Training data: what an AI is made of
Training data is the set of examples (text, images, audio) an AI model studies in order to learn. It shapes what the model knows, what it doesn't, and many of its flaws.
Data as of October 22, 2026
Key figures
| Level | Beginner |
|---|
A model only learns what it is shown
To learn, an AI model needs examples. Those examples are its training data. For a spam filter, they are emails that have already been sorted. For a plant identification app, labeled photos. For a large language model, vast amounts of text from which it learns to predict what comes next in a sentence, during pre-training.
A simple comparison: a student who only studied chapter one will ace questions on that chapter and guess the rest. A model behaves the same way. Whatever is missing from its data is missing from its answers.
Three very practical consequences
1. Knowledge has an expiry date. The data is collected at a given moment. Beyond that point, the model knows nothing unless it can browse the web. This is the knowledge cutoff.
2. Flaws in the data become flaws in the model. If some groups, languages or situations are underrepresented, the AI will handle them less well. This is one of the main sources of AI bias.
3. Your data may count. Depending on your settings, what you type into an assistant may be reused to train later versions. In Europe, the GDPR governs how personal data can be used, and France’s data protection authority, the CNIL, publishes practical guidance on this for AI developers.
A workplace example
A company wants a tool that automatically routes customer requests. If it trains the tool only on messages from a quiet year, the tool may misroute the unusual complaints that pour in during a crisis. The fix is not a “better algorithm” but more varied, better-labeled examples.
Myths worth correcting
“AI has read the whole internet.” The data is huge but selected and filtered, and it does not include private pages or everything ever published.
“The AI keeps a copy of its data and looks things up.” In principle, a model does not store a library it can browse: it retains patterns. It can, however, sometimes reproduce memorized passages.
“More data is always better.” Quality matters as much as quantity, which is why carefully controlled synthetic data is gaining ground.
Before you pick an assistant, look at its privacy options in the AI assistants comparison.
Related terms
- AI bias · A systematic error that makes an AI treat some groups unfairly or repeat stereotypes, often inherited from its training data.
- Knowledge cutoff · The date after which an AI model received no more training data, so it knows nothing about later events.
- Pre-training · The first training phase of a language model: it learns to predict the next piece of text across a huge corpus, with no specific task.
- Synthetic data · Artificially created data, often generated by another AI model, used to train, fine-tune or test a model when real data is scarce.
- GDPR and AI · The EU's General Data Protection Regulation, which also applies to AI tools that process personal data.
- Machine learning · A branch of AI where software learns patterns from examples instead of following rules written one by one by a person.
Try it
Full glossary
Frequently asked questions
Are my chats with an AI assistant used as training data?
It depends on the provider, the plan and your settings. Several assistants let you opt out in their privacy or data controls. Check that setting, especially if you share work information.
Where does the training data for large models come from?
Providers generally mention public web pages, books, code, licensed data and data written or labeled by people. The exact details are rarely published. In the European Union, the AI Act requires providers of general-purpose AI models to publish a sufficiently detailed summary of the content used for training.
Why doesn't an AI know about recent news?
Because its training data stops at a certain date, called the knowledge cutoff. For anything newer, the assistant has to search the web, if it can.
Does bad data make an AI less reliable?
Yes. Wrong, lopsided or overly uniform data produces a model that repeats those flaws. That is why sorting and cleaning data is a big part of the work behind every model.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.