CONCEPT
Training Data as Public Good
The argument—developed most forcefully by Mazzucato and her collaborators—that the corpus of text and knowledge on which AI models are trained was produced by public institutions over centuries, and that its conversion into private proprietary models without compensation is the largest uncompensated extraction of public value in economic history.
A large language model is trained on a corpus of text that includes, by various estimates, trillions of tokens drawn from the accumulated written output of human civilisation. Books, scientific papers, newspaper articles, encyclopaedias, legal documents, software code, online forums, government publications, academic theses—virtually every form of written expression that has been digitised and made accessible. This text was not produced by the companies that train their models on it. It was produced by billions of people over centuries, working within institutions substantially funded by public investment. The books were written by authors educated in publicly funded schools and universities. The scientific papers were produced by researchers funded by government grants, working in publicly funded laboratories. Wikipedia—the largest single source of structured human knowledge ever assembled—was produced by volunteer labour building on the foundation of publicly funded education. The digitisation infrastructure, the internet itself, was