underfitting

advancedwidespread

Underfitting means an AI model is too simple or has learned too little, so it fails to recognize even basic relationships in the training data.

In plain terms

Imagine a student who only memorized for the history exam that "wars are bad." This is true, but they cannot answer any specific question with it. An underfitted model is the same: it sees the world so oversimplified that the essence is missing from it.

In machine learning, a model learns patterns from examples. Underfitting means this learning fails: the model performs poorly even on the examples it was trained on, as well as on new, unknown cases.

Its opposite is overfitting, when the model memorizes the training material but fails in new situations. A good model lies between the two: it understands the rule and does not get stuck on the details of individual examples.

How it works

Developers split the data into at least two parts: the model trains on one and is tested on the other.
During training, the error is measured on both parts. If the model makes many mistakes even on the training data, this is a sign of underfitting.
The reason is usually one of the following: the model is too simple for the task, trained for too short a time, constrained too strictly, or the data lacks essential information.
The fix follows accordingly: a larger or more flexible model, longer training, more and more informative input data, or possibly relaxing overly strict constraints.
Meanwhile, care must be taken that the model does not swing to the other extreme, overfitting.

An everyday example

A real estate website's price estimator calculates only from the floor area, so it values a 60-square-meter apartment in the city center just as much as a house of the same size in a dead-end village. The model is too simple to see the impact of the neighborhood, condition, and floor, so it misses the mark almost everywhere.

Who is affected

developerIt must detect from comparing training and testing errors that the model is too simple, and adjust the model or the data accordingly. → Code

entrepreneur, executiveAn underfitted predictor or recommendation system offers little more than guessing, so the money spent does not pay off.

bank customer, borrowerAn overly simplistic credit-scoring model judges everyone based on a few rough criteria, so it may reject good borrowers and approve risky ones.

patient, doctorAn underfitted diagnostic model fails to detect subtle signals, making it unreliable for decision support. → Health

A common misconception

MythA simpler model is always safer because it cannot "overfit".

In realityAn overly simple model fails just as surely, only differently: it errs systematically, on both the training data and new data. You need only as much complexity as the task requires.

Limits and risks

Poor performance is not always due to underfitting: it can be low-quality, incorrectly labeled, or insufficient data, and in such cases a larger model does not help.
If the essential information is not in the data, neither longer training nor a more complex model will extract it.
While fixing underfitting, it is easy to fall into overfitting, so performance must always be measured on separate data excluded from training.
As a user, you cannot see if a service's model is underfitted; you only notice that its responses are generic and inaccurate.

Similar, but not the same

overfitting – An overfitted model performs excellently on training data but poorly on new data, while an underfitted model performs poorly on both.
bias – Statistical bias is the part of the error that stems from the model's overly simple assumptions. Underfitting models are characterized by high bias, but bias is only an error component, not the phenomenon itself. It should not be confused with societal bias.
hallucination – A hallucination is a confidently presented, fabricated statement from a language model, while underfitting is a general insufficiency in learning.

Related concepts

Broader

machine learning, model training, generalization

Narrower

–

On the edge

overfitting, Bias–variance tradeoff, regularization, training data, validation

A brief history

The term moved from statistical curve fitting to machine learning. The underlying bias–variance dilemma was made widely known for neural networks by Stuart Geman, Elie Bienenstock, and René Doursat in their 1992 study.

Where things stand

As of late 2026, this is a fundamental concept in machine learning, covered in every introductory course. In today's large models, underfitting due to insufficient capacity is rare, but it still occurs in small models, with too-short training or fine-tuning, with overly strong regularization, and with data carrying little information.

Status: 2026 November

Also known as

underfitting · underfitting · undertraining · alulillesztett modell · underfitting · underfit · under-fitting · high bias

Sources

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