Contestability and appeals process

What it does

It gives a wrong decision a way to be corrected. Anyone affected by a decision in scope is told how to challenge it. A challenge is considered by a person who did not make the original decision, who can see the inputs and the reasons, who looks at the case on its merits and who is free to reach a different result. The outcome is explained, and a successful challenge is followed through. The business owner of the decision is accountable; a complaints or review function operates it. It is available for as long as decisions are made.

How to implement it

• Tell people how to challenge, with the decision and in plain words.
• Make the route simple, free and open to those who are not customers.
• Assign reviewers who are independent of the original decision and trained.
• Give reviewers the full record, and do not let review mean re-running the model.
• Set time limits, and explain the outcome with reasons.
• Analyse upheld challenges for systematic faults and report them to the model owner.

Evidence it operates

Design: an appeals procedure; reviewer independence and access arrangements. Operation: volumes, outcomes and times; sampled reviews showing fresh consideration; changes to models or rules that followed upheld appeals; information given to affected people.

Where it fits

After a decision is communicated. It is the primary control for inadequate contestability and redress and wrongful automated refusal. It depends on explanations for affected individuals and AI record-keeping and audit trail, and hands upheld cases to customer remediation for AI errors.

Limits

Most people who receive a wrong decision do not challenge it, so appeals correct a fraction of errors. Reviewers are subject to the same deference to the system as anyone else. An appeal takes time the person may not have. The process corrects individual cases and fixes nothing systemic unless its findings are acted on.