Where training data comes from

A model didn't learn Ruby from nowhere

When an agentic tool from AI tools: introduction (Cursor & Antigravity) suggests scope :published, -> { where.not(published_at: nil) }, it produces that pattern because something like it appeared many times in its training material. Across the industry, that material is reported to include public web pages, open-source code repositories, books, forums and licensed data deals. That's where a model's sense of "what Rails code looks like" comes from. None of it is magic, and none of it escapes the copyright and licensing questions that already applied to the source material.

Public code isn't public domain

A repository anyone can read on a code-hosting site isn't automatically free of copyright. That distinction is easy to miss. Open-source code almost always ships under a license, such as MIT, GPL or Apache, and each has terms. Some require attribution; some require derivative works to carry the same license. "Publicly visible" and "free to reuse however you like" are two different claims, and training data drawn from public repositories inherits whatever licensing questions those repositories carried.

What companies do and don't publish

Companies generally don't publish a complete list of the code and text that trained a given model. They describe broad categories, like "public web data," "licensed partnerships" or "code repositories," but an itemized list isn't the norm. That gap is why lesson 2 calls the licensing question contested rather than settled: nobody outside the company can fully audit the answer.

Why this opens the course

Every claim this course makes about licensing, privacy and company policy traces back to one fact. A model's output is shaped by material it learned from, and most of that material carried a license or ownership claim first. Lesson 2 asks what that means for code. Lessons 3 to 6 turn to a related but separate question: what happens to new code and data you type into these tools.

Key takeaways

• A model's suggestions reflect patterns from its training material, reported to include public web text, open-source code, books and licensed data.
• A public code repository isn't automatically free of copyright; most carry a license (MIT, GPL, Apache and others) with terms.
• Companies generally don't publish an itemized account of what trained a specific model.
• That gap is why the next lesson treats code licensing as an open question.