Copyright and AI Training Data

The most explosive legal question of the 2020s is simple: Is it legal to train an AI on copyrighted work without permission?

On one side, tech giants argue that reading the internet is a fundamental right of learning. On the other side, artists, writers, and publishers argue that this is the largest intellectual property theft in history.

The courts are currently deciding who owns the “atoms” of creativity.

How AI Training Works

To build a model like GPT-4 or Midjourney, companies scrape billions of text snippets and images from the open web. This includes books, news articles, Reddit posts, and deviantArt portfolios.

  • The AI doesn’t “store” copies of these works (mostly).
  • It breaks them down into statistical patterns. It learns how Stephen King writes horror, or how Greg Rutkowski paints dragons, without necessarily keeping the exact file on a hard drive.

The Core Arguments

The Tech Argument: “Fair Use”

AI companies rely on the Fair Use doctrine (specifically in US law). They argue that AI training is transformative.

  1. Intermediate Copying: They admit they copy the files to train, but the output is new.
  2. Learning, Not Copying: They compare AI to a human student. If a human reads every Harry Potter book and writes a wizard story, that’s inspiration, not theft. Why is it different for a machine?
  3. Non-Competing: They argue that a general AI model doesn’t directly compete with the specific book it read.

The Creator Argument: “Unfair Competition”

Artists and authors argue that Fair Use does not apply because the AI does compete with them.

  1. Replacement: If an AI can generate a “Stephen King-style novel” or a “Disney-style illustration” in seconds for free, it destroys the market for the original creator.
  2. Scale: A human can read 1,000 books in a lifetime. An AI reads 10 million in a week. The scale changes the nature of the act.
  3. Data Laundering: They argue that tech companies are profiting from labor they didn’t pay for.

Key Lawsuits and Precedents

The New York Times vs. OpenAI (2023)

The NYT sued OpenAI, showing evidence that ChatGPT could regurgitate entire paragraphs of paywalled NYT articles verbatim. This “memorization” weakens the Fair Use defense, as it looks less like learning and more like unauthorized distribution.

The Artists vs. Midjourney/Stability AI

A class-action lawsuit by visual artists claims that image generators are merely “collage tools” that mishmash copyrighted art. They discovered that some AI images even reproduced the garbled signatures of the original artists, proving the training data was scraped directly from their portfolios.

The Future of Data Licensing

While the courts move slowly, the market is moving fast. We are seeing a shift from “scrape everything” to “pay for everything.”

  • Licensing Deals: OpenAI has signed multi-million dollar deals with Axel Springer, the Associated Press, and Reddit to legally access their data.
  • Opt-Out Mechanisms: Standards like robots.txt or “Do Not Train” tags are being developed, allowing creators to block their sites from AI scrapers.
  • Model Collapse: As the web fills with AI-generated trash, high-quality human data is becoming a scarce, premium commodity.

Conclusion

The era of the “wild west” internet scraping is ending. The future of AI will likely involve a split: “Clean” models trained on licensed, ethical data (for corporate use), and “Wild” models trained on the whole internet (living in the legal grey zones of open source).