Note: This is part three of three related posts.
How did the creators of AI obtain the material that the AI was trained on?
ChatGPT said:
AI developers obtained their training material in several different ways. Some material was purchased or licensed, some was freely available, and some was copied from the internet without obtaining permission from individual copyright holders. The proportions vary considerably among AI companies, and the complete contents of most major training datasets have not been publicly disclosed.
Here are the principal methods.
1. Collecting material from the public internet
This has been one of the most important sources of AI training data.
Developers used automated programs called web crawlers to collect enormous quantities of material from publicly accessible websites.
This included news articles, blogs, discussion forums, educational resources, books available online, and other written material.
One major source was Common Crawl, a nonprofit organization that maintains a massive archive of web content.
Common Crawl makes its datasets available for research and other uses. However, publicly accessible does not necessarily mean free of copyright restrictions.
A newspaper article, for example, can be freely readable on the internet while remaining copyrighted.
AI developers have argued that using such material for training can qualify as fair use under U.S. copyright law. Copyright holders have challenged that interpretation.
2. Using books, including unauthorized digital copies
Books have been particularly valuable for training language models because they contain sustained, carefully edited writing.
Some AI training datasets have included collections of books obtained from unauthorized online repositories.
One widely discussed example is Books3, a dataset containing approximately 196,000 books. It was used in training some AI models and became the subject of copyright litigation.
There have also been allegations and documented instances involving books obtained through shadow libraries such as LibGen.
This is an important distinction: Even if training on copyrighted material can qualify as fair use, obtaining that material through unauthorized copying raises a separate legal question.
3. Licensing material from copyright holders
Some AI companies have entered into commercial agreements with publishers, news organizations, and other content owners.
For example, OpenAI has announced licensing partnerships with publishers including the Associated Press, Financial Times, and News Corp.
These arrangements can involve payment for access to content and permission to use it in developing AI products.
The financial terms of many agreements are confidential.
This approach is much closer to the historical economic relationship you described, in which creators receive compensation through the sale or licensing of their work.
4. Using freely licensed and public-domain material
AI developers also use substantial amounts of material that can legally be copied and reused without negotiating individual copyright licenses.
Examples include older books whose copyrights have expired, government publications, and material distributed under certain open licenses.
Wikipedia is another important source, although its Creative Commons license imposes conditions on reuse.
Computer programming code distributed under open-source licenses has also been used extensively.
5. Using material created specifically for training
AI developers have also paid people to write examples, evaluate responses, annotate material, and help train models.
More recently, developers have increasingly used synthetic data—material generated by AI systems themselves.
These methods reduce reliance on externally created copyrighted works, although synthetic data may still indirectly reflect the material used to train earlier models.
What about ChatGPT specifically?
OpenAI has publicly described three broad sources of training information:
- Information publicly available on the internet.
- Information obtained through partnerships with third parties.
- Information provided or generated by users, human trainers, and researchers.
However, OpenAI has not published a complete inventory of every copyrighted work used to train its major models.
Consequently, I cannot tell you precisely how many books were purchased, how many were licensed, or how many were collected without direct payment to their creators.
And that uncertainty is important to the discussion we’re having.
Returning to your original argument
Your observation was that throughout your life, most of the information you acquired came through economic arrangements that compensated the people who created it.
The development of AI has not consistently followed that model.
Some training material was purchased or licensed. Other material was collected without payment to its creators.
But I think we should distinguish two questions.
Question 1: Was the material obtained legally?
Question 2: Even if it was obtained legally, should its creators receive additional compensation when it is used to train a commercially valuable AI?
Those are separate issues.
And there’s a third question that may be even more consequential:
If a company buys one copy of a book for $20, should it be permitted to use that book to train a model that subsequently serves 100 million people?
A human reader can purchase the same book for $20, absorb its contents, and use that knowledge to create something commercially valuable.
Your original analogy suggests that AI should perhaps enjoy similar freedom.
But your observation about the historical compensation of authors suggests that acquiring the original material legitimately matters.
I’m curious whether knowing how AI developers actually obtained their training material changes your position. It seems to me that your argument could support a distinction between AI companies that lawfully acquire their training material and those that do not, without requiring a completely new theory of intellectual property.
At this point, I’m not sure I have a position.
Leave a Reply