Metacognitive Feedback

Ordinary feedback only says whether a result was right or wrong. Metacognitive feedback goes a step further. It evaluates how someone arrived at that result and how confident they were while doing so. The term originally comes from learning research: metacognition means thinking about one’s own thinking. In work with computer programs that learn from data, this has become its own procedure. Such programs are called models, and they are meant not only to deliver answers but also to assess how reliable those answers are.

Why self-assessment determines trust

Language models like ChatGPT often phrase wrong answers just as confidently as correct ones. Experts call this overconfidence, meaning exaggerated self-assurance. This is dangerous for users because the tone gives no indication of quality. A model that says “I’m not sure about this” is, in practice, often more valuable than one that always sounds decisive.

This is precisely where metacognitive feedback comes in. It trains not only the answer but also the self-assessment. The goal is a system whose confidence matches its actual accuracy rate. If a model states 90 percent confidence for each of a thousand statements, roughly 900 of them should be correct. Experts call this alignment calibration.

The distinction from mere error correction is important. Simply handing a student the correct solution teaches them little. Explaining exactly where in the calculation the reasoning went wrong prevents the same mistake next time. The same principle applies to models, just on a larger scale.

From the reasoning path to evaluating the reasoning path

Technically, the procedure usually begins with a model writing out its intermediate steps. This chain of partial considerations is called Chain of Thought. Instead of stating the result immediately, the model first notes down the individual steps. This is what creates something that can actually be evaluated in the first place.

A second pass then checks this chain. This can be done by a human, by a separate evaluator model, or by the model itself in a new round. Each step is assessed individually, not just the final result. Experts then speak of a process-based reward signal, as opposed to an outcome-based one. The difference is substantial: a model can arrive at the correct number by guessing, even though the path leading there was completely nonsensical.

A second component consists of confidence statements. The model is supposed to indicate, for each statement, how convinced it is. During training, it is then rewarded when high confidence coincides with correct answers, and penalized when an error was presented confidently. A well-known pitfall here is that models learn to always sound uncertain as a precaution. This avoids the penalty without actually improving the assessment. Good training procedures therefore also penalize unwarranted hedging.

From math benchmarks to warnings in the chat window

The effect is most visible in so-called reasoning models, i.e., systems that think for longer before answering. Models of this kind achieve significantly better results on math problems than earlier versions. A key reason is that during training, not just the final result but the entire solution path was evaluated.

In everyday life, you encounter this principle in smaller forms. Some chatbots explicitly reference sources or note that a piece of information is uncertain. Search systems flag answers for which they found no supporting evidence. Even the follow-up question “Do you mean X or Y?” is a consequence of a system recognizing its own uncertainty.

In the news, the term often appears in connection with hallucinations, meaning freely invented claims made by language models. Metacognitive feedback is considered one of the approaches to addressing this problem there. It does not solve it entirely, since a model can also be mistaken about its own uncertainty. But it makes the difference between knowledge and guesswork more visible to users, and that is the decisive point in areas such as medicine, law, or finance.