AI and Copyright: The Lawsuits, the Rulings, and What They Mean for Creators

No court has settled whether AI companies may train on copyrighted work without permission. The rulings so far point in different directions. The docket is growing anyway: one tracker counted 161 US copyright suits against AI companies as of November 4, 2026. This guide maps the decisions that matter, the cases that will decide the rest, and the steps creators can take now.

Quick Takeaways

  • Training is not yet a settled win for either side. Two 2025 district rulings favored AI developers on training. A 2026 appellate ruling went against a non-generative AI tool.
  • Piracy is the bigger legal risk. Judge Alsup held that creating a permanent library from pirated copies was not fair use. That produced the $1.5B Anthropic settlement.
  • Outputs are a separate fight. Memorized lyrics, recognizable characters and regurgitated text raise different claims than training does.
  • Pure AI output stays uncopyrightable. The Supreme Court denied certiorari in Thaler v. Perlmutter on March 2, 2026.
Question Where it stands (Nov 8, 2026)
Is training on books fair use? Yes in two N.D. Cal. rulings, on narrow records. No appellate ruling yet.
Is training on a competitor’s data fair use? No, per the Third Circuit in Ross.
Does pirating training data create liability? Yes. Anthropic paid $1.5B.
Can AI output infringe? Yes, per a Munich court (GEMA). US courts have not ruled.
Is AI-only output copyrightable? No.

The Core Legal Question: Fair Use

US copyright cases over training turn on the four fair use factors. The two that decide most disputes are:

  1. Purpose and character: Is the use transformative?
  2. Market effect: Does the copying harm the market for the original?

Courts have weighed these differently depending on the facts. The facts that matter most are what the AI does, how the data was obtained and whether the output substitutes for the original work.

The Rulings That Shape Everything

Bartz v. Anthropic: Training Passed, Piracy Did Not

In 2025, Judge William Alsup split the case in two. He held that AI companies can use legally purchased books for training, not illegal ones. The claims over downloaded pirated copies headed to a jury.

Anthropic settled before trial. On July 20, 2026, Judge Araceli Martínez-Olguín granted final approval of the $1.5 billion settlement. The key terms:

  • Scale: The Works List contains 482,460 works.
  • Payout: The estimate is about $3,000 per work, before fees and costs.
  • Fees: The court cut class counsel’s fees to about 6.8 percent of the fund, roughly $101.56 million.
  • Scope: Class members release only claims about past acquisition and copying through August 25, 2025. Output claims and future conduct are not released.
  • Data destruction: Anthropic must destroy the original files torrented from LibGen and Pirate Library Mirror.

Kadrey v. Meta: A Win on a Thin Record

Judge Chhabria also found fair use in 2025. The ruling was narrow. It covered Meta’s copying of the thirteen named plaintiffs’ books on that record. Chhabria signaled that plaintiffs with better evidence of market dilution could win. The Copyright Office’s pre-publication report, by comparison, described fair use outcomes as highly fact-dependent.

Plaintiffs have adjusted. The firm Stris & Maher has filed five new suits against Meta and three against SpaceX Superintelligence. It argues that scanning lawfully owned books for training is not fair use. That directly challenges Alsup’s reasoning.

Thomson Reuters v. ROSS: The First Appellate Word

On September 29, 2026, the Third Circuit affirmed summary judgment for Thomson Reuters. It was the first federal appellate decision on fair use in AI training. The panel found ROSS’s use “minimally transformative, at best.” The opinion also carries a built-in limit: the case does not concern a large language model or other generative AI tool.

Comparing the Major Decisions

Case Court Outcome Why it matters
Bartz v. Anthropic N.D. Cal. Training fair use; pirated library not Sourcing drives liability
Kadrey v. Meta N.D. Cal. Fair use for 13 plaintiffs Market dilution is an open theory
Thomson Reuters v. ROSS 3d Cir. Not fair use Competing substitutes lose
GEMA v. OpenAI Munich OpenAI liable Memorization counts as reproduction
Getty v. Stability AI UK High Court Getty lost most claims Model weights are not copies, per that court

The Case That May Decide the Most: In re OpenAI

The US litigation to watch is the OpenAI multidistrict litigation in the Southern District of New York. It consolidates more than a dozen major suits, including the Authors Guild’s and the New York Times’. The schedule:

  • All sides filed summary judgment motions on September 4, 2026.
  • Oppositions were filed October 23.
  • Replies are due November 20, 2026.

The case has drawn unusual outside attention. More than 50 amicus briefs have been filed. The federal government has also taken a side. The Trump administration urged the court to rule that training on publishers’ works is fair use. The Justice Department argued that a ruling against OpenAI would hamper LLM competition, since only the largest companies could afford licensing fees.

Procedural rulings are shaping the case too. Judge Stein refused on August 6 to let the Times add a theory that Microsoft actively encouraged infringement. He said the request came too late. That ruling followed the Times dropping its claim that OpenAI contributed to ChatGPT users’ infringement after a Supreme Court decision raised the bar for platform liability.

Expect a fair use ruling after the briefing closes. Whichever way it goes, an appeal to the Second Circuit is likely.

Outputs, Characters and Lyrics: The Second Front

Training is only one claim. Several cases target what models produce.

  • Lyrics (Germany). The Munich court found OpenAI infringed by training on lyrics and by reproducing them on request. OpenAI is appealing. The ruling conflicts with the UK on a technical point. Munich treated lyrics as contained in the model’s parameters, which contradicts the UK ruling on model weights.
  • Characters (US). Disney, Universal and Warner Bros. accuse Midjourney of enabling mass generation of characters such as Batman and Superman. The case is in discovery. Midjourney is pressing the studios to disclose more about their own AI use.
  • Visual artists. The artists’ class action against Stability AI and Midjourney has slipped. As of August 2026, its listed trial date was April 5, 2027, subject to change.
  • Music. The industry is shifting from litigation to licensing. Warner settled with both Suno and Udio, and Universal settled with Udio. The NMPA struck an industry-wide AI licensing deal with Udio. Meanwhile, UMG and Sony’s settlement talks with Suno hit an impasse in April.

What the Rulings Mean for Creators

Ownership: Human Authorship Is Required

By declining review, the Supreme Court left intact the D.C. Circuit’s holding that copyright requires human authorship. Assisted work is different from autonomous work. The government noted that the Copyright Office has registered hundreds of works incorporating AI-generated material where sufficient human authorship was shown. The boundary remains unsettled. Allen v. Perlmutter tests whether an image made with more than 600 Midjourney prompts qualifies.

Claims: Check Whether You Are in a Class

If your books are on a shadow-library list, you may have a claim. For authors, registration timing matters too. The Anthropic class required Copyright Office registration within five years of first publication.

Exposure: Your Tools’ Terms Are Not Insurance

Creators who use generative tools carry their own risk. Read the indemnification terms before relying on a tool for client work. Many providers do not cover infringement claims.

Practical Workflows

Workflow 1: Block AI Crawlers

A robots.txt entry is a request, not an enforcement mechanism, but compliant crawlers honor it. Add these lines to the site root:

# Block common AI training crawlers (compliant bots only)
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: Google-Extended
Disallow: /

User-agent: CCBot
Disallow: /

User-agent: PerplexityBot
Disallow: /

Document the date you added it. In EU disputes, opt-outs affect whether the text-and-data-mining exception applies.

Workflow 2: Build a Provenance Log

Keep a dated record of every work you register or publish. Use this template:

Field Example
Work title and ISBN/ID Title, ISBN
First publication date 2024-03-12
Copyright registration number and date TX-XXXX, 2024-04-01
AI tools used and what they did Drafting aid only; all text rewritten
Human contribution summary Structure, prose, edits

The Copyright Office asks applicants to disclose AI involvement, and a log makes that disclosure accurate and fast.

Workflow 3: Register Early

Statutory damages depend on timely registration. Register new works promptly instead of waiting for a dispute.

This is general information, not legal advice. Talk to an IP attorney about your specific works.

What to Watch Next

Date Event
Nov 20, 2026 Reply briefs due in In re OpenAI
Nov 20, 2026 Case management conference, In re Salesforce
Nov 23, 2026 Replies due on summary judgment, Nazemian v. NVIDIA
Feb 17, 2027 Reported summary judgment hearing, Andersen v. Stability AI
Pending Midjourney’s request to broaden discovery from the studios
Pending Final Copyright Office training report

Some dates have already moved once. Most of these cases have extended their schedules at least once. Treat every date as provisional.

FAQ

Is training AI on copyrighted work legal?

It depends on the facts, and no appellate court has ruled on generative AI training. Two district courts found fair use for LLM training on books in 2025. The Third Circuit rejected fair use for a non-generative tool built to compete with Westlaw. Acquiring data through piracy is a separate risk.

How much did Anthropic pay authors?

Anthropic agreed to a $1.5 billion class settlement, which received final approval on July 20, 2026. The estimated payment is about $3,000 per covered work, before fees and costs. It covers only past copying of pirated books through August 25, 2025.

Can I copyright something made with AI?

Not if AI alone made it. The Supreme Court left in place the rule that copyright requires human authorship. Works with meaningful human creative contribution can be registered, but you must disclose the AI involvement, and the line for “enough” human input is still being tested.

Can I stop AI companies from using my work?

You can signal an opt-out with robots.txt and metadata, and you can sue or join a class action where your work was copied. A robots.txt file cannot stop non-compliant scrapers. Past copying by crawlers that ignored it is a matter for the courts.

Hot this week

Build Your First Spam Email Classifier

Build a spam email classifier in Python with TF-IDF and Naive Bayes. Copy the code, hit 98%+ accuracy, and ship your first ML model today.

Best AI Music and Voice Generators in 2026

Compare Suno v6, ElevenLabs Music, Udio, Eleven v4, Cartesia and Fish Audio on price, rights and latency. Pick your tool and start creating today.

Train/Test Splits Done Right: Best Practices and Common Mistakes

Avoid data leakage with proper train/test splits. Compare random, stratified, group and time-series methods with Python code. Fix your pipeline today.

AI Engineer vs ML Engineer vs Data Scientist: What’s the Difference?

AI engineer vs ML engineer vs data scientist: compare skills, tools, and roles side by side. Find your best-fit career path and start learning today.

Chain-of-Thought, Few-Shot, and Role Prompts: When Each Technique Helps

Chain-of-thought, few-shot, and role prompts each fix different failures. Learn when each works, when it backfires, and copy the templates.

Topics

Build Your First Spam Email Classifier

Build a spam email classifier in Python with TF-IDF and Naive Bayes. Copy the code, hit 98%+ accuracy, and ship your first ML model today.

Best AI Music and Voice Generators in 2026

Compare Suno v6, ElevenLabs Music, Udio, Eleven v4, Cartesia and Fish Audio on price, rights and latency. Pick your tool and start creating today.

Train/Test Splits Done Right: Best Practices and Common Mistakes

Avoid data leakage with proper train/test splits. Compare random, stratified, group and time-series methods with Python code. Fix your pipeline today.

AI Engineer vs ML Engineer vs Data Scientist: What’s the Difference?

AI engineer vs ML engineer vs data scientist: compare skills, tools, and roles side by side. Find your best-fit career path and start learning today.

Chain-of-Thought, Few-Shot, and Role Prompts: When Each Technique Helps

Chain-of-thought, few-shot, and role prompts each fix different failures. Learn when each works, when it backfires, and copy the templates.

Outliers: How to Detect and Handle Them Without Losing Signal

Learn outlier detection with Z-score, IQR, Isolation Forest, and LOF. Copy the Python code and handle outliers without losing real signal.

Hotend and Extruder Basics: How Filament Becomes Plastic Layers

Learn how a hotend and extruder turn filament into layers. Compare direct vs Bowden, nozzles, and temps, then tune yours. Read the full guide.

Multiboard vs Pegboard vs Gridfinity: Which Wall Storage System Should You Print?

Multiboard, pegboard or Gridfinity? Compare load, filament, print time and cost, then pick the right wall storage system to print. Read the guide.

Related Articles

Popular Categories