A new artificial intelligence model calls into question the use of protected content in the training of language systems. Mr. Chatterbox, a chatbot created exclusively with public-domain texts from the 19th century, shows that it is possible to develop AI without resorting to copyrighted data. Engineer Trip Venturella developed the model and trained it on more than 28,000 works from the British Library, published between 1837 and 1899. The system uses nearly 2.93 billion tokens and does not incorporate any content after that date.
More than 28,000 works from the British Library
The model, publicized by researcher Simon Willison, uses a corpus of more than 28,000 works from the British Library, published between 1837 and 1899. In total, it uses nearly 2.93 billion tokens, all of them from texts prior to that date.
This is VibeVoice, Microsoft's bet on open source voice AI
With about 340 million parameters, similar in size to intermediate versions of models such as GPT-2, the Mr. Chatterbo file takes up barely 2 GB and can run locally, without the need for a connection to external servers. However, its performance is far from current standards. The model's responses, although stylistically consistent with Victorian language, are imprecise and, in many cases, lack practical usefulness. According to Willison, interacting with the system is more like a probabilistic string of text than a structured conversation.
The data challenge in artificial intelligence
The experiment underscores the need for large volumes of data in the development of language models. Recent studies suggest that, to achieve competitive performance, the amount of training data should be twenty times the number of parameters in the model, which in this case would imply more than double the corpus used.
Even so, the project shows that it is possible to train functional systems using only copyright-free content, an issue that is increasingly relevant among technology companies and content creators. Another notable aspect of the project is its accessibility. Mr. Chatterbox can be installed and run locally using open source tools, which makes it easier for developers and enthusiasts to use without the need for complex infrastructure.
Willison himself stresses that this type of initiative could pave the way toward more transparent models that respect copyright, although he acknowledges that there is still a long way to go to reach the level of usefulness of current commercial systems.