Exam-style
A retailer's customer-facing support assistant, built on Amazon Bedrock over the company's own help articles, tells a customer her order cannot be returned. She knows the item is inside the return window and that the answer is wrong. The interface gives her the paragraph and nothing else. The team is asked what to add so a customer in her position is treated properly. Which change does that?
Reveal the answer
C. Add a control on every answer that lets her mark it wrong and say why, feeding a queue a support person works through, and label the assistant as AI, show the help articles the answer was drawn from, and give her a way to reach a human
Sort these five by who they are for. A guardrail filter strength, a model evaluation job, a Model Card and a temperature setting are all real controls, and every one of them is aimed at the builder or the auditor. The guardrail tightens what the system will emit, the evaluation job measures a model against a dataset before it ships, the Model Card is documentation an assessor reads, and temperature is a decoding knob. None of them reaches the customer holding a wrong answer about her own order. The principles of human-centered design for explainable AI ask a different question: what can the person affected see, challenge and escalate? Two named principles answer it. User-feedback mechanisms give her a way to disagree in place, a thumbs-down with a comment box, wired into a queue somebody actually works, because a control that collects complaints and answers none of them is worse than no control. AI decision transparency covers what the interface tells her up front: that she is talking to an AI system rather than a person, which help articles the answer rested on so she can check them herself, how confident the answer is when the system knows, and how to reach a human who can override it. Neither one improves the model. Both change what happens to the customer the model got wrong. The feedback is worth having twice over, because a queue of marked-wrong answers is a labelled dataset of the assistant’s failures, and it feeds the next evaluation set and the next round of customisation. So the four other options are not wrong so much as later: they are what the team does with what the feedback control tells them.
Q. A customer reads a wrong answer from the support assistant and the interface gives her a paragraph and nothing else. Stronger guardrail filters, a model evaluation job, a feedback control plus disclosure, a Model Card, or a lower temperature?
A. The feedback control plus disclosure: let her mark the answer wrong into a queue a person works, tell her she is talking to an AI, show the articles it drew on, and name a route to a human. The other four never reach her.
Why? Ask who each control serves. Guardrail filter strength, a Bedrock model evaluation job, a Model Card and temperature all sit on the builder’s side of the glass, or the auditor’s. They shape what the system emits, measure it before launch, or document it for somebody assessing the organisation. The principles of human-centered design for explainable AI are about the far side of the glass, and the guide names two of them. User-feedback mechanisms let the affected person say the answer is wrong, in the place she read it, into a queue that a human works through; a control that goes nowhere is a complaints box with the lid welded shut. AI decision transparency is what she is told without asking: that an AI produced this, what it was working from, which sources the answer traces back to, and who to escalate to. Then the loop closes back on the builder. Marked-wrong answers are the cheapest labelled failures a team will ever get, better than most monitoring at finding faults, and they become the evaluation set for the next model.