A recommendation engine suggests cutting the marketing budget for a high-performing product but gives no reason. Even if the recommendation could be useful, users may see an arbitrary black box and stop relying on it. Trust depends not only on an AI system’s output, but on whether people can understand its limits, judge its advice, and remain in control. section 11.4
AI makes trust harder to design for because open-ended inputs can produce unpredictable outputs, including wrong predictions. The goal is not to promise certainty. It is to help users rely on the system appropriately: show what it can and cannot do, make uncertainty visible, and plan for mistakes rather than treating them as exceptions. section 10.0 section 12.3
Build trust across the product experience and its operation:
- Choose automation users actually want. Consider accuracy, the effort needed to check outputs, and whether taking over a task would remove valued human judgment. Start with high-confidence opportunities before more complex ones. section 10.1.1
- Explain consequential outputs. In high-stakes workflows, users need enough insight to assess a recommendation and discuss it with others; transparency includes explainability, interpretability, and accountability. section 11.4
- Preserve meaningful control and invite feedback. Be candid about limitations, provide guidance for using the system, and give users ways to help improve its data, experience, and performance. section 12.3 section 12.3.3
- Back the interface with governance. Protect data and models, test for bias, maintain human oversight and audit trails, and monitor compliance. Security or privacy failures can undermine trust regardless of the interface. section 11.6 section 11.0
A forecast a team can challenge
A supply-chain manager receives forecasts with confidence intervals rather than unexplained figures. The team can adjust predictions using its own insights and consult documentation describing the data sources and algorithms. These choices make uncertainty and human control part of the workflow, not a reassurance added after users raise concerns. section 12.3.3
Address concerns directly, including fears about how the system works or whether automation threatens jobs. Honest answers may not remove every objection, but ignoring those concerns can damage adoption and satisfaction. Trust is maintained through both product decisions and continuing conversations with users. section 12.3.3 section 12.3
