AI

Fed on the Internet

Every AI model you use was trained on something. Here's where that something actually comes from, and why it matters more than most people realize.

James

— 4 min read
Share
Fed on the Internet
Every word the model learned came from somewhere.

Every AI model you use was trained on something. Here's where that something actually comes from, and why it matters more than most people realize

The Fuel Nobody Talks About

You've heard a lot about what AI can do. You've probably heard less about what it was built on.

Every large language model, including ChatGPT, Claude, Gemini, and Llama, learned to write, reason, and respond by consuming enormous volumes of text. We're talking about datasets so large they would take tens of millions of years for a human to read. Meta has publicly acknowledged one of its recent models was trained on roughly 15 trillion tokens of text. That's not a typo.

The question worth asking is: where did all of that come from? The honest answer is that it came from the internet, from books, from you, and the rules around all of it are still being written in real time.

🚀 MARTY SAYS

"Every model is only as good as what it was fed. That means understanding AI also means understanding its diet."

The Four Main Ingredients

Most AI models pull from the same general recipe, with some variation by company:

1. The open web, crawled without asking
The foundation of nearly every major model is a massive sweep of publicly available internet content. Common Crawl, a nonprofit that has been archiving the web since 2008, provides a dataset that almost every major AI lab has used. OpenAI runs its own crawler called GPTBot. Anthropic runs one called ClaudeBot. Google uses Google-Extended. These bots visit websites, read the content, and add it to the training pile. Website owners can technically block them using a file called robots.txt, but compliance is voluntary, and not every company respects it.

2. Licensed deals with publishers and platforms
As the easy web data gets mined out, companies have turned to licensing agreements. OpenAI has deals with publishers including the Financial Times, Le Monde, and Axel Springer. Stack Overflow has partnered with multiple AI companies to provide technical content. Disney signed a deal with OpenAI late last year allowing its characters to be used in AI-generated video. These deals are typically confidential. You rarely find out about them unless they leak or get announced as press releases.

3. Books, legally murky and heavily used
Multiple AI companies have been sued over their use of books in training data. Courts are actively sorting through whether training on copyrighted books constitutes fair use. In Bartz v. Anthropic, a federal judge ruled that AI training on lawfully acquired books was "quintessentially transformative" and qualified as fair use, but that storing pirated copies did not. That case settled for $1.5 billion in 2025, with final approval in May 2026, covering approximately 482,000 works at roughly $3,113 per book. It is the largest publicly reported copyright recovery in US history.

The music industry is using the same playbook. Universal Music Publishing Group, Concord Music Group, and ABKCO filed a $3.1 billion lawsuit against Anthropic in January 2026 over training on song lyrics. BMG filed a separate suit in March. Authors have sued Meta. Artists have sued Google. According to the Copyright Alliance, more than 70 AI copyright infringement lawsuits had been filed as of early 2026.

4. Your conversations
This one surprises people. A Stanford study found that all six leading US AI companies, including Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI, have harvested user conversations to train their models, with varying degrees of transparency about opt-outs. The policies differ. Anthropic's default on paid plans is not to train on your conversations unless you opt in. Others are less clear. The key word to look for in any AI tool's privacy policy is "training". If it's in there, your conversations may be fuel.

It is also worth knowing that what a company promises about your conversations can be overridden by a court order. As part of the New York Times' ongoing lawsuit against OpenAI, a federal court ordered OpenAI to produce 20 million ChatGPT conversation logs in January 2026. Zero Data Retention enterprise customers were exempt. The lesson: privacy policies describe defaults. Lawsuits can override them.

THE PRACTICAL PICTURE

The data problem isn't going away. It's getting harder. Researchers have identified that high-quality public web data is already heavily mined. That's part of why AI companies are racing to expand into markets like India, where hundreds of millions of people are coming online for the first time and generating fresh human-produced data at scale. The next wave of training data may come from populations who don't yet know they're part of the equation.

Safe Harbor: Three Things You Can Do This Week

  • Find the training opt-out on the AI tool you use most. This is different from general privacy settings. Look specifically for "use my conversations to train models" or "improve our services." It's often two or three menus deep, and it's often on by default.
  • Use your robots.txt file if you run a website. If you publish content online and don't want it used to train AI models, you can add crawler blocks for GPTBot, ClaudeBot, Google-Extended, and others. It's not a guarantee, but it signals your preference.
  • Write down three categories of information you will never put into an AI tool. Examples: financial account details, medical specifics, anything involving someone else's confidential information. Having the list written down makes it easier to defend in the moment.

Next week: AI's hidden vocabulary. We used words like training and tokens throughout previous issues. Next week we slow down and actually define them.

Read more