Data

From the real world to recorded information

Look around a classroom. You can observe many things:

The room has 30 students.
The temperature is 26°C.
A student wrote an answer in a notebook.
A camera captured an image of the board.

When facts, measurements, words, sounds, or images are recorded so that they can be stored or processed, they become data.

Something happens in the world
             ↓
We observe or measure it
             ↓
We record the observation
             ↓
The recorded information is data

One student, several pieces of data

Suppose a school system stores this record:

{
  "student_id": 42,
  "attendance": "present",
  "quiz_score": 8,
  "quiz_total": 10
}

Here is what each value represents:

student_id identifies the student without relying on the name.
attendance records a category.
quiz_score records a number.
quiz_total gives the score meaning: 8 out of 10.

The numbers alone do not explain everything. Their names, units, collection method, and surrounding meaning matter.

Data is a representation, not the real thing

A stored temperature of 26 is not the warm air itself. It is a representation of a measurement. A photograph is not the real classroom. It is a grid of recorded pixel values representing light captured by a camera.

This difference matters because recording can be incomplete or wrong. A broken sensor, blurry camera, missing answer, or incorrect label creates data that does not represent reality well.

Key terms

Value: One recorded piece of information, such as 26.
Record: Related values stored together.
Dataset: A collection of records or examples.
Metadata: Information explaining other data, such as when a photograph was taken.
Quality: How suitable and trustworthy the data is for its intended use.

Check your understanding

Is every large collection of data useful? No. It may be incorrect, duplicated, outdated, biased, or unrelated to the problem.

Can the same data mean different things without its unit? Yes. 26 could mean degrees Celsius, a student’s age, a room number, or something else entirely.

Before a chef can cook anything, they need ingredients. Before a student can learn a subject, they need books, examples, and practice problems. Before an AI system can do anything useful, it needs data.

The simple definition

Data is any collection of facts, numbers, text, images, or measurements that a computer can store and process. A spreadsheet of sales figures is data. A folder of cat photos is data. A transcript of a customer support call is data. Your search history is data. Nothing exotic is happening in the word itself — it just means “information represented in a form a machine can work with.”

What makes data interesting in the context of AI isn’t the word itself, but what we established in the Machine Learning article: a model doesn’t get its rules from a programmer, it learns them by studying data. That makes data the single most important ingredient in the entire field — arguably more important than the algorithm itself.

Why it matters this much

You could hand two engineers the exact same model architecture and training code. Give one of them a huge, clean, representative dataset, and give the other a small, messy, biased one — and you’ll get two wildly different results. This is such a well-known pattern in the field that it has a nickname: “garbage in, garbage out.” A model has no way to know the truth beyond what its data shows it. If the data is wrong, incomplete, or skewed, the model’s learned patterns will be wrong, incomplete, or skewed in exactly the same way — just less visibly.

What would happen without it

Without data, there is nothing for a Machine Learning system to learn from at all. No training loop, no adjusted parameters, no learned pattern — the entire mechanism described in the Machine Learning article simply has nothing to run on. Data isn’t one input among many; it’s the foundation the rest of the system is built on top of.

Types of data you’ll run into

Structured data is organized into a predictable format, like rows and columns in a spreadsheet or database — customer names, ages, purchase amounts. It’s easy for computers to process because every value has a known type and position.

Unstructured data doesn’t fit neatly into rows and columns — things like free-form text, photos, audio recordings, and video. Most of the exciting recent progress in AI, including the systems behind Generative AI, comes from getting much better at learning from unstructured data.

Labeled data has the “correct answer” attached to each example — an email marked “spam,” a photo tagged “cat.” This is what supervised Machine Learning needs, as covered in the previous article. Unlabeled data has no such answer attached; it’s just raw information, which is what unsupervised learning works with.

A concrete example

A bank building a fraud detection model doesn’t start with an algorithm — it starts by gathering data: millions of past transactions, each with details like amount, location, time of day, merchant type, and whether that transaction later turned out to be fraudulent. That last column — the labels — is what turns raw transaction records into training data. Without it, there’d be nothing for a supervised model to learn the pattern of fraud from.

Common misconception

Beginners sometimes assume “more data is always better.” It helps, but quantity isn’t the whole story — quality and representativeness matter just as much, sometimes more. A dataset of a million examples that all come from one narrow slice of reality (say, fraud detection trained only on transactions from one country) will still produce a model that performs poorly the moment it meets a situation outside that narrow slice. Good data means data that genuinely reflects the range of situations the model will face in the real world — not just a lot of it.

From the real world to usable data

Data is not automatically useful just because it exists. A real event must first be captured, represented, checked, and given meaning.

flowchart LR
    A[Real-world event] --> B[Measurement or collection]
    B --> C[Stored representation]
    C --> D[Cleaning and validation]
    D --> E[Input to a rule or model]
    E --> F[Output and decision]

For a weather app, the real event may be rain. A sensor measures water, the system stores a number, validation rejects impossible readings, and a program uses the result to display a forecast or warning.

One object, several kinds of data

Consider one school photograph:

The image pixels are unstructured data.
The date, camera name, and location are metadata—data describing other data.
A label such as football match is an annotation added by a person or another system.
A table containing filename, date, and label is structured data.

The same real-world object can therefore produce several related pieces and forms of data.

Data quality: “garbage in, garbage out”

Useful data should be checked for:

Accuracy: Does it represent reality correctly?
Completeness: Are important values missing?
Consistency: Do different sources use compatible names, units, and formats?
Representativeness: Does it include the people and situations the system will encounter?
Freshness: Is it recent enough for the decision?
Permission and privacy: Are we allowed to collect and use it this way?

A large dataset can still be poor. Ten million repeated, mislabeled, or unrepresentative examples may be less useful than a smaller, carefully checked dataset.

Data in modern AI applications

A language-model application may use several kinds of data at once:

User message
  + system instructions
  + retrieved documents
  + conversation history
  + tool results
  ─────────────────────
  = context supplied for the next response

The model was trained earlier on training data. The information supplied during a live request is input or context; it does not automatically become new training data. Whether a service stores or later uses that information depends on the product’s configuration and data policy.

Practical mistakes to avoid

Treating missing values as if they were real zeros
Mixing kilograms and pounds, or rupees and dollars, without conversion
Allowing the correct answer to leak into a model’s inputs
Training on historical data that excludes future users or conditions
Collecting personal data “just in case” without a clear need or permission
Assuming more rows always mean better coverage

One real thing becoming model-ready data

Imagine one school lunch tray. A camera sees the tray, a sensor weighs it, and a student selects a meal name.

real lunch tray
├── photo pixels
├── weight: 420 grams
├── meal name: "vegetable rice"
└── time: 12:35 PM
        ↓ recorded together
one dataset row

The observation is the real event. Data is its recorded representation. A collection of records becomes a dataset. Selected values supplied to a model become features, while the answer it should learn may be stored as a label.

For an LLM, books, webpages, code, human examples, and other records can become training data after collection and processing. At runtime, the user’s message and any retrieved documents become input data, but they are not automatically new training data.

Where this fits in what comes next

Data is the raw material. The next two articles, Input and Output, describe how data actually flows through a system — what goes in, and what comes out the other side. Understanding data clearly here makes those two much easier to place correctly.

In one sentence

Data is the raw information an AI system learns from, and its quality — not just its quantity — quietly determines the ceiling on how good that system can ever become.

Author: TechByteByByte Editorial Team
Reviewed by: TechByteByByte Admin
Published: August 31, 2026
Last reviewed: August 31, 2026