How an AI copyright lawsuit works, step by step

AI copyright cases are still copyright cases. They follow the same procedural arc as any other federal copyright suit, with three places where they tend to break differently.

Stage 1: registration and pre-suit

A copyright suit in the United States cannot proceed past the pleadings without a registered copyright. For news publishers and individual authors, this is a paperwork step. For data — datasets, scraped corpora, output works — registration is harder and a real strategic constraint. Statutory damages are only available for works registered before infringement begins, which is one reason plaintiffs structure complaints around specific timestamps.

Stage 2: complaint and forum

Most AI copyright cases are filed in one of four districts: the Southern District of New York (publishers and the Authors Guild MDL), the Northern District of California (foundation-model labs), the Central District of California (image and music), and Delaware (corporate-residence-based filings). Forum matters because each court has developed distinct preliminary postures on AI cases.

Stage 3: motion to dismiss

The first decisional choke point. AI defendants typically argue: (1) the plaintiff has not pled substantial similarity between training inputs and outputs; (2) the plaintiff has not registered the specific works; (3) the conduct is fair use as a matter of law. The first ground is where most early-2024 dismissals happened. By 2025, most AI cases survive 12(b)(6).

Stage 4: class certification (in class actions)

The second choke point. Defendants argue that whether each plaintiff's specific work was in the training data is an individualized question that defeats commonality. Plaintiffs argue that whether the corpus included unlicensed copies is a common question. Bartz v. Anthropic certified; Andersen v. Stability AI partially certified after substantial narrowing.

Stage 5: summary judgment

The fair-use determination, when it comes, comes here. Bartz's bifurcated summary-judgment ruling — training potentially transformative; downloading pirated copies not — is the template most courts now follow. Fair use is a question of law applied to undisputed facts, which is why most AI courts treat it at SJ rather than letting it go to a jury.

Stage 6: settlement or trial

Trials are rare. Once SJ resolves the headline doctrinal questions, cases settle on damages. The Bartz $1.5 billion settlement is the canonical example: the doctrinal question (lawful acquisition) was resolved at SJ, then the parties priced the willful-damages exposure and settled.

What changes vs. a normal copyright case

Three things. First, discovery into training corpora is unprecedented in scope — courts have ordered production of crawl logs, dataset manifests, and model checkpoints. Second, expert evidence on substantial similarity is harder than in traditional copyright (you cannot easily play a song and ask whether it sounds like a record). Third, statutory damages math swings outcomes more aggressively because the corpus sizes are enormous.

Frequently asked questions

How long does an AI copyright lawsuit take?

From filing to disposition typically runs 18-30 months for cases that settle, 36-48 months for cases that reach summary judgment, and longer if there is an appeal. Bartz v. Anthropic settled roughly two years after filing.

What does it cost to file an AI copyright lawsuit?

Federal filing fees are nominal. Litigation costs vary enormously — a single-plaintiff defamation case may cost low six figures; a books-class action with extensive discovery into training corpora can cost tens of millions on each side.

Can I sue an AI company without registered copyrights?

You can file, but you cannot recover statutory damages or attorney's fees without timely registration. Actual damages and injunctive relief remain available, but the economics rarely work for individual plaintiffs without registration.

What is the difference between actual damages and statutory damages in AI cases?

Actual damages require proof of lost licensing revenue or profit attributable to the infringement — hard to prove for training data. Statutory damages are set by 17 USC § 504(c) and range from $750 to $150,000 per work for willful infringement, regardless of actual harm. The willful-statutory math is what drives most AI settlements.