Skip to content
Submit incident
AII-1001-99273
Well documentedThree or more independent publishers report the core facts.

Llama 3.1 reproduces copyrighted books including Harry Potter

Researchers found that Meta's Llama 3.1 AI model can reproduce large portions of copyrighted books verbatim, including 42% of the first Harry Potter novel, raising concerns about potential copyright infringement liability for the company.

16 June 2025 Β· Meta Β· Intellectual property and appropriation
16 June 2025
AII-1001-992733 sources ↓
Quick facts
Category
Intellectual property and appropriation
Country
United States
Source
Ars Technica
Date
16 June 2025
Drafted by
Raidu Incident Detection

What happened

Researchers from Stanford, Cornell, and West Virginia University discovered that Meta's Llama 3.1 model can reproduce large portions of copyrighted books with high accuracy. The study found that Llama 3.1 has memorized around 42% of the first Harry Potter book and can reproduce 50-word sections correctly nearly half the time.

Llama 3.1 was the most prone among five major AI models analyzed to output copyrighted content, including exact excerpts from The Hobbit and 1984. The model was trained on the Books3 dataset, which contains approximately 200,000 copyrighted publications that Meta engineers downloaded using a torrent.

Reported impact

Legal experts warn that the model's ability to reproduce verbatim copyrighted material creates significant legal liability for Meta. The findings raise concerns that Llama 3.1 functions more like a repository of copyrighted works than a system producing transformative outputs based on learned patterns.

Index analysis

By comparison, Llama 1 65B, released in February 2023, had memorized only 4.4% of the same Harry Potter book, highlighting a significant increase in verbatim retention in the newer model. Experts suspect the increased memorization could result from repeated exposure to the same texts during training, possibly sourced from fan sites, academic analyses, or online reviews.

Governance controls

A copyright lawsuit against Meta by authors including Sarah Silverman produced court messages showing that Meta engineers expressed concern while downloading Books3 from a torrent on a corporate laptop. Legal experts estimate that if just 3% of the Books3 dataset were found to be infringing, Meta could owe nearly $1 billion in statutory damages.

Classification

Category
Intellectual property and appropriation
Industry
Media, social platforms and marketing
Country
United States
Harm status
Potential harm
Intent
Unintentional
Timing
Post-deployment
Harm types
Economic or property
Harmed
Business
Failure modes
Training data misuse
AI systems
Llama 3.1

Sources and evidence

This record was researched and written by the Index, based on the following original reporting.

Ars Technica
Study: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book
20 June 2025
The Hans India
Meta's Llama 3.1 AI model reproduces copyrighted book text, raising legal alarms
16 June 2025
Futurism
Lawyers Just Discovered Something About Meta's AI That Could Cost Zuckerberg Untold Billions of Dollars
17 June 2025
Jul 2026Nigeria opens investigation into tech giants over media content exploitationUnder investigationIntellectual property and appropriationYahoo actualitΓ©sJul 2026Delhi High Court Restrains AI and Deepfake Misuse of Ravi Kishan's IdentityConfirmedIntellectual property and appropriationThe Times of Indiathehindu.comAsian News International (ANI)+1Jun 2026Bombay High Court allows Preity Zinta to sue over AI-generated deepfakesConfirmedIntellectual property and appropriationThe Times of IndiaThe Indian ExpressTimesNow+5Jun 2026Judge allows copyright lawsuit over Meta's use of pirated adult films for AI trainingUnder investigationIntellectual property and appropriationMashableInternational Business Times UKMay 2026Journalists and narrators sue tech giants for using voiceprints to train AIAllegedIntellectual property and appropriationnews.bloomberglaw.comCourt House News ServiceBiometric Update+4