AI Glossary / AI and machine learning / Beginner

Feedback loop

Definition

A feedback loop is the closed circuit in which a system's outputs affect its future inputs: what the system does changes the world it then reads, for better or worse.

What it means

Open systems act and learn nothing; looped systems act, observe the consequences, and act again differently. The loop is the unit of learning, but it is also the unit of distortion, because a system inside a loop is partly reading its own handwriting.

Virtuous loops compound improvement: the model predicts, outcomes are recorded, retraining improves the model, better predictions generate better outcomes to learn from.

Vicious loops compound error: a fraud model blocks a population, so no data on that population's good behavior ever arrives, so the model grows ever more certain they are all fraudulent, having made its own evidence impossible.

Recommendation systems live this daily: they show what they believe you like, you click from what is shown, and the clicks confirm what the system believed, whether or not it was true.

For a business, spotting which loop you are inside is survival knowledge: virtuous loops are assets that compound, vicious loops are debts that compound, and they look identical on a dashboard until someone traces the circuit.

There is a quieter third kind, the stabilizing loop: thermostats and inventory replenishment, where feedback pulls the system back to target. Not every loop is a drama; the steady ones are just invisible.

How it works

Map the circuit: what the system outputs, what the world does in response, and what of that returns as input.

Ask the self-evidence question: could the system's actions be manufacturing the data that confirms it, by hiding alternatives or shaping behavior.

Break vicious loops deliberately: exploration traffic, randomized holdouts and human review populations keep counter-evidence flowing.

Strengthen virtuous loops: shorten the delay between action and recorded outcome, because fast loops learn faster.

Monitor loop health: diversity of inputs over time, share of decisions overridden, and whether the system's world is shrinking to its own reflection.

Three practical examples

• Credit: approvals generate repayment data that improves the model, if declined applicants are not made permanently invisible.
• Recommendations: exploration slots that sometimes show unranked items, so clicks reveal taste rather than rank position.
• Hiring tools: tracking hires' outcomes without letting the tool's rejections erase all evidence about the rejected.

Think of it this way

Think of a microphone placed before its own speaker. The speaker's output re-enters the microphone, comes out louder, re-enters again, and within seconds the room is howling. A feedback loop without damping is that howl, in data.

The limit: the cure is not earplugs but distance and mixing. Randomness, holdouts and human judgment are the mixer's controls that keep the loop from singing only its own note.

Worked example

unbreaking a fraud loop

Suppose a bank's fraud model has blocked a whole merchant category for a year, and its certainty about that category keeps rising. The figures are illustrative.

An analyst asks the fatal question: how would we know if we were wrong, given that no transactions in the category are ever approved.

The bank runs a controlled experiment: 2 per cent of category transactions are approved under watch, and most prove legitimate.

The model is retrained with the exploration data, blocking falls 60 per cent in the category, and confirmed fraud barely moves.

Illustrative case study

A fictional content platform let its recommendation loop run untended, and within a year its top content was uniformly extreme, because each round of engagement fed a narrower, louder selection.

Engagement was up and advertiser trust was down, and unwinding the loop took a year of deliberate exploration.

The invented lesson: a loop optimizing a metric will consume everything that is not the metric. Decide what the loop is not allowed to eat.

Common misconceptions and mistakes

• Feedback loops are always good. They compound whatever they contain, wisdom or error, and the vicious kind manufactures its own confirmation.
• Accurate predictions prove the loop is healthy. Accuracy on self-selected data proves only that the system recognizes its own shadow.
• Loops only affect AI. Hiring, lending, policing and pricing have run human feedback loops for centuries; AI just runs them faster.

Business use

Audit each automated decision system for its loop: what evidence can no longer arrive because of the system's own actions.

Fund exploration as a line item: a small share of traffic devoted to learning beats a large share devoted to confirmation.

Watch input diversity as a health metric; a shrinking world is a loop closing in on itself.

In strategy reviews, ask of any improving metric: is the system learning, or is the loop shrinking the world until the metric is easy? The two look identical from inside.

People also ask

What is a runaway feedback loop?

One where the system's outputs keep amplifying their own inputs without damping, like the microphone howl, until the system occupies a tiny extreme corner.

How do holdouts break loops?

By preserving a population the system cannot influence, they keep generating evidence the system did not manufacture.

Are feedback loops visible in dashboards?

Rarely directly; they show as narrowing input diversity and rising self-confirmation, which need deliberate measurement.

Check your understanding

Why can a highly accurate model still be trapped in a vicious feedback loop?

Answer: Because its accuracy is measured on data its own actions selected, so it can be perfectly right about the world it created and wrong about the world outside it.

Related terms

• Feedback
• Reinforcement learning
• Exploration
• Selection bias
• Retraining
• holdout

Learn next

AI for Business Owners explains how to spot and break the loops that make automated systems confirm their own mistakes: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

More in AI and machine learning

• Federated learning - Federated learning is training a shared model across many devices or organizations without moving their raw data to one place: the training goes to the...
• Feed-forward neural network density model - A feed-forward neural network density model uses a standard layered network to estimate the probability distribution of the data itself, so it can score how...
• Feed-forward neural network regression - Feed-forward neural network regression is using a standard layered network to predict a number, such as a price, a demand or a risk score, from input features.
• Feedback neural network - A feedback neural network includes connections that allow earlier internal activity or later-layer information to influence subsequent processing.
• Filters, random fields, and maximum entropy model - The filters, random fields, and maximum entropy model, usually called FRAME, is a probabilistic model developed for texture analysis and synthesis.
• Finite mixture model - A finite mixture model describes data as coming from a fixed number of underlying groups, each with its own distribution, so one dataset can be read as...

Sources

[S2] Google, Machine Learning Glossary; [S3] NIST AI Resource Center glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Feedback Next: Feedback Loop Bias