What is a feedback loop?
A feedback loop forms when the output of an AI system influences the data later used to train, tune, or evaluate it. The loop usually runs through people and operations. A model makes a prediction, someone acts on it, the outcome is recorded, and the record becomes training or evaluation data for the next version. Generative AI has the same loop, since every accepted, edited, or rejected draft records what a user wanted. Reinforcement learning from human feedback (RLHF), which model developers use to train chat models, is one formal version of it.
Some loops are built on purpose to improve a system. Others form without anyone designing them, when a model’s decisions change which data gets collected.
How a feedback loop works
A deliberate loop has four parts. First, the application captures a signal with each output, such as whether a dispatcher at a managed service provider kept the suggested ticket category or changed it. Second, the signal is stored with the input, the model version, and the time. Third, a human-in-the-loop review checks a sample, because quick corrections made under pressure contain their own mistakes. Fourth, the reviewed examples update the system through retraining, fine-tuning, new prompt examples, or revised retrieval content, and the new version is tested on held-out cases before release. Active learning improves the loop by sending the cases the model is least sure about to reviewers first.
Feedback loops in practice
A regional grocery chain forecasts demand for each item and store, and the forecast sets replenishment orders. When the forecast runs low, the shelf empties early and recorded sales stop at whatever was in stock. The next model trains on those capped sales and reads them as weak demand, so the store orders even less. The fix is to record stockouts alongside sales, so the model learns from estimated demand instead of from what the shelf happened to hold. Similar loops appear in fraud detection, where only flagged transactions get investigated, and in recommendations, where items that are shown collect the clicks that make them look popular.
What to watch for
- Capture feedback inside the normal workflow, since few users fill in a separate rating form.
- Check whether the model’s own decisions determine which outcomes get recorded, which is a form of selection bias.
- Where the business can tolerate it, leave a small random sample of decisions outside the model’s control, so there is unbiased data to compare against.
- Review model monitoring results after each update, since a loop can reinforce errors as well as correct them.