Learn AI /Model Quality and Evaluation /Underfitting
Underfitting
- Intermediate
- Model Quality and Evaluation
- Technical readers
Also called: underfit.
Back to the milestone you came from
In five seconds
What it is
A model misses important patterns because it is too simple, too constrained, or too weakly trained for the task.
In plain words
The system is too basic for the job, so it misses real patterns and performs badly even on examples it has already seen.
An everyday example
Suppose you estimate every journey using only the distance. The rule ignores traffic, road type, weather and time of day, so it gives poor estimates even for journeys you have already observed. Adding relevant information and allowing a more flexible rule can capture patterns that distance alone misses.
A document search rule is a hand-written method, while a machine learning model adjusts numerical parameters from data. Underfitting can also come from strong regularisation, low model capacity, or too little training, not only from using too few keywords.
Remember thisHigh error on both training and validation data often points to underfitting: the model may need better features, greater capacity, more effective training or less restrictive regularisation.
Understand it
Understand it
Why it matters
A model that underfits can look safe because it is simple, but it can still mislead decisions by missing important patterns and producing consistently weak results.
Too simple, about right, and too complex
After viewing this, the reader should be able to say that underfitting means a model is too simple to capture the real pattern, so it performs poorly even on training data.
01
Same data
Each panel starts with identical data points, including a curved underlying trend with some noise.
Draw the same x-y scatter plot in all three panels, with faint grey dots and a dashed light-blue curved reference trend behind them.
02
Underfitting
A model that is too simple misses the trend, so errors stay high on both training and test data.
In the left panel, draw a straight orange line through the scatter that clearly misses the curve. Add badges: 'training error: high' and 'test error: high'.
03
Good fit
A model with suitable complexity captures the main pattern without chasing random variation.
In the centre panel, draw a smooth green curve close to the dashed reference trend. Add badges: 'training error: low' and 'test error: low'.
04
Overfitting
A model that is too flexible follows noise in the training data, so test performance worsens.
In the right panel, draw a dark red highly wavy line passing near many dots. Add badges: 'training error: very low' and 'test error: high'.
Show three side-by-side panels with the same scatter plot. Left: a straight line misses a curved pattern, labelled underfitting. Middle: a smooth curve follows the main trend, labelled good fit. Right: a wavy curve follows noise, labelled overfitting. Under each panel, show training error and test error as high/high, low/low, and very low/high.
Hover or tap each panel to highlight its error pattern and show a short note: too simple, balanced, or too complex.
Describe this without the diagram
The diagram compares three models using the same data. In all panels, the real pattern is curved. In the left panel, a straight line cuts across the data and misses the curve. It is labelled underfitting, with high training error and high test error. In the middle panel, a smooth curve follows the main trend and is labelled good fit, with low training and test error. In the right panel, a very wavy line bends to follow many individual points and is labelled overfitting, with very low training error but high test error. The key message is that underfitting means the model is too simple, so it performs badly even on examples from the training set.
With reduced motion: A static three-panel comparison shows the same data with an underfit straight line, a well-fit smooth curve, and an overfit wavy curve, plus text labels for training and test error under each panel.
A machine learning model takes input data, such as past examples with labels, and tries to learn a pattern that can be used on new cases. Underfitting happens when the pattern it learns is too crude to capture what really matters in the data.
What goes in can be training data, chosen features, and model settings such as model size or training time. What happens is that the model fails to learn enough structure. This may occur because the model form is too simple, the useful inputs were left out, the regularisation is too strong, or training stopped too early.
What comes out is poor performance even on the training data, not only on new test data. That is a key clue. If a system cannot do well on examples it has already been shown, it may be underfitting rather than overfitting.
Underfitting is often confused with overfitting. They are different. Overfitting means the model matches the training data too closely and then fails on new data. Underfitting means it does not capture enough signal in the first place. Both reduce generalisation, but for opposite reasons.
How “Predicting journey time from distance alone” maps onto the real thing
| In the story | In the AI system |
|---|---|
| Search tool using only the exact word "lunch" | A model that is too simple or uses too few useful features |
| Missing documents that use other phrases | Failing to capture important patterns in the data |
| Bad results even on the team’s test set of known documents | Poor fit visible already on training or development data |
A worked example
A liaison office tests a model to flag implementing partner progress reports that require follow-up before clearance. The first version considers only whether a report was late and how long it is.
Starting from: Labelled past reports are divided into training and validation sets. The initial model has two features and is strongly constrained.
- The model makes many errors on the training set because timely or concise reports can still contain unresolved delivery, access or procurement issues.
- Validation error is also high and broadly similar to training error. The problem is not a large gap between the two sets, but weak performance on both.
- The duty officer and programme officer identify additional relevant features, including repeated output slippage, unexplained activity changes and outstanding follow-up points.
- After retraining with the richer features and suitable model capacity, error falls on both the training and validation sets. The office then checks performance across report types before considering operational use.
Result: The initial model was underfitting. Its features and effective complexity were insufficient to learn the target pattern, so it performed poorly even on the material from which it was trained.
In a United Nations workplace
At close of business, a duty officer in a liaison office reviews an implementing partner's progress report for the front office. A triage model uses only submission delay and report length to flag files needing follow-up. It misses reports showing repeated output slippage, unexplained revisions to activities and unresolved procurement issues.
The model is easy to describe, but its limited features cannot capture the patterns that drive follow-up. If it performs poorly on both the reports used for training and a separate validation set, underfitting is a likely diagnosis.
Before you try thisProgress reports may contain personal, confidential or procurement-sensitive material. They must not be entered into an external tool that has not been approved for that material. Testing should use an approved environment, appropriate access controls and human review before any flag reaches the front office.
Where you may hear it
- A model evaluation meeting before a pilot is cleared
- A vendor demo in a procurement meeting
- A review of validation results with the information management officer
- An ICT governance board discussion on model performance
Avoid confusion
Avoid confusion
Often confused with
Overfitting
A model fits its training examples so closely that its performance drops on new cases.
Accuracy
The share of cases a system labels correctly out of all the cases it was tested on.
Bias (statistical)
Statistical bias is a systematic gap between a model’s expected prediction across possible training samples and the true target, often caused by restrictive assumptions.
Common misconceptions
Common misconception: If a model is simple, it is automatically better and more trustworthy.
Correction: Simple models can be easier to inspect, but they can still miss important structure. A model that misses the main signal may produce unreliable outputs even if its logic is easier to describe.
Common misconception: Underfitting only matters if test accuracy is low.
Correction: A main sign of underfitting is that training performance is also poor. If the model cannot fit known examples reasonably well, it may be too limited for the task.
Common misconception: More data always fixes underfitting.
Correction: Extra data helps some problems, but not if the model class is too weak, the features are uninformative, or regularisation prevents learning the pattern.
Quick check
Which pattern most strongly suggests underfitting?
Underfitting usually shows up as weak performance on both training and test data because the model has not captured enough of the real pattern. Very low training error with worse test error is more typical of overfitting.
Go deeper
Go deeper
The simple version
The system is too basic for the job, so it misses real patterns and performs badly even on examples it has already seen.
The technical version
Underfitting occurs when model capacity, features, training, or effective complexity after regularisation is insufficient to capture the target pattern, raising error on training and validation data and sometimes reflecting high statistical bias.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This official framework helps readers place model performance problems within broader risk management and governance, and it suits practitioners who need organisational guidance.
Intermediatelink checked 2026-07-31
EncyclopaediaWikipedia search: Underfitting (opens in a new tab)Wikipedia
This search results page helps readers find multiple Wikipedia entries or related topics on underfitting, and it suits readers who want a broad starting point rather than a single source.
search results, not one pagelink checked 2026-07-31
CourseClassification: Accuracy, recall, precision (opens in a new tab)Google
This course resource helps readers compare evaluation measures such as accuracy, recall, and precision, and it suits learners who need context for judging whether poor results may reflect underfitting.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Underfitting (opens in a new tab)YouTube search
This search results page helps readers find a range of video explanations of underfitting, and it suits people who prefer visual or spoken teaching rather than text.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Underfitting (opens in a new tab)Reddit search
This search results page helps readers see informal community questions and examples about underfitting, and it suits readers who want practical discussion but should be read critically.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
High error on both training and validation data often points to underfitting: the model may need better features, greater capacity, more effective training or less restrictive regularisation.