Underfitting
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
A failure mode where a model is too simple to capture the real pattern in its data, performing poorly on the training set and on new data alike.
What is Underfitting?
Underfitting is a failure mode where a machine learning model is too simple to capture the real pattern in its data, so it performs poorly on the training set and on new data alike. Where an overfit model memorizes, an underfit model never learns.
It is the easier of the two failures to spot. Bad performance on the training data itself is the giveaway; no held-out test is needed to see it.
Causes of Underfitting
- Insufficient model capacity: a linear model asked to fit a curved relationship has no shape that matches the data, no matter how long it trains.
- Missing features: if the inputs leave out the factors that actually drive the outcome, the pattern is not there to be learned.
- Excessive regularization: a complexity penalty turned up too high flattens the model until it ignores real structure along with the noise.
- Training cut short: a model stopped before it converges behaves like a simpler one, whatever its architecture allows.
Underfitting vs Overfitting
An underfit model has learned too little from its training data; an overfit model has learned too much. The tell is where the errors sit: underfitting produces high error everywhere, while overfitting produces low training error and a wide gap to everything else.
| Criterion | Underfitting | Overfitting |
|---|---|---|
| Root cause | Model too simple for the pattern | Model too complex for the data |
| Training error | High | Low |
| Error on new data | High | High |
| Typical fix | More capacity, better features, less regularization | Regularization, more data, early stopping |
Example of Underfitting
Predict house prices with a linear regression that uses one input: the number of bedrooms. Training on 5,000 past sales, the model settles on a single rule of thumb – roughly a fixed price increase per bedroom.
The errors are large immediately, on the training sales themselves. A two-bedroom apartment in the city center and a two-bedroom house in a declining suburb get the same prediction, because location, floor area, and condition were never inputs. Adding those features, or switching to a model that handles interactions between them, is what fixes it; more training data of the same shape would not.
Related AI terms: Overfitting · Bias-Variance Tradeoff · Regularization · Learning Curve · Feature Engineering
Did you like the Underfitting gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About