Feedback Loop Amplification

Feedback loop amplification occurs when the outputs of a system are used as inputs to the process that generated those outputs, causing the system's existing patterns to self-reinforce over successive iterations. The term distinguishes amplifying feedback — where errors or biases compound — from homeostatic feedback, which corrects deviations. In the context of automated decision-making systems, feedback loop amplification is a primary mechanism by which initial biases in training data become entrenched and enlarged in deployed systems: a model trained on historically biased decisions produces biased outputs, which inform future data collection, which trains the next version of the model on more biased data. The loop does not stabilize; it drifts in the direction of its initial bias. The distributional shift this causes is invisible to standard accuracy metrics measured on the current distribution, because the distribution itself is shifting under the measurement. Detecting feedback loop amplification requires longitudinal analysis across model versions — a practice rarely built into deployment evaluations.