Researchers tested leading AI models for copyright infringement using popular books, and GPT-4 performed worst.

A group of researchers, former employees of Meta, dedicated their 2023 to developing a new artificial intelligence software, called 'Patronus Ai'. The aim of the company, of the same name, is to evaluate and test the famous LLMs (large language models) - currently recognised as the technology carrier behind AI products.

Recently, Patronus AI released a new tool called 'CopyrightCatcher', aimed at supporting research into how often leading artificial intelligence models generate content based on originally copyrighted examples. At the same time, the company published the results of a test carried out in this regard.

The test took into consideration OpenAI's GPT-4, Anthropic's Claude 2, Meta's Llama 2 and Mistral AI's Mixtral . The candidate with the highest level of copyright infringement turned out to be GPT-4, with 44% of results containing protected content.