Thumbs up/down, edits, retries, and support tickets are all feedback. Closed loops improve the product; unfiltered loops can amplify bias, sycophancy, or popular but wrong answers. Teams usually split online evaluation from offline retraining so a few noisy days of traffic cannot rewrite the model.