Explaining decisions made with AI (ICO & The Alan Turing Institute)
A three-part practical guide co-badged by the UK’s data protection regulator and the Alan Turing Institute on explaining AI-assisted decisions to the people affected by them: Part 1 covers the basic concepts, Part 2 walks through building and delivering an explanation task by task, and Part 3 covers organizational roles and policy.
License: Open Government Licence v3.0 — https://www.nationalarchives.gov.uk/doc/open-government-licence/version/3/
Key points
- Six types of explanation, each answerable independently and combinable depending on context: rationale (the reasons behind the decision, in accessible terms), responsibility (who’s accountable and who to contact for a human review), data (what data was used and how), fairness (steps taken to ensure unbiased, equitable treatment), safety and performance (accuracy, reliability, robustness), and impact (effects on the individual and wider society).
- Every explanation type splits further into process-based (how the system was designed and governed, in general) versus outcome-based (why this specific decision came out the way it did) — the same distinction applies across all six types.
- Five contextual factors decide which explanation types to prioritize: domain (the sector/setting — criminal justice vs. e-commerce carry very different expectations), impact (how severe the consequences are for the individual), data (what was used to train and decide), urgency (how much time the person has before they must act), and audience (their expertise level, and any accessibility/reasonable-adjustment needs).
- Explanations should be layered: give the prioritized explanation type(s) up front, and make the rest available in further layers (expanding sections, tabs, or linked pages) rather than delivering everything at once — explicitly framed as avoiding “explanation fatigue.”
- Treat delivering an explanation as a dialogue, not a one-way disclosure — people should be able to discuss and clarify the explanation with a competent human being, not just receive a static statement.
- Some explanation types can and should be delivered before a decision is made, not only after: process-based explanations generally, plus the outcome-based responsibility, impact, and data explanations. Rationale, fairness, and safety-and-performance explanations are usually decision-specific and only available afterward.