Fotoğraf: Dev Jadiya, Wikimedia Commons (CC BY 4.0)
What Data Are AI Models Trained On?
From text scraped off the internet to openly licensed archives: where the data models learn from comes from, and the debates it has set off.
Nova AI News Editor
August 15, 2026 · 1 min read
The Raw Material: The Internet Itself
Most large language models are trained on enormous piles of text collected from publicly accessible web pages. Archives like Common Crawl, along with Wikipedia, forum discussions, code repositories, and digitized books, are the known components of that pile. But "collected from the internet" says nothing about the quality of the data.
Cleaning Removes More Than Half of It
Most of the collected data doesn't actually get used. Duplicate pages, spam, automatically generated text, and low-quality translations are filtered out. In languages like Turkish, where the data volume is far smaller than in English, that filtering is even more critical: when a high share of the data is discarded, what's left directly affects the model's command of the language.
Why the Copyright Debate Won't Go Away
A text being openly accessible on the internet doesn't mean it can be freely used as training data. Writers, news organizations, and visual artists have taken that distinction to court. Some model developers are moving toward licensed data agreements, while others are releasing models trained entirely on openly licensed data sets.
Synthetic Data: A Solution, or a Vicious Circle?
One proposed answer to data scarcity is training new models on text the models themselves produced. Used in a controlled way it works, but taken too far it reduces the model's diversity and copies errors from one generation to the next. Researchers call this "model collapse."
Conclusion
What a model doesn't learn is determined by its training data just as much as what it knows. Transparency about sources isn't yet an industry standard, but pressure from users and regulators is pushing in that direction.
Related Articles
How Big Is AI's Energy Bill?
The energy cost of model training versus everyday use, the load on data centers, and the concrete steps being taken to cut consumption.
Read more→
Where Is AI Actually Used in Game Development?
From enemy behavior to procedural world generation and voice acting, AI's real role inside game studios.
Read more→
AI in Finance: How Is Fraud Caught?
What are the risk models working behind a payment approved in seconds actually doing — and why do they sometimes block innocent transactions?
Read more→Comments
No comments yet — be the first to comment.