Data
The row-by-row record that becomes the raw material AI learns from
- Data is a record of something that happened, written down. Everything AI learns from comes from here.
- One row is one event, and each column in that row is a trait of that event.
- It's not only numbers. Text, photos, sound, and click trails all become the same raw material once they're captured in a column.
- When a row has the correct answer written in alongside it, that's labeled data. Supervised learning grows by eating rows like these.
- Whatever never got written down might as well never have happened. A missing record is invisible to AI from the very start.
Contents
1The analogy
Open a budget notebook and each line holds one entry: May 3rd, subway, $1.75. The line below it: May 3rd, a sandwich, $6. One line by itself means almost nothing. But once a hundred lines pile up, this month's transit spending and whether weekends spike become obvious on their own.
Data is exactly this, line by line. Whatever happened before it got written down can't be worked with, no matter how big it was, and the moment it gets written down, it becomes something you can count, compare, and learn from. Skip the date on one line, or mistype an amount, and the total gets thrown off right along with it — the ledger is only ever as honest as what actually made it onto the page.
2In detail
One row is an event, each column is a trait
A line in a budget notebook has the date, where it was spent, the amount, and the payment method sitting side by side. One row holds "one thing that happened that day" in full, and each column is a piece of that event. When data gets drawn as a table, the horizontal lines are called rows and the vertical ones columns.
Building AI, these columns get separate names. The columns used to predict something are called features; the column holding the answer to guess is called the label. The same table becomes a completely different problem depending on which column you set as the answer — set the amount as the answer and it's a spending-prediction problem, set the payment method as the answer and it's a payment-method problem.
How you design the columns matters a lot for performance. Leaving "May 3rd" as a raw date is different from pulling out the day of the week and whether it's a weekend into a new column. Spending spiking on weekends only becomes visible once there's a weekday column to look at.
It doesn't have to be a number
A photo is data too. Every dot on screen is recorded as a brightness value, and a photo is those values lined up in a grid. Sound is the same, a line of values rising and falling over time, and text gets broken into pieces and swapped for numbers.
The trail left behind while using an app is data too. Which screen got opened, how many seconds someone stayed, what got skipped without a tap — all of it piles up as rows. A machine writing its own record without a person lifting a finger makes up a large share of today's data.
Different shapes still funnel into one method. Whatever the raw material, it has to be turned into a bundle of numbers before it can go into a model, which is exactly why one model can handle photos, sound, and text together.
Rows with an answer, and rows without
Imagine marking some lines in a budget notebook by hand: "this one was necessary," "this one was a waste." That mark is a label. With enough labeled rows, you can teach an AI "a line like this one is a waste."
Attaching labels is usually a person's job — naming an object in a photo, marking the emotion in a sentence, writing down what a sound actually is. It takes a lot of hands and a lot of cost, and often labeling turns out to be harder than collecting the data in the first place.
So methods for using unlabeled records as they are grew alongside it too — grouping similar rows to find clusters, or hiding part of a sentence and having the model learn by guessing what's missing, making up its own answer to learn from.
Whatever never got written down might as well never have happened
Suppose a budget notebook only records card payments and skips cash. The monthly total it produces will come out lower than reality, not because the math was wrong, but because the raw material was already leaning one way.
The same thing happens with AI. Collect only photos taken in daylight, and night stops existing as far as the model is concerned. Collect only one region's accent, and every other way of speaking never gets a chance to be learned. A model doesn't imagine and fill in what isn't in the data. The gap just stays empty, and a wrong answer comes out of that empty spot.
That's why looking at data means checking what's missing just as much as what's there. Who never got recorded, and what situation is entirely absent, checking that often matters more than the performance score.
Cleaning takes longer than collecting
Real records are always messy. Dates are formatted differently line to line, the same store name gets spelled a few different ways, and some rows have an empty amount column. Some lines got entered twice by accident.
Straightening all of this out is called data cleaning: deciding what to do with blank cells, merging different spellings of the same thing into one, and looking into values that seem wildly out of range. This step is usually where the most time goes in actual practice.
When the raw material isn't consistent, results wobble no matter how good the model is. That's also why cleaning the data often improves performance more than switching models does.
3More precisely
Data is a collection of values recorded in a fixed format from something observed. Data with clear rows and columns, like a table, gets called structured data; data without that kind of frame, like a photo or a document, gets called unstructured. A collection gathered and organized specifically for training, for a particular purpose, gets its own separate name: a dataset. Data and dataset get used interchangeably a lot, but the first is the raw material itself, and the second is a package of that material curated for one specific problem.
The analogy breaks down in places. A person writing their own budget notebook has a rough sense of what's missing; data used by AI is usually something someone else collected long ago, and it can be hard to even know what standard was used to filter it. A single line in a budget notebook carries meaning on its own, but the values a model receives have that meaning stripped away into plain numbers, which makes it much harder for a person to scan by eye and spot something off. That's why documenting where and how data was collected, and what got left out, matters so much in practice.
4Try it yourself
- Teachable Machine ailearn.space A few photos you take become training material immediately — add more photos or change the background and watch how the result shifts
- Google's Machine Learning Crash Course (Korean-language) ailearn.space Walks through, in diagrams, how a table of rows and columns splits into features and labels
5Common misconceptions
It's easy to think data means numbers, but actually text, photos, sound, and click trails are all data too, and today more of what AI learns from actually comes from this side.
It's easy to think data is neutral because it's just the plain facts, but actually a person's choices are already baked in from the moment someone decides what to record and what to leave out.
It's easy to think more collected is always better, but actually piling up more duplicated, wrong, or lopsided records can make performance worse, not better.
7One-line summary
In shortData is the record, line by line, of what happened in the world, and AI can only ever learn what made it onto that line.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02