Feedback Loop

A feedback loop exists when the outcome of a process feeds back into the process itself. A simple example is heating with a thermostat. The thermostat measures the temperature in the room. If it’s too cold, it heats more strongly. This raises the temperature, and the thermostat immediately measures that very temperature again. The result thus becomes the new input, in an endless circle. Such circles exist in nature, in the economy, and indeed in computer systems that learn from data.

Why learning systems stall without feedback

A computer program that learns patterns from data is often called a model. Such a model has no inherent way of knowing whether its answers were good. It needs a signal from outside. This is exactly what a feedback loop provides. Without it, a system remains stuck at the level of the day it was built.

That is why access to feedback is an economic advantage for companies. Whoever has many users gathers many clues about what works. This makes the product better, which in turn attracts more users. Experts call this the data advantage. It explains why large providers are often able to maintain their lead.

But the same mechanism can also cause harm. If a system is systematically off the mark and these errors flow back unchecked, they amplify themselves. A recommendation system that once shows someone extreme videos learns from the clicks that this is exactly what is wanted. Next time it shows even more of it. This is called a self-reinforcing loop.

From user click back into the model

The cycle usually has four stages. First, the system produces an output, such as an answer or a recommendation. Second, someone reacts to it: through a click, a thumbs-up, a correction, or an abandonment. Third, this reaction is stored and evaluated. Fourth, it flows into the next training round and changes the system’s behavior.

With chatbots, this often runs through a method abbreviated RLHF, spelled out as Reinforcement Learning from Human Feedback. Humans rate several of the model’s answers and rank them by quality. From these ratings, a second model emerges that is meant to distinguish good answers from bad ones. The main model is then trained to satisfy this rating model.

It is important to distinguish between two time scales. Some loops run immediately: you correct an autocomplete suggestion, and the suggestion changes within the same conversation. Others run slowly. The ratings of millions of users are collected and only incorporated into a new model months later. A common misconception is the assumption that a chatbot learns permanently from every single conversation. Usually, it does not.

From Spotify to model collapse

You encounter feedback loops constantly in everyday life. Streaming services suggest music and measure whether you keep listening or skip. Map services predict traffic jams and observe how quickly users actually progress. Spam filters learn from every email you mark as unwanted. In all these cases, your reaction is the fuel of the system.

In news about AI, the term currently appears frequently in a particular variant. More and more texts and images on the internet themselves originate from AI systems. New models are trained on this data and thus learn from their own outputs. Researchers call this model collapse: quality slowly declines because errors and monotony build up across generations.

That is why companies deliberately build brakes into their loops. These include spot checks by humans, test groups without personalization, and data sources that are demonstrably human-generated. A feedback loop is therefore neither good nor bad in itself. What matters is which signal is fed into it and who controls the cycle.