overfitting

basicwidespread

Overfitting occurs when an AI model memorizes its training examples, performing well on them but poorly on new, unknown cases.

In plain terms

Think of a student who memorizes the solutions to last year's exam papers word for word. They repeat the old tasks flawlessly, but as soon as the question is phrased slightly differently, they get stuck. They did not understand the material, they just memorized the specific examples.

The same happens in machine learning: the model learns from examples, and the goal is for it to capture the general rule behind them. In overfitting, it instead treats the random details, errors and idiosyncrasies of the examples as rules too.

This is why it is misleading: the developer sees excellent results on their own data, yet the system disappoints when real people use it in real situations. Overfitting is not a useful method but an error that must be continuously monitored and prevented during training.

How it works

Developers split data into at least two parts: the model trains on one, while the other is set aside and used only for verification. Often, a third part is separated for final measurement.
During training, the model becomes increasingly accurate on the training data as it gradually adjusts its internal settings to match them.
Beyond a certain point, it no longer learns only the general pattern but also memorizes the noise and random coincidences of individual examples.
This is revealed by the fact that it improves further on training data but begins to deteriorate on held-out, never-before-seen data.
Defense is possible with more and diverse data, simpler models, stopping training at the right time, or methods that penalize overly complex solutions.

An everyday example

A resume screening program learns from a company's few dozen past hiring decisions. It infers that a good candidate attended a particular university, because by chance many such people were among the few examples. When you apply from a different school with excellent experience, the system ranks you lower because you do not fit the memorized pattern.

Who is affected

developerIt must detect and prevent issues, otherwise a model that looks good in testing will fail in production. → Code

entrepreneur, executiveIf you buy an AI system, the accuracy seen in the demo is worthless until you test it on your own, real data.

patient, doctorIf a diagnostic system has learned the specific characteristics of a single hospital's images (such as the characteristic image noise of the equipment there), it may be less accurate with other machines and other patients than its measurements suggest. → Health

bank customer, investorA credit assessment or stock market model overfitted to past data treats past coincidences as rules, leading to poor decisions in new situations. → Money

A common misconception

MythThe longer and more precisely a model is trained on its own data, the better it becomes.

In realityNear-perfect results on training data prove nothing on their own; they could simply indicate rote memorization. What matters is how well the model performs on cases it has never seen before.

Limits and risks

Overfitting is not visible from the outside: it is only revealed by testing on independent data not used during training.
Accuracy figures in ads and news reports can be misleading if the test tasks or close variants of them were included in the training material.
It cannot be completely ruled out, only mitigated. Careful checks do not protect against everything: in situations that differ significantly from the training data, even a model that is not overfitted may perform poorly.
If defenses against overfitting are too strong, the model may swing to the other extreme and fail to learn real relationships.

Similar, but not the same

underfitting – An undertrained model has learned too little or is too simple, so it performs poorly even on training data, not just on new ones.
hallucination – A hallucination is a confidently presented but fabricated claim by a language model, whereas overfitting is a training error that can affect any machine learning system.
bias – Bias is the adoption of systemic one-sidedness in data, while overfitting is the memorization of random details, though the two often occur together.

Related concepts

Broader

machine learning, training

Narrower

–

On the edge

underfitting, generalization, regularization, early stopping, data breach, training data, tesztadat, validation, benchmark

A brief history

The phenomenon originates from statistics, where it was known long before modern AI in the context of curve fitting: an overly complex model also follows measurement noise. Machine learning inherited the concept, and with the spread of neural networks, it became a fundamental question in training.

Where things stand

Overfitting checks will be a routine part of all serious AI development by the end of 2026. For large language models, the debate focuses mainly on how much training data containing test tasks distorts public benchmark results.

Status: 2026 November

Also known as

overfitting · overfitting · overtraining · overfitting · overfitting · overfit · over-fitting · overtraining

Sources

Wikipedia: Overfitting

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.