AI training data – what it is and why it's in demand

What is AI training data?

AI training data is the material a model learns from. A language model learns how language works by processing huge amounts of text. An image model learns to recognize subjects through images and descriptions. A model can never be better than the training data it has been given.

Training data can be:

• Text – articles, books, documentation, emails and chats.
• Images and video – photos, illustrations, videos with descriptions.
• Audio – speech, podcasts and recorded calls.
• Code – source code and technical documentation.
• Structured data – tables, registers and business data.

A collection of text used for training is often called a corpus. Read more about the term in what is a corpus.

Three types of training data

Not all training data is used the same way. It helps to distinguish between three stages.

Pre-training

During pre-training, the model learns language and general knowledge from very large amounts of data. Volume and breadth matter most here. Published texts, books and news archives are typical sources.

Fine-tuning

After pre-training, the model is trained further on smaller, more specific data. The aim is to teach it a particular task, domain or way of answering. Quality and relevance outweigh volume here – for example, support tickets with good answers or documentation from a specialist field.

Data for AI agents and reinforcement learning

The newest demand is for AI agents – systems meant to carry out work tasks, not just answer questions. Training them requires data that shows how real people work: how a case is handled from start to finish, how decisions are made and followed up.

AI companies are increasingly buying workplace data such as Slack, email and support tickets for exactly this purpose. In 2025–26, Fast Company, Gizmodo and Forbes, among others, reported on US startups that sold their archives to AI companies after shutting down. Dig deeper in the guide why AI agents are trained on company data.

Why Nordic and European language data is scarce

English dominates the data large models are trained on. Smaller European languages – Swedish, Danish, Norwegian, Finnish and many others – make up a tiny share, and it shows in how well models write and understand them.

A few examples of how the gap is being handled:

• AI Sweden built GPT-SW3 on a Nordic corpus.
• OpenEuroLLM has had to generate synthetic data because there is too little authentic text.
• The National Library of Norway found that licensed books and newspapers improved Norwegian models.

The issue is also political. In Sweden, for example, the government has appointed a national coordinator for Swedish language models, whose remit includes compensation for rights holders.

For European companies, this means authentic text in local languages – especially business communication that has never been published – has a value that is hard to replace.

What makes training data valuable?

Not all material is equally interesting. What usually matters:

• Language – Nordic and other smaller European languages are rare and therefore in demand.
• Uniqueness – data that isn't already on the open web adds something new.
• Reasoning and sequences – threads where problems are solved over time are worth more than isolated messages.
• History and volume – more years and more examples give more to learn from.
• Structure and metadata – dates, categories and outcomes make the data easier to use.
• Clear rights – what you clearly have the right to license is easier to broker.
• Little personal data – the less that needs to be de-identified, the simpler the deal.

Rights and rules

Training data is covered by both copyright and data protection. Under the EU DSM Directive, text and data mining is permitted unless the rights holder has reserved those rights. The EU AI Act (Article 53) requires providers of general-purpose AI models to respect such reservations and publish a summary of their training content.

If the data contains personal data, GDPR applies. That is why licensing, rather than free collection, is the way rights holders can get paid for their data. More on how a license works on our page about data licensing. This is general information, not legal advice.