Feedback Loop

A feedback loop is a cycle. A system does something, the result is measured, and that measurement flows back into the next decision. The English word “feedback” means exactly that: a return signal. A thermostat on a heater is the simplest example. It measures the room temperature, heats when it’s cold, and because that makes it warmer, it switches itself off again. As soon as such a feedback signal exists, a system changes its own environment — and afterward reacts to that changed environment.

When recommendations confirm themselves

Almost every major digital product runs on such cycles. A video platform shows you a video, you click, and the click counts as proof that the recommendation was good. So more videos of the same kind follow. This can be pleasant, because the suggestions fit better. But it can also mean that you end up seeing only a narrow slice of what’s out there.

The core problem: the system evaluates itself using data it generated itself. Nobody can click on something that was never shown. So it looks as if it’s unpopular. Experts call this a self-confirming loop. It explains why recommendation systems often become narrower instead of broader.

For companies, feedback loops are at the same time the biggest competitive advantage. Whoever has many users gets a lot of feedback. A lot of feedback makes the product better, which in turn attracts more users. In the business press, this cycle is also called a data advantage. It is one reason why market leaders in the software business are so hard to catch up with.

Positive and negative feedback

Technically, two kinds are distinguished. Negative feedback dampens: the stronger the deviation, the stronger the counter-reaction. The thermostat belongs to this category, as does a car’s cruise control. Such cycles stabilize themselves and oscillate around a target value.

Positive feedback, on the other hand, amplifies itself. One example is the whistling that occurs when a microphone is too close to a speaker. The speaker outputs what the microphone picks up, and the microphone picks up what the speaker outputs. Within fractions of a second, this turns into a deafening sound. “Positive” here doesn’t mean good — just self-reinforcing.

With AI systems, there’s an added twist: the cycle doesn’t turn in seconds, but over months. A language model learns from texts found on the internet. More and more of these texts are themselves written by AI. If new models are trained on this, they learn from their own outputs. Researchers warn that quality can decline over several generations as a result. This effect is discussed as model collapse.

From chat ratings to credit scoring

Feedback loops are encountered most directly in chatbots. Next to every answer there’s a thumbs-up and a thumbs-down. Providers collect these ratings and use them to fine-tune the model. This procedure is called Reinforcement Learning from Human Feedback, often abbreviated as RLHF. Simplified: people indicate which answer they prefer, and the model is adjusted in that direction.

It gets tricky when decisions about people are part of such a loop. A credit-lending system rejects certain groups more often. The bank never learns from rejected applications whether those people would have paid reliably. So the data remains one-sided, and the model confirms its own bias. Something similar has been documented in predictive policing: where more checks happen, more is found, so checks continue there.

A common misconception is confusing a feedback loop with training. Training is a single computational process. A feedback loop is the pattern surrounding it, in which output and input are connected. That’s why regulations like the EU’s AI Act mandate ongoing monitoring after launch. The goal is to detect when a system spirals out of control through its own effects.