In August 2025, Anthropic agreed to pay $1.5 billion to settle a class-action lawsuit from authors who accused the company of training Claude on pirated copies of their books — the largest publicly disclosed copyright settlement in AI history. A year later, the underlying legal question that produced that settlement is still unresolved: is it actually legal to train a large language model on copyrighted books in the first place?
According to TechCrunch, the practice is widespread across the industry, but U.S. case law on it remains inconsistent, with separate lawsuits against OpenAI, Meta, and Anthropic producing different rulings on overlapping facts. That inconsistency is the actual headline here — not that training on books happens, which everyone already knew, but that the courts can't yet agree on when it's allowed.
For anyone building products on top of a foundation model, that's not an academic detail. It's a standing liability question with no fixed answer.
Why the courts disagree
The split runs through a single legal concept: fair use. U.S. copyright law lets courts weigh whether a use is "transformative" enough to bypass a rights holder's permission, and judges have reached different conclusions about where AI training falls on that spectrum. Some rulings have treated the transformation from "book text" to "statistical model weights" as sufficiently different from the original to qualify as fair use. Others have drawn a harder line around how the training data was obtained, treating the training itself and the sourcing of the data as legally separate questions.
That second distinction is what actually sank Anthropic's position enough to force a settlement: the company had trained on legally purchased and scanned books, which courts were comfortable calling fair use, but it had also trained on books downloaded from pirate libraries — and piracy doesn't get a fair-use defense no matter how transformative the downstream model is.
Same question, three different defendants
OpenAI and Meta are fighting versions of the same fight, with authors' groups and individual writers arguing that both companies trained on shadow libraries of pirated books without a license. Meta in particular has faced scrutiny over training data sourced from bulk pirated book archives. The claims against all three companies are structurally similar — copyrighted books, ingested without a license, used to train a commercial model — but they're proceeding through different courts, different judges, and different procedural postures, which is exactly why the case law hasn't converged.
A few things are consistent across the litigation so far:
- Buying and scanning a book yourself has fared far better in court than downloading it from a pirate source, even when the end use is identical.
- Settlements are increasingly the industry's preferred outcome over waiting for a definitive appellate ruling — Anthropic chose to pay rather than risk a broader precedent.
- No single decision so far functions as binding precedent for the whole industry, since each case turns on its own facts and its own circuit.
What this means for anyone shipping on top of these models
If you're building a product on GPT, Claude, Llama, or any other foundation model, you don't control what that model was trained on, and right now nobody outside the labs has a definitive answer to whether that training was fully lawful. That has a few concrete consequences worth planning around:
- Vendor contracts and indemnification clauses matter more than they used to — check whether your model provider indemnifies you against training-data claims, not just output claims.
- Regulatory and legal exposure is a moving target, not a fixed cost, since a future ruling or settlement could change the terms your provider operates under with little warning.
- Model providers with murkier data-sourcing histories carry more tail risk for downstream builders, even if that risk hasn't materialized into an obligation yet.
None of this is a reason to freeze feature work. It's a reason to know which foundation model your product depends on and to keep half an eye on that model's litigation exposure the same way you'd track a critical vendor's uptime.
AiiN's takeaway
The fair-use question won't get a clean, industry-wide answer soon — it's being litigated case by case, defendant by defendant, and the outcomes so far already diverge depending on how each company sourced its training data. In our estimation, the piracy-versus-licensing distinction is likely to end up mattering more than the training method itself, since it's the one factor that has already moved a major lab to settle for well over a billion dollars rather than litigate it out. For builders, the practical move isn't to wait for legal clarity that isn't coming soon — it's to treat model-provider due diligence as part of the stack you're responsible for, right alongside uptime and pricing.