An AI system is reliable when it performs its intended job consistently and correctly under the conditions where it is used. For a prediction system, that means more than producing plausible answers: its errors must be understood in terms of what they cost people and organizations. section 11.11 section 10.2
Fraud detection
A payment system must catch fraudulent transactions without rejecting too many legitimate ones. In the cited analysis, one in five rejected online transactions was a false positive, and more than half of those wrongly rejected customers subsequently took their business to a competitor. Reliability therefore depends on both detecting fraud and limiting mistaken declines. section 11.11
The stakes rise when automated decisions can cause physical harm. In a 1991 incident, an autonomous target-acquisition system fired at another ship’s countermeasures, and four rounds struck the ship. Such cases warrant caution about the practicality of autonomous systems, not speculation that they act with human intentions. section 9.4
For decisions that directly affect people, reliability is one part of a broader test that also includes fairness, privacy, transparency, and explainability. Systems should be validated before use in consequential settings. More resilient AI remains a goal, not a reason to assume present systems are dependable without evidence. section 10.2 section 11.0
