Overfitting means an AI model has learned its practice examples too exactly, a bit like memorising them, so it does well on those but badly on anything new. The opposite, a model that copes well with new examples, is said to generalise.
Think of a student who memorises last year’s exam answers word for word. Give them the same paper and they score full marks. Change the questions slightly and they are lost, because they learned the answers, not the subject. That student has overfitted.
In a sentence: “The model scored perfectly on the practice questions but failed the new ones, a classic case of overfitting.”
What the word means
The plain point: the model fits its training data so closely that it stops being useful elsewhere.
Google’s Machine Learning Crash Course defines it in nearly those words, and offers its own comparison: “Overfitting means creating a model that matches (memorizes) the training set so closely that the model fails to make correct predictions on new data.” … “An overfit model is analogous to an invention that performs well in the lab but is worthless in the real world.”1
The goal of training is the opposite: a model that works on things it has never seen.2
Why it happens
The plain point: the practice examples were not like real life, or the model was too complicated for them.
Google names two broad causes: training data that does not represent real life, and a model that is too complex.2
- Unrepresentative data: if every cat photo in training was taken indoors, the model may struggle with a cat in a garden.
- Too complex a model: a model with lots of room to spare can store quirks of its examples, including noise, instead of the general pattern.
The usual check is simple. Builders keep some examples back and test the model on those afterwards.3 Doing well on practice but badly on the held-back set is the warning sign.
Why it matters to you
The plain point: a high score on a test the model has effectively seen tells you little.
The same idea applies to the scores AI companies publish. Benchmarks, the standard tests used to compare models, are public, so their questions can leak into training data. A 2024 study wrote a fresh set of school maths problems in the style of a well-known test and compared results.4 It found “several families of models showing evidence of systematic overfitting”, while “many models, especially those on the frontier, show minimal signs of overfitting”.5 How AI is tested covers the limits of benchmarks more widely.
Judge an AI tool on your own tasks, not only on its published test scores. Your work is the one exam it cannot have memorised.
Often confused with
- Underfitting: the opposite failure, where the model has not even learned its practice examples. Google: “An underfit model doesn’t even make good predictions on the training data.”6
- Hallucination: a chatbot confidently stating something false. It is a different problem with its own causes. See hallucination meaning.
- Bias: errors that fall harder on some groups of people. Unrepresentative data can cause both. See AI bias meaning.
Questions people ask
What causes overfitting?
Google names two usual causes: practice data that does not look like real life, and a model that is too complex. In its words: “The training set doesn’t adequately represent real life data” or “The model is too complex.”2
Overfitting vs underfitting: what is the difference?
An overfit model has learned its practice examples too well; an underfit one has not learned them well enough. Google: “An underfit model doesn’t even make good predictions on the training data.”7
How do you spot overfitting?
Test the model on examples it never saw in training. Google: “You should test a model against a different set of examples than those used to train the model.”3 A big gap between practice and test scores is the warning sign.
Learn more on this site
- Learning from examples: supervised learning meaning.
- Testing your own way: How to test whether a prompt change actually helped.
- How models are made: How an AI is actually built.
- The basics: the AI glossary and AI explained for complete beginners.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. What is overfitting?
- A model that is too small to hold anything
- A model that learned its practice examples too exactly and does badly on new ones
- A model that runs too slowly on a laptop
Show the answer
Answer: A model that learned its practice examples too exactly and does badly on new ones. Like a student who memorises last year’s exam answers, the model learned the answers, not the subject.
2. How do builders usually check for overfitting?
- They test the model on examples kept back from training
- They train it for twice as long
- They ask the model whether it overfitted
Show the answer
Answer: They test the model on examples kept back from training. Doing well on practice examples but badly on the held-back set is the warning sign.
3. What does the page advise when judging an AI tool?
- Trust the highest published test score
- Pick the tool with the longest name
- Try it on your own tasks, not only its published scores
Show the answer
Answer: Try it on your own tasks, not only its published scores. Public test questions can leak into training data, and your own work is the one exam it cannot have memorised.
Related terms: Mixture of Experts (MoE) · Model collapse · Open weights · Parameters (weights) · Pretraining · Reinforcement learning
- 1Google for Developers, Overfitting, Machine Learning Crash Course, last updated 3 Dec 2025, read at source 1 Oct 2026“Overfitting means creating a model that matches (memorizes) the training set so closely that the model fails to make correct predictions on new data.” … “An overfit model is analogous to an invention that performs well in the lab but is worthless in the real world.” ↑
- 2Google for Developers, Overfitting, read at source 1 Oct 2026“A model that generalizes well makes good predictions on new data.” ↑a“Very broadly speaking, overfitting is caused by one or both of the following problems: The training set doesn’t adequately represent real life data (or the validation set or test set). The model is too complex.” ↑b“Very broadly speaking, overfitting is caused by one or both of the following problems: The training set doesn’t adequately represent real life data (or the validation set or test set). The model is too complex.” ↑c
- 3Google for Developers, Datasets: Dividing the original dataset, Machine Learning Crash Course, last updated 3 Dec 2025, read at source 1 Oct 2026“You should test a model against a different set of examples than those used to train the model.” ↑a“You should test a model against a different set of examples than those used to train the model.” ↑b
- 4Zhang et al., A Careful Examination of Large Language Model Performance on Grade School Arithmetic, arXiv, 1 May 2024, read at source 1 Oct 2026“GSM1k is designed to mirror the style and complexity of the established GSM8k benchmark”. ↑
- 5Zhang et al., arXiv 2405.00332, abstract read at source 1 Oct 2026; it adds that “some models may have partially memorized GSM8k”.↑
- 6Google for Developers, Overfitting, read at source 1 Oct 2026.↑
- 7Google for Developers, Overfitting, Machine Learning Crash Course, read at source 1 Oct 2026“An underfit model doesn’t even make good predictions on the training data.” ↑
Zhang et al., A Careful Examination of Large Language Model Performance on Grade School Arithmetic, arXiv, 1 May 2024, read at source 1 Oct 2026.
The comparisons on this page (a student who memorises last year’s exam, the indoor cat photos) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK