPs directly that the register was still incomplete, while also telling them it believed it had captured most cases [5]. Both statements might be true. Neither can be checked, because DSIT has never published the one figure that would let anyone else check them: the total number of algorithmic systems actually within scope of the mandate. A register that cannot say how incomplete it is has swapped a compliance target for a promise.
Credit where due
The register is not standing still. It held 110 entries for the whole of central government in February 2026 and 152 today, seven months later, a real rate of growth rather than a static list [6]. DSIT has also not tried to hide the gap: it gave the Committee a direct account of why some records are harder to publish than others, telling MPs that "departments find the process challenging as they do not want to provide information that might help fraudsters" [5]. That is a genuine operational tension, not a fig leaf: a public record explaining exactly how a fraud-detection algorithm flags claims is also a manual for anyone trying to beat it. The Committee did not accept that as a reason to stay incomplete indefinitely, and neither should a reader, but it is a real dilemma rather than an excuse invented for this piece.
Verdict
Four stars, from two documented mistakes. The Algorithmic Transparency Recording Standard asks a reasonable question of government: before you let an algorithm shape a decision about someone, tell the public it exists. Making that mandatory in February 2024 was the right call. What government has not done, in the two and a half years since, is build a way to check whether the mandate is being met. There is no penalty in the policy for a department that ignores it, no published figure for how many systems should be on the register, and no way to distinguish "DSIT is confident it has recorded the majority of cases" from a department guessing. The growth in the register from 110 to 152 records this year is real and worth crediting. It is also, on the public record, the only number anyone outside government can actually check.
Sources
- Algorithmic Transparency Recording Standard: mandatory scope and exemptions policy, GOV.UK / DSIT, published 17 December 2024
- Use of artificial intelligence in government, National Audit Office, HC 612, Session 2023-24, 15 March 2024
- Algorithmic Transparency Recording Standard Hub, GOV.UK, published 5 January 2023, last updated 8 May 2025
- Treasury Minutes: Government Response to the Committee of Public Accounts on the Eighteenth to the Twenty-Second reports from Session 2024-25, CP 1341, HM Treasury, June 2025
- Government use of data analytics on error and fraud, House of Commons Public Accounts Committee, HC 891, 23 March 2026
- Find out how algorithmic tools are used in public organisations, GOV.UK, accessed 15 September 2026