SagelixSAGELIX
How it worksStoriesThe archiveBlog
Log in
  1. Sagelix
  2. Blog
  3. Where what an AI model knows comes from
Back to blog
Para empresasPublished on November 18, 20265 min read

Where what an AI model knows comes from

Las manos de un profesional firmando un documento

«Data from the internet» is ceasing to be an answer. What audits, the law and the courts say about the provenance of the data AI is trained on.

S

Sagelix

Sagelix

Image generated with artificial intelligence.

For years, the answer to «what was this trained on?» was «data from the internet». That answer is ceasing to hold, for reasons that can be documented.

Not even curated data comes with its licence

An audit reviewed more than 1,800 text datasets, of the kind used to train and fine-tune models. In the most widely used repositories, more than 70% did not state the licence under which they were distributed, and more than 50% stated it incorrectly (Longpre et al., 2024).

It is not talking about material crawled indiscriminately. It is talking about collections prepared and published to be reused. Not even there is it known with what permission the content arrives.

Publishers are withdrawing permission

Another study reviewed 14,000 web domains present in three large training corpora. In a single year, more than 28% of the most important and best-maintained sources in one of them became restricted. If each site’s terms of use are taken into account, 45% of that corpus is already restricted (Longpre et al., 2024b). We consulted this paper as a preprint.

The internet, therefore, is not a neutral source. A growing share of those who write on it have said no.

The law and the courts are already asking

  • Europe. Since August 2025, providers of general-purpose models must have a policy for complying with copyright and publish a sufficiently detailed summary of the content used for training (AI Act, Article 53). In July 2025 the Commission published the template for that summary, which asks for the main data collections to be listed and the remaining sources to be explained. The obligation falls on whoever makes the base model, not on whoever builds a product on top of it.
  • United States. In July 2026 a court approved a 1.5 billion dollar settlement between an AI company and book authors, of about 3,000 dollars per work, with the obligation to destroy the copies obtained from two pirate libraries. What was paid for was the origin of the material. The figure comes from the authors’ guild, which is an interested party.
  • Germany. In November 2025 a Munich court ruled that memorising song lyrics inside a model infringes the right of reproduction. The ruling is not final, and we know it from a law firm’s summary, not from the original text.

The framework is not settled. The United States Copyright Office notes that there are already systems trained only on licensed or public domain works. At the same time it acknowledges that the growth of the licensing market does not prove that licensing everything is viable at scale.

Why it matters that there is a person behind it

There is a technical reason as well as the legal one. Training models on content generated by other models produces defects that accumulate and are not corrected, and the authors of that paper conclude that data coming from real people will be increasingly valuable (Shumailov et al., 2023).

There is also evidence that material from a specific field improves results in that field: training only on biomedical text gave substantial gains over starting from a general model (Gu et al., 2021). That study compares texts, not knowledge contributed by professionals.

Good sources do not guarantee correct answers

It is worth not overstating the case.

  • Legal tools built on professional sources were wrong between 17% and 33% of the time, and their authors conclude that what the vendors promised was overstated (Magesh et al., 2024).
  • A well-steered general model outperformed specialised medical models on multiple-choice exams (Nori et al., 2023).
  • We have not found any study that compares, in a real product, knowledge contributed by veteran professionals with data from the web.

Having a clear provenance does not make a system get things right. It makes it possible to know where what it says comes from, and with what permission.

What it means for those who contribute what they know

What changes is the question. An AI product is going to be asked, by its customers and increasingly in the courts, to say where what it asserts comes from. Knowledge with an identified author and written permission for use can answer it; knowledge that was simply collected cannot.

At Sagelix each volume carries the name of the professional who contributed it, and each purchase leaves a record of the permission for use with its conditions.

Sources

  • Longpre et al. (2024). «A large-scale audit of dataset licensing and attribution in AI». Nature Machine Intelligence 6. Preprint
  • Longpre et al. (2024b). «Consent in Crisis: The Rapid Decline of the AI Data Commons». arXiv:2407.14933
  • Regulation (EU) 2024/1689, Article 53. Text
  • European Commission (2025). Template for the public summary of training content
  • Authors Guild (2026). Statement on the approval of the settlement in Bartz v. Anthropic
  • CMS (2025). Summary of the judgment of the Landgericht München I, 42 O 14139/24
  • U.S. Copyright Office (2025). «Copyright and Artificial Intelligence, Part 3: Generative AI Training», pre-publication version
  • Shumailov et al. (2023). «The Curse of Recursion». arXiv:2305.17493
  • Gu et al. (2021). «Domain-Specific Language Model Pretraining for Biomedical NLP». arXiv:2007.15779
  • Magesh et al. (2024). «Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools». arXiv:2405.20362
  • Nori et al. (2023). «Can Generalist Foundation Models Outcompete Special-Purpose Tuning?». arXiv:2311.16452
Share

Related articles

Why Every Industry's AI Needs Real Experts
AI Trends

Why Every Industry's AI Needs Real Experts

5 min
Essential Digital Tools for Senior Professionals
Tools

Essential Digital Tools for Senior Professionals

1 min
Knowledge gap: what is lost when the person who knew retires
AI Trends

Knowledge gap: what is lost when the person who knew retires

2 min

Did you enjoy this article?

Join Sagelix and preserve your professional experience for future generations.

Start for free

Stories from the archive

© 2026 Sagelix — the living archive

A Quantum Howl project
PrivacyTermsContact