Learn AI /Model Quality and Evaluation /Overfitting
Overfitting
- Starter
- Model Quality and Evaluation
- Technical readers
- Programme & data teams
Also called: overfit.
Back to the milestone you came from
In five seconds
What it is
A model fits its training examples so closely that its performance drops on new cases.
In plain words
It learned the practice examples too specifically, so it does not cope well with fresh ones.
An everyday example
Someone learns to identify a dog from dozens of photographs taken in the same living room. They appear to be excellent at the task, but most photographs include the same striped carpet. When shown new pictures of the dog outdoors, they struggle because they had partly learnt to recognise the carpet rather than the animal. The familiar background was useful for the old photographs but did not generalise to new ones.
A student can consciously choose study methods, while a model does not choose in that human sense; it adjusts mathematical parameters. Also, overfitting can happen even when the model has not memorised every example word for word, because it may still rely on unstable patterns.
Remember thisExcellent results on training examples do not prove that a model will work on new cases. Prefer the model that performs reliably on genuinely unseen data.
Understand it
Understand it
Why it matters
A good-looking score can lead people to trust a system that will fail in new situations. That can waste money, miss real problems, and create unfair or unsafe decisions if the model is deployed without proper evaluation.
Training score up, real-world score down
After using this explainer, the reader should be able to say why a model can improve on training data while getting worse on new data.
An interactive chart shows two lines across training rounds: a training score and a validation score. Controls let the reader increase model complexity, add or reduce noise, and move through rounds of training. A small panel of example satellite road images updates to show which visual cues the model appears to rely on, such as road cracks versus irrelevant image artefacts. The key teaching point is the split between learning a useful pattern and learning quirks.
Readers can drag a model-complexity slider, scrub through training rounds, toggle noisy data on or off, and press a button to stop at the best validation round.
Describe this without the diagram
This interactive explains overfitting using text and values instead of graphics. At first, both training and validation scores rise together as the model learns useful patterns. When model complexity becomes too high or training continues too long, the training score keeps improving but the validation score stops improving and then falls. A text panel explains that the model is now relying on unstable clues, such as image artefacts, rather than the real feature of interest, such as road damage. Controls let the user change complexity, training round, and noise level, and each control updates the score values and explanation in words.
With reduced motion: A static version shows three side-by-side states: early training, useful fit, and overfit. Each state includes fixed score values, a non-animated line chart, and text describing which cues are useful and which are irrelevant.
A machine-learning model learns patterns from examples. The goal is not to repeat those examples perfectly, but to work well on new cases it has not seen before. That ability is called generalisation.
Overfitting happens when the model locks onto details that do not hold up outside the training data. These details can include noise, accidental quirks, or patterns that only appear in one dataset. The model may score very highly during training but do worse in real use.
What goes in is a set of past examples, such as satellite images labelled "damaged road" or "not damaged road". During training, the model adjusts itself to match those labels. If it adjusts too closely, it may learn irrelevant clues, such as a certain image border, season, or camera source, instead of the road damage itself.
What comes out is a misleading picture of quality: strong results on familiar data, weaker results on new data. This is why people compare performance on separate validation or test sets, not only on the data used for training. Overfitting is different from data leakage, where information that would not be available at prediction time slips into training or testing.
How “Recognising the carpet, not the dog” maps onto the real thing
| In the story | In the AI system |
|---|---|
| memorising one old test paper | fitting very closely to the training data |
| doing very well on that same practice paper | high training performance |
| struggling on the new exam | poor performance on unseen data |
| learning the broader subject | generalisation |
A worked example
A programme officer in a regional bureau assesses a model that routes incoming progress reports to review queues before the weekly portfolio meeting.
Starting from: Historical progress reports prepared with an older template for training, and a separate set of more recent reports using revised formats for validation.
- As the model is trained, its error on the historical reports becomes very low.
- At first, its error on the separate reports also falls, suggesting that it is learning useful features.
- With further training, error on the historical reports continues to fall, but error on the separate reports begins to rise.
- Review shows that the model has relied on old page headers and table positions rather than the substantive content used to choose a review queue.
Result: The later version is overfitted. It performs better on the reports it learnt from but worse on new reports. The programme officer retains the earlier version, which generalises more reliably across formats.
In a United Nations workplace
During a records retention review, a legal officer in a UN Country Team working group examines a model that assigns files to retention categories. Its results are nearly perfect on the shared-drive records used for training, but deteriorate on newer files from other offices, just as a note for the file is due for clearance by close of business.
The model may have learnt filename conventions, templates and folder structures peculiar to the training records rather than features that indicate the correct retention category. Testing it on records from other offices and periods shows whether those patterns generalise.
Before you try thisRecords may contain personal, human resources, procurement-sensitive or security-sensitive material. They must not be uploaded to a tool that has not been approved for such material. The model must not make retention or deletion decisions on its own, since an error could destroy an official record prematurely or retain sensitive information beyond the authorised period.
Where you may hear it
- At an ICT governance board reviewing a pilot whose training results are much stronger than its validation results
- In a procurement meeting when a vendor demo relies on examples used during model development
- At a regional bureau portfolio meeting after performance drops on the latest reporting cycle
- During first reading of an evaluation report comparing results on training and unseen data
Avoid confusion
Avoid confusion
Often confused with
Underfitting
A model misses important patterns because it is too simple, too constrained, or too weakly trained for the task.
Generalisation
The ability of a model to give useful results on new cases, not just on the examples used to build or tune it.
Data leakage
Data leakage happens when future, target or held-out test information improperly influences model training or evaluation, making performance look better than it is.
Common misconceptions
Common misconception: If training accuracy is very high, the model must be good.
Correction: High training accuracy only shows the model matches data it has already seen. What matters for real use is how it performs on separate unseen data.
Common misconception: Overfitting only means memorising examples exactly.
Correction: Exact memorisation is one form, but overfitting can also mean learning weak or accidental patterns that do not repeat outside the training set.
Common misconception: A bigger, more complex model is always better.
Correction: More complexity can capture real structure, but it can also capture noise. Past a certain point, extra complexity may reduce performance on new cases.
Common misconception: Overfitting and data leakage are the same thing.
Correction: They are related but different. Overfitting is fitting too closely to training data; data leakage happens when information from the answer or test data improperly enters training or evaluation.
Quick check
Which result most strongly suggests overfitting?
A widening gap where training keeps improving but validation gets worse is the classic warning sign. It suggests the model is fitting the training examples too closely instead of learning patterns that generalise.
Go deeper
Go deeper
The simple version
It learned the practice examples too specifically, so it does not cope well with fresh ones.
The technical version
Overfitting occurs when a model has low error on training data but higher error on unseen data because it captures noise or sample-specific patterns rather than signals that generalise.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This link gives a risk-management perspective on testing and monitoring AI systems and suits programme, governance, and assurance audiences.
Intermediatelink checked 2026-07-31
EncyclopaediaOverfitting (opens in a new tab)Wikipedia
This link gives a broad reference-style overview of overfitting and suits readers who want a quick definition plus related concepts.
Intermediatelink checked 2026-07-31
CourseClassification: Accuracy, recall, precision (opens in a new tab)Google
This link helps readers place overfitting beside common evaluation measures and suits beginners comparing model scores.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Overfitting (opens in a new tab)YouTube search
This link is a YouTube search results page for video explanations of overfitting and suits readers who prefer to compare several teaching styles before choosing one.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Overfitting (opens in a new tab)Reddit search
This link is a Reddit search results page showing community discussions about overfitting and suits readers who want informal questions, examples, and debate, with careful fact-checking.
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
Excellent results on training examples do not prove that a model will work on new cases. Prefer the model that performs reliably on genuinely unseen data.