Publishing Algorithmic Transparency Records
The UK Government has introduced the Algorithmic Transparency Recording Standard (ATRS). This standard requires public bodies to publish structured records describing any automated system used to make individual decisions affecting members of the public. The record is not a free-form description; it follows a specific format that allows citizens and researchers to compare algorithms across government.
An algorithmic transparency record describes what the system does, who it affects, what decisions it supports, what data it uses, how accurate it is, how it was tested for bias, and what human oversight exists. The record is published online in a format citizens can access and search. Some departments publish their records on gov.uk. Others use their own websites or central repositories.
The record includes mandatory fields. The system name, the organisation responsible, the date it went live, and the date of last update are basic identifiers. The purpose of the system must be stated clearly. “To allocate housing” is too vague. “To prioritise properties for disabled adaptations based on severity of need and queue length” is specific enough for oversight.
The affected population must be named. “All residents of the borough” is clear. “Potentially vulnerable people” is too vague. The record must state how many people a year are affected by the system and what decision is made. A transparency record for a benefits system might state: “This system determines whether applicants for housing benefit receive discretionary hardship payments. In 2023, the system assessed 2,400 cases and recommended hardship payments in 480 cases.”
The record describes the data used. What variables does the algorithm consider? Are they legally relevant to the decision? A housing allocation algorithm might use housing need (relevant), queue position (relevant), local connections (relevant), but if it also uses postcode (potentially problematic as a proxy for race or other protected characteristics), this must be disclosed and justified.
The record includes performance metrics. What is the system’s accuracy? Accuracy can be misleading, so the record should include precision and recall for key decision types. For a benefits system, recall (the proportion of eligible cases the system correctly identified) may be more important than precision. The record should state whether accuracy has been tested separately for different demographic groups.
The record discloses what testing has been done for bias. Has the system been tested to identify disparate impact on ethnic minorities, disabled people, women, or other groups? What was found? If disparate impact was found, how was it addressed? If bias testing has not been completed, the record should say so. Transparency means admitting what you do not know.
The record describes human oversight. Are decisions made by the algorithm final, or do caseworkers review them? If caseworkers do review, how often do they change the algorithm’s recommendation? If caseworkers rarely override the system, this suggests high algorithmic autonomy and high risk. The record should disclose this.
The record includes contact information for citizens who want to challenge or question the system. This might be a dedicated email address or the standard complaints process for the organisation. It should be simple for a citizen to find out how to query a decision made by the algorithm.
Publishing a transparency record is not optional. The standard applies to any automated system that makes or materially informs individual decisions about benefits, housing, enforcement, or other matters affecting the public. A system that shortlists candidates for a public sector job is in scope. A system that filters freedom of information requests is borderline but likely in scope if it affects which requests are responded to.
Record publication creates accountability. Journalists, MPs, researchers and advocacy organisations read these records. If a record reveals that an algorithm has never been tested for bias, public pressure will build. If a record shows that a system makes errors more often for disabled people, campaigners will question why it is still in use. This scrutiny improves algorithm design.