The Legal Reckoning Over AI‘s Training Data: Copyright Lawsuits Highlight Key Challenges
The recent lawsuits filed by Sarah Silverman, Christopher Golden and Richard Kadrey against tech giants OpenAI and Meta have sparked an intense debate over the use of copyrighted works in training datasets for artificial intelligence models. The authors allege that the companies‘ ChatGPT and LLaMA language models were trained on troves of their books and other writings, scraped from the internet without permission. As these high-stakes cases move forward, they are laying bare the lack of clear legal frameworks governing AI training data – and the tensions between protecting intellectual property and enabling transformative machine learning research.
The Allegations: Large Language Models Trained on Illicit Data
At issue in the lawsuits is the question of whether OpenAI and Meta violated the authors‘ copyrights by ingesting their works into the datasets used to train ChatGPT and LLaMA. The complaints allege that the companies obtained the texts from "shadow library" websites that offer unauthorized downloads of books and articles, such as Bibliotik, Library Genesis, and Z-Library.
ChatGPT and LLaMA are examples of large language models (LLMs), a type of machine learning system that has been behind many of the biggest breakthroughs in AI over the past few years. LLMs are trained on massive amounts of text data, often scraped from the internet, which they use to build complex statistical representations of language. This allows them to perform tasks like question-answering, summarization, and generation of coherent text in the style of the training data.
The scale of the datasets used to train state-of-the-art LLMs is staggering. For example, OpenAI‘s GPT-3 model, on which ChatGPT is based, was trained on a dataset of nearly 500 billion tokens (roughly equivalent to 1 trillion words of text), sourced from a filtered version of the Common Crawl web scrape, digitized books, and Wikipedia. Similarly, Meta‘s LLaMA models learned from a 1.4 trillion token dataset called TheBlob, which includes data from Common Crawl, Wikipedia, Project Gutenberg, arXiv, and dozens of other online sources.
Given this vast scope, it‘s perhaps unsurprising that copyrighted works would end up in these datasets. But the authors argue it‘s a clear violation for the companies to reproduce their writings for commercial gain without permission or compensation. As evidence, the lawsuits point to ChatGPT‘s ability to generate detailed summaries of the authors‘ books, which they say is a sign the model memorized substantial portions of their texts verbatim during training.
The Legal Landscape: Fair Use and Copyrights for AI
The key legal question in these cases is likely to be whether OpenAI and Meta‘s use of the authors‘ copyrighted material to train their AI models constitutes fair use. Under U.S. copyright law, reproducing protected works without a license is permissible for certain purposes such as criticism, teaching, research, or other transformative uses that don‘t simply substitute for the original. Courts evaluate fair use claims by weighing four main factors:
- The purpose and character of the use, including whether it is commercial or non-profit and whether it is transformative
- The nature of the copyrighted work (e.g. factual vs. creative)
- The amount and substantiality of the portion used relative to the work as a whole
- The effect of the use on the potential market for or value of the copyrighted work
Past court rulings have found that some types of machine learning training can be fair use, such as using thumbnail images to train visual search algorithms. But LLMs pose a knottier case, given the breadth of content they ingest (entire books, not just snippets) and their ability to reproduce significant chunks of that data in their outputs.
Some legal experts believe OpenAI and Meta have a strong argument that ingesting the authors‘ works to train the models is a transformative use, since the purpose is not to simply republish the texts, but to create a general-purpose AI system that can engage in open-ended language tasks. The companies may also contend that the commercial nature of the models is outweighed by their significant research and societal benefits.
However, critics argue that even if the purpose is transformative, the sheer scale of the data ingestion and the models‘ memorization of the works pushes it beyond the bounds of fair use. They say training LLMs is more akin to creating an unauthorized searchable database of copyrighted material. There are also concerns about the potential market harm to authors, if readers turn to free AI chatbots that can provide detailed book summaries instead of purchasing the original works.
Toward Sustainable AI Training Data Practices
Cases like Silverman, Golden and Kadrey‘s lawsuits make clear the urgent need for updated legal frameworks and industry standards around the use of copyrighted material in AI training datasets. While some companies have adopted their own policies (Google, for instance, forbids the use of copyrighted works without permission), there is a lack of clear guidelines across the field. This has left individual AI researchers and startups uncertain about their legal exposure and raised concerns about a "data laundering" ecosystem of unauthorized web scraping.
Legal experts have proposed several potential solutions, including:
- Updating fair use doctrine to create a clearer safe harbor for transformative AI training, with guardrails to protect authors‘ rights
- Establishing a compulsory licensing scheme, where AI companies could use copyrighted works for training in exchange for reasonable royalty payments to authors/publishers
- Encouraging voluntary licensing agreements between AI developers and content owners to provide legal certainty and ensure fair compensation
The European Union is considering some of these approaches in its draft AI Act, which would create new obligations for companies to document the intellectual property rights associated with their training datasets and restrict the use of illegal content.
On the technical front, there are also efforts underway to develop better methods for filtering out copyrighted material from web-scraped datasets and documenting dataset provenance. Groups like the Partnership on AI and the AI Now Institute have called for increased transparency around AI training data practices.
Some experts even envision training AI systems themselves to honor copyrights, by fine-tuning LLMs on legally-licensed data and instructing them to avoid reproducing protected works. But this remains an unsolved challenge, given the models‘ ability to internalize and imitate the data they are exposed to during training.
Balancing Innovation and Authors‘ Rights
As the lawsuits against OpenAI and Meta wind through the courts, they will have significant implications not only for the companies involved, but for the future direction of AI development. At stake is the balance between protecting the legitimate rights of authors and fostering the kind of groundbreaking machine learning research that many see as key to advancing the field.
Few would dispute that writers like Silverman, Golden and Kadrey deserve to be compensated for their creative works and have a say in how they are used. But at the same time, training cutting-edge AI systems requires large, diverse datasets that inevitably bump up against copyright limitations. Overly restrictive rules around training data could risk slowing the pace of innovation and concentrating power in the hands of a few large tech companies with the resources to license datasets.
Threading that needle will require good-faith collaboration between AI developers, content creators, legal experts and policymakers. We need updated legal frameworks and industry practices that allow for the responsible development of AI training datasets while ensuring authors are fairly treated. This could include steps like:
- Requiring companies to obtain licenses for a substantial portion of their training data and compensate authors for usage
- Setting clear thresholds for how much of a copyrighted work can be ingested into a dataset under fair use
- Mandating detailed documentation of dataset contents and provenance to improve transparency
- Developing technical tools to help identify and filter out copyrighted material
- Creating mechanisms for authors to opt out of having their works included in training data
Importantly, any new policies will need to be crafted with input from a wide range of stakeholders, including smaller AI startups, independent researchers, and authors and artists from diverse backgrounds. We must avoid a future in which only tech giants have the legal resources to train frontier AI models.
The Future of Generative AI and Copyright
Ultimately, the Silverman, Golden and Kadrey lawsuits are an important inflection point in the rapidly-evolving relationship between copyright law and artificial intelligence. As generative AI systems become more sophisticated and widely deployed, clashes over training data and intellectual property are likely to intensify.
On one side are creators concerned about protecting their livelihoods and artistic integrity in a world of AI chatbots and automated content generation. On the other are technology companies and researchers eager to push the boundaries of what‘s possible with machine learning, but who risk overstepping ethical lines in the process.
Striking the right balance will be a difficult but necessary challenge. If we can establish clear legal frameworks and responsible industry practices, AI systems trained on properly-licensed data could become a powerful tool for augmenting and inspiring human creativity – while still respecting the rights of the authors and artists who provide the raw materials.
Imagine literary chatbots that help writers hone their prose, visual generators that remix images in novel styles, and music AIs that compose new melodies – all while ensuring the original creators are recognized and compensated. That‘s the kind of generative AI future worth striving for.
But realizing it will require ongoing good-faith dialogue between the AI community, the creative industries, and the public. The Silverman, Golden and Kadrey lawsuits are an important test case, but they are only the beginning of a much larger societal conversation about the role of copyright in the age of ubiquitous AI.
As we wrestle with these issues, one thing is clear: in a world of machine-generated everything, the legal and ethical foundations of our creative culture are shifting beneath our feet. It‘s up to all of us – creators, technologists, policymakers, and citizens alike – to ensure those foundations remain strong. The future of both human and artificial creativity may depend on it.
