Feedback Loop
A feedback loop is an arrangement in which part of a system's output is measured and routed back as an input that influences what the system does next, so that cause and effect form a circle rather than a line. The concept is generic and appears wherever a quantity affects its own future: a room heating itself toward a thermostat setting, a population growing in proportion to its size, a microphone squealing next to its own loudspeaker, or a model whose predictions change the data it will later be trained on. Understanding which kind of loop is present, and how fast it acts, is often the difference between a stable system and one that oscillates or runs away.
Negative feedback pushes a system back toward a target, and it is the basis of regulation. A thermostat compares measured temperature with a setpoint and switches the heater to reduce the difference; the error signal shrinks as the room approaches the goal. James Watt's centrifugal governor on steam engines, widely used from the late eighteenth century, did the same for speed, with spinning weights rising as the engine sped up and closing the valve, and James Clerk Maxwell's 1868 paper "On Governors" gave the first mathematical analysis of why such devices sometimes settle and sometimes hunt wildly. Harold Black at Bell Labs conceived the negative-feedback amplifier in 1927, trading raw gain for stability and low distortion, and Harry Nyquist's 1932 stability criterion supplied a test for when a loop would stay well behaved. Modern control theory, from the proportional-integral-derivative controller in industrial plants to the autopilot in aircraft, extends this lineage, and the broader study of feedback in living and mechanical systems is the subject of cybernetics.
Positive feedback amplifies deviations instead of correcting them, and it drives growth, switching and instability. A bank run, a viral post, an avalanche and compound interest all share the structure in which a change causes more of the same change. Positive loops are also constructive: they make a neuron fire in an all-or-nothing way, a flip-flop circuit hold a bit, and a chain reaction sustain itself. Real systems usually contain several loops of both kinds, and behavior depends on their strengths and delays. A negative loop with too much gain or too long a delay overshoots, corrects too late, and then oscillates, which is why a poorly tuned shower, a controller or a learning-rate setting can swing instead of settling.
Machine learning is built from feedback loops, and it also creates them. The training of a neural network is a negative loop: the loss measures error, backpropagation computes how each weight contributes to it, and gradient descent nudges the weights to reduce it, repeated over many steps; the learning rate acts as the gain, and a value that is too high produces the same overshoot and oscillation seen in a badly tuned controller. Reinforcement learning closes the loop through an environment, with rewards as the feedback signal, and methods that fine-tune language models from human preference ratings place people inside the loop. The more troublesome loops arise after deployment, when a model's outputs change the world that produces its future training data. A recommender that promotes what is already popular generates more clicks on those items, which the system reads as evidence they are good. Predictive policing offers a sharper case: researchers including Kristian Lum and William Isaac, in a 2016 article, and Danielle Ensign and colleagues in 2018, showed that dispatching patrols to places where past crime was recorded produces more records there, which can entrench an initial bias. Work published in 2024 on model collapse by Ilia Shumailov and co-authors describes a similar loop in which models trained on the output of earlier models lose the tails of the original distribution.
Several misconceptions cloud the topic. The words "negative" and "positive" describe the sign of the loop, not its desirability: negative feedback is generally the stabilizing and often welcome kind, while positive feedback is what makes a bank run catastrophic yet also what makes a memory cell work. Another mistake is to assume that more feedback is always better; delayed or noisy feedback can destabilize a system that would behave perfectly well without it, and a feedback signal that is a poor proxy for the real goal leads optimization toward the proxy. Finally, evaluating a deployed system offline, on frozen data, can hide loops that only appear when the system's decisions are allowed to act on the world, which is why monitoring and experiments on live behavior matter.
Read feedback loop as the basic circuit of adaptation and instability alike: whenever an output can reach its own input, the questions that count are the sign, the strength and the delay of the return path, and those three properties explain thermostats, training runs and recommendation engines in the same language.