Training Data

An AI program doesn’t get its abilities typed in by programmers. It learns them from examples, and these examples are called training data. For a text program, these are vast amounts of written language: websites, books, forums, news articles. For a program meant to recognize cats in photos, it’s millions of images where someone has previously noted what’s depicted in them. The program searches these examples for patterns and stores them in a kind of memory made of numbers. What never appeared in the examples, it can later barely handle either.

Why the dataset determines quality

In computer science there’s an old saying: garbage in, garbage out. For AI, it applies especially strictly. A program has no way to check whether what it has learned is correct. It adopts whatever appears frequently in the data, including all errors and biases.

A well-known example: hiring systems that learned from old personnel files. If these files show that almost only men were hired, the program considers male applicants to be the better candidates. It has no opinion; it simply extrapolates the past. Experts call such a built-in skew bias, meaning distortion.

Quantity matters too. Large language models were trained on amounts of text that a human could not read in a thousand lifetimes. Precisely for this reason, good data has become an economic commodity. Publishers, image databases, and forums sell licenses, and several major lawsuits are underway over content used without permission.

From raw text to finished dataset

Raw data is almost never directly usable. First it’s collected, often with programs that automatically crawl websites. Then it’s filtered: spam removed, duplicates removed, corrupted files removed, and often violence and hate speech as well. In some projects, only a small portion of the collected material survives this cleanup.

Then comes preparation. Texts are broken down into small building blocks so the program can process them. Images are brought to a uniform size. It’s also important to hold back a portion of the data. This test data is later used to check whether the program has truly understood something or has simply memorized it.

Some of the data is created through human labor. People label images, rate answers, or write sample solutions. This is expensive, so companies are increasingly relying on synthetic data: examples generated by an AI itself. This works surprisingly well, but carries a risk. If models are repeatedly trained on the output of other models, errors compound like a photocopy of a photocopy.

Training data in headlines and products

When AI is debated in the news, it very often concerns training data. Authors and artists complain because their works were used without permission. Newspapers like the New York Times have therefore initiated lawsuits, while other publishers have instead signed agreements. The European Union’s AI Act also requires providers to disclose what they trained on.

In everyday life, you encounter the term in the fine print of apps. When you uncheck a box in a chatbot's settings, you’re usually deciding precisely this: whether your inputs may serve as training data. For companies, this is a sensitive topic, because trade secrets can end up with a provider this way.

A common misconception: that training data is stored in the model like in a database. The model stores patterns, not copies, and cannot look anything up specifically. This is precisely why it sometimes invents details. Newer systems get around this by additionally searching real documents when answering – but these are no longer training data, they’re reference sources.