Copyright and the limits of AI
The question over machine learning tools (commonly known as AI) and their use of intellectual property is a key test of our time. That is why recently in Parliament, I spoke in a debate around machine learning and copyright. This follows contact from several concerned constituents whose livelihoods depend on them being able to protect their copyrighted works. And as I have continued to conduct my own research, I too am concerned that intellectual property must be protected in an age of machine learning.
Large Language Models, a form of machine learning such as ChatGPT, have already used pirated material and copyrighted material without consent to train their models. These are massive companies using other people’s work without explicit permissions of the original content owner or recompense for their work. This impacts the stability of artists, writers, actors and musicians and the control they have over their creative work. I believe that people should keep ownership of the work they create and it’s a well-established right that people retain copyright with limited exceptions for education or critique. We have clear copyright laws and collective licensing schemes, yet these have been ridden roughshod over by machine learning developers.
Protecting intellectual property is not only important as a principle but also it is also important economically. The UK’s creative industries are a growth area contributing £124.6bn in 2022 alone. If creatives can’t monetise their work and protect their intellectual property, then that growth stalls and the industry is damaged. So I believe it’s important that we create use frameworks that align with our values and don’t roll over for mega corporations who really don’t care for our values.
Meta, who own Facebook, have argued that individual works that are used to train machine learning tools have no value in themselves as they, individually, barely affect the performance of Large Language Models. As a Vanity Fair article pointed out, this is a bit like an orchestra arguing against paying individual musicians because a solo bassoon can’t play the whole of an orchestral piece. I’d also point out that in any other arena, it doesn’t matter what the thing you steal is worth, it’s still theft.
Whilst I am very excited about the potential of machine learning to help with current scientific and other problems, particularly in use of large datasets in health, I’m also very concerned that ordinary users of programmes like ChatGPT don’t understand the limitations of these tools. For example, ChatGPT can and does invent information. The best example I heard was on the Skeptics with a K podcast. Presenter Mike, a huge Doctor Who fan, asked ChatGPT for examples of Doctor Who episodes over time. The programme came back with examples of episodes that simply didn’t exist. They were plausible, believable, but totally false. I am worried that this limitation is simply not understood and could do real harm to people’s understanding of the world.