Data drift occurs when real-world inputs change from the distribution represented in training data. In a language-model app, unpredictable user inputs can produce errors, inaccurate responses, or harmful outputs. section 6.4.2 section 11.3.1
Monitor user behavior and incoming data continuously. Use what you observe to adapt prompts and examples, or trigger updates when imbalances emerge; guardrails can mitigate risks but do not replace monitoring. section 6.4.2 section 11.3.1
Two production signals
An app’s free-text inputs differ from its curated training data and lead to problematic responses. Separately, changing real-world data reveals demographic imbalances in a recommendation system. section 6.4.2 section 11.3.1
