140 - Generative AI and Copyright, with Chris Callison-Burch

Training Data

Chris Callison-Burch, a professor at the University of Pennsylvania, highlights the complexities of AI training data and copyright issues. He explains that generative AI models, like OpenAI's ChatGPT, are trained on vast datasets, often containing copyrighted material, which raises legal concerns. Chris argues for the preservation of fair use in AI training, likening the process to human learning and emphasizing its necessity for innovation 1. He acknowledges the potential for AI outputs to infringe on copyright, citing examples where models reproduce copyrighted content verbatim 2.

Generative AI is trained on huge amounts of data. Large language models are now trained on roughly 1 trillion words.

--- Chris Callison-Burch

Chris stresses the need for legislation that balances innovation with copyright protection, advocating for technical solutions to minimize copyright violations in AI outputs 2.

   

Model Capabilities

The capabilities of AI models have advanced significantly, as Chris Callison-Burch notes, with generative AI reaching a transformative stage. He shares his initial shock at the capabilities of models like ChatGPT, which can perform complex tasks such as language translation and document summarization 1. Despite initial fears about the impact on academic research, Chris remains optimistic about AI's potential to enhance productivity and creativity 1.

This is a truly transformative technology that will shape many aspects of our lives. I hope that it is for the better.

--- Chris Callison-Burch

He emphasizes the importance of understanding AI's technical aspects and advocates for legislation that supports innovation while addressing potential risks 3.