Exploring Memorization and Copyright Violation in Frontier LLMs: A Study of the New York Times v. OpenAI 2023 Lawsuit

The New York Times v. OpenAI lawsuit underscores copyright infringement concerns in large language models (LLMs). Studies measure OpenAI's models' tendency to memorize content, with larger models, especially those over 100 billion parameters, showing significantly higher memorization capacity. This has practical and legal implications, affecting training methods and copyright infringement claims. LLMs trained on copyrighted data risk infringing by reproducing or regurgitating articles. Research quantifies ChatGPT's memorization, focusing on extracting articles from its outputs. The study's main contributions include assessing LLM memorization, quantifying claims in the lawsuit, and discussing legal implications. The text discusses legal aspects of memorization in LLMs, focusing on copyright infringement. It highlights that LLMs, trained on copyrighted content, risk infringing by reproducing or regurgitating articles. The idea/expression doctrine in copyright law is applied, distinguishing between abstract ideas and their original expression. OpenAI argues that memorization is a bug, potentially influencing defenses to infringement claims, especially regarding fair use and contributory infringement. The paper examines three sets of articles from the New York Times, aiming to assess the likelihood of these articles being part of the LLM's training set. It also mentions the inability to reproduce strong claims of near-exact matches, possibly due to differences in models or missing elements in the attack process.