Transparent AI in government: what good evidence looks like

By AiVACO · 28 October 2026 · 7 minute read

In short

Transparency in government AI is not a policy statement. It is the ability to show, for a specific case, what the AI was given, what it relied on, what it produced and who made the decision. Good evidence is attributable, contemporaneous, complete (refusals included), tamper evident, reconstructable and understandable to someone who is not a technician.

What does transparency actually require?

A public body can be asked, by a citizen, a court, an auditor, a select committee or a journalist, how a particular decision was reached. If AI played any part, the answer has to cover what it was given, which sources it relied on, what it produced, what a person did with it and who was accountable. A general description of the technology does not answer that. A record of the specific case does.

Six properties of a record that holds up

Attributable. Every entry says who or what did it, including which version of the rules and which model.
Contemporaneous. It is written at the time, before the output is used, not reconstructed later.
Complete. It includes the outputs that were refused or escalated, not only those that were released.
Tamper evident. Any alteration to an earlier entry can be detected. An append only record in which each entry is cryptographically chained to the one before gives you this.
Reconstructable. Years later, someone can follow the path from the question to the decision.
Understandable. A reviewer who is not a data scientist can read it and understand what happened.

What the record of one decision contains

Take a caseworker who asks an assistant to summarise the correspondence on a file. A defensible record shows the question, the documents the summary was drawn from and their versions, the checks that were run and their results, the risk tier assigned, whether the output was released, who reviewed it, and whether they accepted or rejected it. If the case later comes under challenge, that single record is the answer.

What to publish and what to keep

There are two audiences. The public needs to know that AI is used, for what, and with what safeguards. In the UK, the Algorithmic Transparency Recording Standard is the way central government bodies publish that information, and wider public sector bodies are encouraged to use it. The detailed per case record is internal. It is held securely, disclosed where the law requires, and available to auditors and reviewers.

Procurement questions

Can the supplier produce the full record of any single output on request?
Is the record held inside our environment, under our control, for the retention period we set?
Are refusals and escalations recorded as fully as releases?
If we change supplier, can we keep the records and read them without the supplier's software?
Does the supplier train its models on our data?

Where public bodies go wrong

Publishing a transparency statement while keeping no per case evidence behind it.
Logging inputs and outputs but not the sources, rules and versions in between.
Treating a human reviewer's click as evidence of oversight, without recording what they were shown.
Relying on a supplier's cloud logs that the department cannot export.

Frequently asked questions

What is the Algorithmic Transparency Recording Standard?

It is a UK government standard for publishing information about the algorithmic tools a public body uses, including what they are for and how they are overseen. It applies to central government bodies, and other public sector bodies are encouraged to adopt it.

What is the difference between a transparency record and an audit trail?

A transparency record is public and describes a tool in general. An audit trail is internal and describes what happened in each individual case. A public body needs both.

Does every AI output in a public body need human review?

Not every one. The depth of review should match the consequence of a wrong output. Low consequence outputs can be released with a full record, while outputs that influence decisions affecting people should be held for a named, accountable person.

Sources and further reading

This article is general information, not legal or regulatory advice. Regulatory texts change, and some discussed here are drafts: check the current official text before relying on any detail.