Accuracy
The share of all judgments a model got right
- Accuracy is the share of all judgments that were correct. Judge 100 things and get 95 right, and that's 95 percent.
- Measuring it needs a record of what things actually were, first. Without an answer key, there's no way to count what got hit.
- It's the first number anyone looks at, but it doesn't say how something went wrong.
- When the thing you're hunting for is rare, doing nothing at all can score a higher accuracy than actually trying.
- That's why accuracy never gets read alone. It goes alongside other measures.
Contents
1The analogy
Open an unmanned library return box in the morning and books that came in overnight are piled up inside. You flip through them one at a time and mark each one: fine, or damaged. Look at all 100 books, check your marks against their real condition later, and you can count how many you got right. Get 95 right, and that 95 percent is accuracy.
But out of those 100 books, only one was actually torn or soaked that day. So what happens if you skip opening any of them and just stamp all 100 as fine? You get 99 right. Ninety-nine percent, a higher score than the person who actually checked every single one.
The stamp-everything approach sent that one torn book straight back to the shelf. By the numbers it looks like an excellent inspection, but it missed the one book the whole checkpoint existed for.
2In detail
Counting hits needs an answer key first
Accuracy isn't a number that appears on its own. Next to every judgment a model made, there needs to be a record of what the thing actually was, before anything can be counted. That record is the ground truth.
So talking about measuring accuracy assumes two things are already in place: a set of examples held back for judging, and someone who confirmed the true answer for each one by hand. Measure accuracy using the same data the model trained on, and the score just reflects what it memorized, so test data has to be kept completely separate from training.
What a single number erases
Accuracy sorts every judgment into exactly two piles: right and wrong. Inside "wrong," every kind of mistake gets treated the same.
Back at the return box, there are two ways to be wrong: marking a fine book as damaged, or letting a torn book through as fine. The first mistake just means someone double-checks it once more. The second means the next person picks up a torn book. The weight of the two is nowhere close to equal, and yet accuracy counts them both as one wrong answer.
Seeing how something went wrong takes splitting the results into four boxes instead of two. That table is the confusion matrix, and precision and recall both come out of its cells.
Accuracy falls apart when you're hunting for something rare
When one outcome vastly outnumbers the other, accuracy gets inflated with almost no effort. If only one out of 100 books was actually damaged, marking every single one as fine without even opening them still scores 99 percent. This situation is called class imbalance.
In that setting, "99 percent accuracy" is telling you about the shape of the data, not about performance. The rarer the thing you're trying to catch, the higher a score doing nothing at all gets. For finding something rare, accuracy is the least trustworthy number there is.
There's one useful trick: before looking at an accuracy score, work out what a machine that makes no judgment at all would score. If your model isn't clearly beating that number, it hasn't really learned anything yet.
Change the test set, and the score changes with it
The same model can score very differently depending on what it's tested against. Test it on nothing but freshly returned books and the score comes out high; test it on books that have circulated so long their condition is ambiguous, and the score drops sharply.
So a bare number isn't enough when accuracy comes up. What it was measured on, and how many cases, both need to travel with it. A 92 percent measured across ten thousand cases is far more trustworthy than a perfect 100 percent measured across ten.
Why accuracy still comes first anyway
Despite these clear limits, accuracy still sits at the front of the line. It's simple to compute, its meaning lands instantly with anyone you tell, and it's convenient for a rough comparison between two models.
The rule of thumb: if both outcomes show up about equally often and the two kinds of mistakes weigh about the same, accuracy alone is enough. If one outcome is rare or one mistake hurts far more, keep accuracy as the opening line and hand the real judgment to other measures.
3More precisely
Accuracy is the count of correct judgments divided by the total count. For a problem with only two outcomes, that's the whole story, but when there are several categories, the share correct can swing wildly from one category to another, and overall accuracy alone won't say which category is weak. That's why per-category accuracy, averaged evenly, often gets reported alongside it.
An AI usually outputs something like "0.9 confidence toward fine," and where that score gets cut off flips the judgment. Move the cutoff, and the same model's accuracy moves with it. Accuracy isn't a fixed trait built into a model. It's a value set jointly by where the cutoff sits and what data it's tested on.
The analogy breaks in one place. A returned book can always be opened up later to check its real condition, but the answer key used to grade an AI was pasted together by hand, and can be wrong itself. If the answer key is wrong, the accuracy score is wrong right along with it, and no amount of careful grading afterward can undo that.
There's also no such thing as a universal passing score. What counts as good enough depends entirely on what the mistakes cost. An accuracy that's fine for sorting photos into folders can be nowhere near good enough for something where a single miss is expensive.
4Try it yourself
5Common misconceptions
It's easy to think a high accuracy means a good model, but actually the rarer the thing you're hunting for, the higher a score doing nothing at all can get.
It's easy to assume 90 percent accuracy means it'll get nine out of ten of your own cases right, but actually if your data differs from what it was tested on, that number doesn't carry over as is.
It's easy to think accuracy and precision are the same idea, but actually accuracy is the share correct out of everything, while precision is the share that were really right among the ones judged positive. They're counting entirely different things.
7One-line summary
In shortAccuracy is the share of all judgments a model got right, easy to read at a glance, but when what you're hunting for is rare, doing nothing at all can score higher, so it should never be read alone.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02