The training-data problem
Lesson 1 of 5 in AI and Intellectual Property: Training Data, Outputs, and the Litigation Wave.
Every large model was trained by copying. Web pages scraped and stored, books ingested, images downloaded, all reproduced across preprocessing pipelines and GPU clusters — and copying is exactly the act copyright law regulates. That single technical fact spawned the largest coordinated litigation wave the technology industry has seen: dozens of suits by authors, newspapers, artists, music publishers, and stock-photo agencies against essentially every major model developer.
Be precise about where copying happens, because the legal analysis differs at each point:
- Corpus assembly — works are downloaded and stored. Clear reproduction; the question is whether an exception (fair use, a TDM exception) excuses it.
- Training — works are processed into weights. Developers argue the model learns statistical patterns, not the works; plaintiffs argue the works are copied throughout training and sometimes retained.
- The model itself — do the weights "contain" the works? Mostly no in a literal sense, but memorisation is real: models can regurgitate training text and near-duplicate images, which is how the New York Times exhibited a hundred near-verbatim article excerpts.
- Outputs — a generated image or paragraph can be substantially similar to a training work, creating ordinary infringement exposure for whoever generates or distributes it.
Keep those four loci separate and every case in this module becomes legible: each lawsuit is a fight about which locus, under which national law, with which exception.
Key terms: fair use, text and data mining exception, foundation model, training, content provenance
| Case | Core claim | What it established (as of this writing) |
|---|---|---|
NYT v. OpenAI/Microsoft (S.D.N.Y., filed Dec 2023) | Training on millions of Times articles + outputs regurgitating them near-verbatim | Survived dismissal in large part (2025); the marquee test of fair use for generative training on news. Discovery battles over output logs; merits undecided — watch it |
Thomson Reuters v. Ross (D. Del., decided Feb 2025) | Training a legal-research AI on Westlaw headnotes | First US ruling to reject fair use for AI training: factors 1 and 4 favoured Thomson Reuters — Ross built a direct market substitute. Non-generative facts; on interlocutory appeal |
Bartz v. Anthropic (N.D. Cal., 2025) | Training Claude on books, some purchased, some from pirate libraries | Training on lawfully acquired books held fair use ("transformative"); retaining pirated copies was not — followed by a landmark $1.5B settlement for the pirated-library class |
Kadrey v. Meta (N.D. Cal., 2025) | Training Llama on authors’ books | Meta won summary judgment on the record presented — but the court stressed plaintiffs failed to develop the market-dilution theory, signalling training claims can win on better evidence |
Getty Images v. Stability AI (UK High Court, judgment Nov 2025) | Stable Diffusion trained on 12M Getty images; watermarks appearing in outputs | Getty dropped its main training claims mid-trial (training happened outside the UK); secondary-infringement claim failed because model weights are not an "infringing copy"; narrow trademark win on watermarks. Parallel US case continues |
Andersen v. Stability AI (N.D. Cal., ongoing) | Artists’ class action over image-model training and style imitation | Core copyright claims allowed to proceed (2024); testing whether a diffusion model is itself an infringing work |
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.