A negative feedback loop occurs when a system’s poor outputs shape the data used for future training, making later outputs worse. A recommendation system can serve unrelated content, receive only rare clicks, and then learn from those distorted logs. section 2.4
Treat the loop as a second-order consequence when estimating mistake costs. It can turn an initially weak recommendation into a continuing business problem, while similar effects—such as word-of-mouth after unjustified fraud bans—can damage growth. section 2.4
Poor recommendations reinforce themselves
A recommendation model produces unrelated suggestions. Users rarely click them, so future training data contains mostly rare clicks from a poor experience. A new model trained on those logs can repeat or worsen the original problem. section 2.4
