Training Data and Copyright
Large language models are trained on terabytes of text from books, websites, and repositories. Most of this content is copyrighted. The models don’t replicate exact text but learn statistical patterns. Does this violate copyright?
Legal arguments exist on both sides:
- Fair Use Argument: Training models on copyrighted content for research is transformative and doesn’t replace the original
- Copyright Infringement Argument: The entire value of the model derives from the copyrighted training data without compensation
Lawsuits are underway. New York Times sued OpenAI and Microsoft. Authors sued for similar reasons. Outcomes will establish precedent that affects how AI is developed.
Regulatory Approaches
The EU’s approach in draft AI legislation requires explicit permission for training on certain data. The UK and US are watching legal proceedings before establishing rules. This patchwork creates uncertainty—what’s legal in one jurisdiction might be illegal in another.
Practical Implications for Developers
Use openly licensed training data when possible: Creative Commons and other open licenses enable training without copyright concerns.
Document your training data: Transparency about what data was used reduces risk and enables licensing discussions if needed.
Respect opt-out requests: If creators ask their work not be used for AI training, respecting this shows good faith.
Anticipate regulation: Future rules might require attribution or licensing fees. Systems designed with this in mind adapt better.
The Creator Perspective
Artists, authors, and musicians rightfully worry that AI trained on their work diminishes demand for their services. A model generating cover art means fewer commissions for illustrators. A model generating text means fewer writing projects.
But entirely excluding human-created content makes AI less capable. The challenge is balance—enabling AI advancement while ensuring creators benefit from their contributions.
Possible Resolution Paths
Mandatory licensing: Training on copyrighted content requires licenses and compensation, similar to music licensing.
Pooled arrangements: AI companies fund pools that compensate creators, distributed based on contribution estimates.
Open licensing movements: Creators voluntarily license work for AI training in exchange for credit and recognition.
Technological solutions: Watermarking content enables tracking when it’s used for training, facilitating licensing agreements.
The resolution will likely combine legal clarity, industry standards, and technology enabling better tracking and attribution. The organizations and countries that navigate this transition thoughtfully will gain competitive advantage in AI development.