Public AI transparency record

What it does

It lets people outside the organisation see how AI is being used on them. A public record lists the significant systems, in plain language: the purpose, the decisions it informs, the kinds of data involved, the human oversight in place, how to obtain an explanation or raise a concern, and who is responsible. Statements made elsewhere about the organisation's AI are kept consistent with it. A senior owner is accountable for accuracy; the governance function maintains it. It is updated when a significant system is launched, changed or retired, and reviewed at least annually.

How to implement it

• Set criteria for which systems are listed, starting with those that affect the public.
• Use a standard template, such as the UK public sector's algorithmic transparency recording standard.
• Draw each entry from the inventory, so the record and the inventory agree.
• Describe training data at a level that answers reasonable questions.
• Have each entry approved by the system owner and reviewed by legal.
• Check marketing and investor statements against the record.

Evidence it operates

Design: publication criteria and a template. Operation: the published record with dates; change history tied to system changes; approval records; evidence of a claim corrected to match the record.

Where it fits

At deployment and through operation, drawing on the inventory. It supports training data transparency gap and overstated AI claims by fixing in public what the organisation actually does. It relies on dataset documentation and the AI inventory.

Limits

A record describes systems at a level that may be too general to be useful, particularly where detail is withheld for security or commercial reasons. Publication is a legal expectation for some public bodies and a choice for most firms. A record that is not kept up to date becomes a misstatement.