Read the following passage and mark the letter A, B, C or D to indicate the correct answer
AI Ethics and Copyright Assessment
Amid growing concerns from artists, writers, and other creators about AI models infringing on their copyrighted works, recent research has found that top AI models generate copyrighted content at an alarmingly high rate. Patronus AI, a startup founded by former Meta researchers, released a tool called CopyrightCatcher on Wednesday to detect potential copyright violations in large language models (LLMs).
The company evaluated four major AI models: OpenAI’s GPT-4, Anthropic’s Claude 2.1, Mistral’s Mixtral, and Meta’s Llama 2. Of these, GPT-4 generated the most copyrighted content at 44% of the prompts, followed by Mixtral at 22%, Llama 2 at 10%, and Claude 2.1 at 8%. Patronus AI tested these models using books under copyright, including Gone Girl by Gillian Flynn and A Game of Thrones by George R.R. Martin. While some generations may fall under fair use in the U.S., GPT-4 completed book texts 60% of the time and generated the first passage 26% of the time. Mixtral generated the first passage 38% of the time and completed texts 6% of the time. Llama 2 generated both first passages and completed texts 10% of the time, while Claude 2.1 completed texts 16% of the time and generated the first passage 0% of the time.
Rebecca Qian, cofounder and chief technology officer at Patronus AI, expressed surprise that GPT-4, widely regarded as the most powerful model, produced copyrighted content on 44% of constructed prompts. OpenAI, Mistral, Meta, and Anthropic did not immediately respond to requests for comment. Patronus AI emphasized the importance of detecting these reproductions to avoid legal risks and protect company reputations, noting that LLMs, which are trained on data including copyrighted works, can easily generate exact reproductions.
Question 3: In the study, which book was mentioned as part of the evaluation for potential copyright violations?