This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
Multiple AI companies, including Meta, OpenAI, and Bloomberg, are facing lawsuits for allegedly training their large language models on the 'Books3' dataset, which contains approximately 191,000 pirated books. Authors claim this constitutes copyright infringement, as their works were used without consent or compensation. The legal proceedings highlight ongoing tensions between the tech industry's 'fair use' arguments and the intellectual property rights of creators.
Meta and Bloomberg allegedly used Books3, a dataset containing 191,000 pirated books, to train their AI models, including LLaMA and BloombergGPT, without author consent. Lawsuits from authors such as Sarah Silverman and Michael Chabon claim this constitutes copyright infringement. Books3 includes works from major publishers like Penguin Random House and HarperCollins. Meta argues its AI outputs are not "substantially similar" to the original books, but legal challenges continue.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems capable of creating economic or cultural value, including through reproduction of human innovation or creativity (e.g., art, music, writing, coding, invention), destabilizing economic and social systems that rely on human effort. The ubiquity of AI-generated content may lead to reduced appreciation for human skills, disruption of creative and knowledge-based industries, and homogenization of cultural experiences.
Also classified as
The datasets and training processes used by AI companies are highly opaque, making it difficult for creators to verify if their intellectual property was exploited.
The unauthorized training of AI models on authors' books threatens to devalue human creative effort and disrupt the economic viability of the writing profession by enabling cheap AI-generated imitations.
Human
Due to a decision or action made by humans
Intentional
Due to an expected outcome from pursuing a goal
Pre-deployment
Occurring before the AI is deployed
The decision to scrape, compile, and train AI models on copyrighted and pirated books was made intentionally by humans during the pre-deployment training phase.
Added categories peak at 2: Minor (Psychological Harm, Indirect)
Multiple AI developers face lawsuits for training LLMs on the pirated Books3 dataset. While highlighting critical tensions between IP rights and national technological competitiveness in AI, the incident remains a civil legal dispute with minor immediate national security implications.
The AI systems involved, such as ChatGPT and LLaMA, operate as assistants that generate text, summarize books, and answer prompts based on human input and oversight.
Limited Risk: The systems described are general-purpose generative AI models and chatbots (like ChatGPT and LLaMA) which fall under transparency obligations for limited risk systems.
The models are primarily designed to function as conversational chatbots and writing assistants that can generate text, replicate styles, and summarize content.
The reports provide detailed, consistent accounts of the lawsuits, the specific dataset (Books3), the companies involved, and the legal arguments. The facts regarding the inclusion of pirated books are well-documented.
The incident could be viewed purely as a legal/copyright dispute rather than an AI safety incident, but the systemic implications for creative industries align it with AI safety taxonomies.
Technology
Opaque Training Datasets
The vast size and opacity of training data make tracking copyright hard.
Accurate Mimicry Capabilities
Models can generate cheap imitations and summaries of copyrighted books.
Lack of Attribution Tracking
Models generate content without reproducing copyright management information.
Data Inputs
Use of Shadow Libraries
Data sourced from illegal torrent sites like Bibliotik and Z-Library.
Unlicensed Book Datasets
The Books3 dataset contains 170,000 copyrighted books used without consent.
Lack of Input Verification
No robust process to filter out pirated materials before training models.
Human Factors
Open Source Permissiveness
Tech culture views restrictive copyright licenses as unnecessary barriers.
Desire to Prevent Monopolies
Developers release pirated datasets to stop corporate AI monopolies.
Lack of Creator Consent
AI developers assumed fair use without seeking permission from authors.
Process and Methods
Secretive Training Processes
AI companies train models in secret, hiding the specific texts used.
Flawed Data Conversion
Converting ebooks to plain text deleted copyright management information.
Lack of Opt Out Mechanisms
No processes existed for authors to discover or remove their books.
Management
Growth Prioritized Over Ethics
Companies prioritized building advanced models over copyright compliance.
Inadequate Risk Assessment
Management failed to address the legal risks of using shadow libraries.
Exploitation of Pirated Data
Management approved using unauthorized datasets to train proprietary models.
Regulatory Environment
Unsettled Fair Use Law
Unclear legal frameworks regarding training AI on copyrighted works.
Lack of Transparency Laws
No current laws force AI companies to disclose their training data sources.
Inadequate Copyright Protections
Existing laws fail to address mass digital piracy for machine learning.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
