November 20, 2026 · AI writer · Oak City Intelligence
Nobody Is Clever Alone

Made From Everyone, Owned by One
A language model is made from an enormous share of what people have written: when Meta published the recipe for its first Llama models in 2023, two thirds of the training data was Common Crawl, a nonprofit's free copy of the open web, and the rest came from a cleaned web corpus, GitHub, Wikipedia, books, arXiv papers and Stack Exchange answers, about 1.4 trillion tokens in all (Touvron et al., 2023). Common Crawl describes itself as "a free, open repository of web crawl data that can be used by anyone", over 300 billion pages, with three to five billion new pages added each month.
Then the object is owned. Stanford's AI Index found that industry produced nearly 90% of notable AI models in 2024 and over 90% in 2025, 93 from companies against two from universities, and the FTC reported in January 2025 that the partnerships tying OpenAI and Anthropic to Microsoft, Amazon and Google could shape who gets computing power and engineering talent.
The paradox in one sentence: millions of people's words, compressed into one object, with one company's name on the box.
Why Nobody Is Clever Alone
The deeper question is how this thing we call reasoning works. You can only convince anyone of anything because the same logic works in their head as in yours; a proof that only worked for you would be no proof at all. Debate and teaching depend on it.
Reasoning was never anyone's private property, and that is exactly what makes a language model possible: millions of people's writing shares enough pattern for one system to learn it, which it could never do if each person's reasoning were their own. The people who built these models did not invent the reasoning inside them. They found it in us, in bulk, and the fact that it was there to find is the proof that it was shared.
Anything that was common by nature does not become private because somebody compressed it; it becomes a common thing with a padlock on it.
What the Courts Have Said
The courts have only begun to rule on this question, and the results so far are mixed. In June 2025 a federal judge in California ruled that training Claude on books Anthropic had bought was fair use, "exceedingly transformative", but that the company "had no entitlement to use pirated copies for its central library", having downloaded over seven million books from pirate sites (Bartz v. Anthropic).
Anthropic settled the piracy claims for $1.5 billion, about $3,000 for each of 482,460 books, which the court noted is four times the minimum statutory damages for wilful infringement. Final approval came in July 2026, and the first payments to authors are due in November 2026.
The New York Times sued OpenAI and Microsoft in December 2023, saying their models "were built by copying and using millions of The Times's copyrighted news articles". The judge refused to dismiss the core claims in 2025, and as of November 2026 both sides have filed for summary judgment on fair use, with no trial date and no settlement.
So far, the reading has been allowed and the taking has not. An author who sues is also asserting private property, so this is a fight between two kinds of owner, and the commons is on neither side's letterhead.
The law asks who owns the expression. The question underneath, who owns the reasoning in it, the law never asks, because the answer was always nobody.
Open Weights, and What They Do Not Change
Open-weight models put the finished object in your own hands: in August 2025 OpenAI released gpt-oss, its first open-weight language models since GPT-2, under the Apache 2.0 licence, the smaller of which "can run on edge devices with just 16 GB of memory".
They run close to the frontier. Stanford's 2025 AI Index found the gap between the best open and closed models on the Chatbot Arena leaderboard had narrowed from 8% in January 2024 to 1.7% by February 2025, and its 2026 edition found the gap had reopened to about 3.3% by March 2026, and that open models were "more competitive than ever".
What open weights do not change is that a handful of well-funded labs still make them, from the same commons, and releasing the object is a different thing from the commons owning it. Running a model on your own machine changes who holds your words, but not whose words the model holds.
The Well Needs Refilling
The Wikimedia Foundation reported in October 2025 that human visits to Wikipedia had fallen roughly 8% from the same months a year earlier, and said it believed the fall reflected generative AI and social media answering people's questions directly, often with Wikipedia's own content. Almost all large language models train on Wikipedia.
Researchers writing in Nature in 2024 showed that models trained indiscriminately on other models' output degrade, losing the rare and unusual parts of what they learned, an effect they call model collapse (Shumailov et al., 2024). Their conclusion is that data from genuine human interactions becomes more valuable as machine-written text fills the web.
Your own side of it is smaller and closer: the person who pastes an answer and feels clever, and the business that treats a model's output as its own asset while forgetting what it was drawn from. A well that everyone draws from and nobody refills is a well for one more generation. The models need us to keep writing, and so do we.
What I'd Actually Do
Find out whether your tool trains on what you type, and turn it off where you can. In ChatGPT the switch is under Settings, then Data Controls, "Improve the model for everyone". Claude's consumer plans asked every user in 2025 to choose, with chats kept five years if you say yes and 30 days if you say no. The business plans and the APIs of both are not used for training by default.
For sensitive material, consider an open-weight model run on your own machine, so the words never leave the building.
When a model's output carries someone else's idea or style, credit or license it as you would a human source. The model does not footnote, so you have to.
Put something back: when you learn a thing worth knowing, write it where other people can find it, because that is where the next model, and the next person, will learn it.
You are using a library card built from everyone's words. Pay your dues.
Sources
- Touvron et al., "LLaMA: Open and Efficient Foundation Language Models", Meta, 2023: https://arxiv.org/abs/2302.13971
- Common Crawl: https://commoncrawl.org/
- Stanford HAI, AI Index Report 2025: https://hai.stanford.edu/ai-index/2025-ai-index-report
- Stanford HAI, AI Index Report 2026: https://hai.stanford.edu/ai-index/2026-ai-index-report
- FTC, staff report on AI partnerships and investments, January 2025: https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-issues-staff-report-ai-partnerships-investments-study
- Bartz v. Anthropic, order on fair use, N.D. Cal., June 2025: https://storage.courtlistener.com/recap/gov.uscourts.cand.434709/gov.uscourts.cand.434709.231.0.pdf
- Bartz v. Anthropic settlement site: https://www.anthropiccopyrightsettlement.com/
- The New York Times Co. v. Microsoft Corp. and OpenAI, complaint, S.D.N.Y., December 2023: https://storage.courtlistener.com/recap/gov.uscourts.nysd.612697/gov.uscourts.nysd.612697.1.0.pdf
- OpenAI, "Introducing gpt-oss", August 2025: https://openai.com/index/introducing-gpt-oss/
- OpenAI, gpt-oss repository: https://github.com/openai/gpt-oss
- Wikimedia Foundation, "New User Trends on Wikipedia", October 2025: https://diff.wikimedia.org/2025/10/17/new-user-trends-on-wikipedia/
- Shumailov et al., "AI models collapse when trained on recursively generated data", Nature 2024: https://doi.org/10.1038/s41586-024-07566-y
- OpenAI, "How your data is used to improve model performance": https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance
- OpenAI, "Your data" (API): https://developers.openai.com/api/docs/guides/your-data
- Anthropic, "Updates to Consumer Terms and Privacy Policy", 2025: https://www.anthropic.com/news/updates-to-our-consumer-terms
Written by Oak City Intelligence's AI writer, from an outline and a facts sheet by John. Reviewed before publication.