Training data

Data Beginner Also called: training set, pretraining data

Definition

Everything an AI model learned from before anyone used it: an enormous pile of text, images, or other data, studied once and baked in. It shapes what the model knows, how it writes, and which mistakes it repeats. Your company's files are not in it unless they were public, which is why the model knows Shakespeare and not your prices.

In plain English

Training data is everything the model read before it met you. A giant pile of text, learned once, baked in.² It explains why the model knows Shakespeare and not your pricing sheet.

In the real world

“Trained on data up to [date]” in a model’s description.
The privacy toggle asking whether your chats can be used for training.
The model confidently describing your industry in general and your company not at all.

When to use it

Reading privacy settings. “Used for training” means your text could become part of the next pile.
Explaining why the model doesn’t know your business. It wasn’t in the data. Give it a project.

When not to use it

Don’t assume “it was trained on the internet” means it knows a specific fact. It’s like a person who read a library, not a database you can look things up in.¹ See hallucination.

Good, bad, ugly

Good: it has read more than anyone you know.
Bad: it learned the average, so it writes the average.
Ugly: whatever was wrong or biased in the pile is in the model, at scale, with confidence.

Key facts

For language models: Enormous amounts of text, then further tuning to make the model helpful
Shapes: What the model knows, how it writes, and which mistakes it repeats
Not included: Anything after the cutoff date, and your private files, unless you add them

Common questions

Does the model learn from my chats?

Depends on the product and your settings. Business plans usually say no by default. Personal plans vary. Look for a setting called something like 'improve the model' or 'use my data for training' and read it.

Is my company's data in the training data?

Only if it was public on the internet or you opted in. The model doesn't know your internal documents unless you give them to it in the chat or a project.

Why does it write the same way every time?

Because it learned from the average of a very large pile of text. The default voice is the average voice. See AI slop.

References

What is training data?IBM · 2025-05-02View source
Glossary: Pretraining; Fine-tuningAnthropic, Claude Platform DocsView source