Copyright & the Training-Data Wars

Whether training on copyrighted works is infringement turns on fair use's four factors, and the argument runs through precedent: Authors Guild v. Google held that scanning books to build a search index was transformative — the copy was an intermediate step to a non-expressive use.

The AI industry's argument extends this: training copies works to extract patterns (non-expressive learning), not to serve the works themselves. Plaintiffs counter on two flanks — market harm (models can generate content that substitutes for the originals) and memorization (models can regurgitate training text, which looks less like 'learning' and more like stored copies being re-served). The doctrine is genuinely contested and the courts are splitting the question into pieces rather than ruling on 'AI training' wholesale.