Does AI simply help itself to the internet?

For a language model to learn language, it needs text in a quantity that only the internet can provide: websites, forums, encyclopedias, news. Programs called crawlers download this content automatically. Website operators can state in a file called robots.txt that they do not want this. Whether everyone respects that is disputed. In the EU, providers of large AI models have had to summarize since August 2025 which data they used for training.

The second way is live search. The AI reads current pages and summarizes them. This makes answers more up to date, but not automatically more correct, because it also picks up what is on poor-quality pages.

For publishers, photographers, and authors, this has economic consequences: their content is used without readers coming to their site. Courts in several countries are dealing with this.