Feedback and complaints channel for AI

What it does

It gives the organisation the evidence only people can provide. Users, staff and those affected by a system have a simple way to say that an output was wrong, offensive or harmful. Each report is tagged to the system and version, assessed, and passed to someone who can act. Reports are also analysed together, because a model repeats its faults. The system owner is accountable for acting; the complaints function runs the process. It operates throughout the system's use.

How to implement it

• Put a visible way to report in the interface, at the point of use.
• Provide a route for people affected who are not users.
• Tag complaints that involve an AI system, in the general complaints process as well.
• Triage by severity, and send serious reports into incident response.
• Analyse reports in aggregate for repeated faults and affected groups.
• Tell the reporter what was done, and add confirmed faults to the test suite.

Evidence it operates

Design: a feedback mechanism in each system; tagging rules in complaints handling. Operation: report volumes and categories; triage and response times; pattern analysis; changes made as a result; reports that became incidents.

Where it fits

In operation, alongside technical monitoring. It supports AI monitoring and alerting gap and post-market monitoring gap by catching what measurement misses, and supplies evidence to customer outcome monitoring and the post-market monitoring plan.

Limits

Most people who meet a faulty output do not report it, and those harmed by a decision often do not know an AI system was involved. Reports are skewed towards articulate users. A channel that produces no visible response stops being used. Frontline staff have to recognise an AI-related complaint for the tagging to work.