The rise of large language models (LLMs) like ChatGPT has brought incredible advancements in natural language processing. But alongside these benefits comes a potential threat to privacy: the ability of these models to identify the authors of anonymous texts. A new research paper introduces AIDBench, a benchmark designed to evaluate just how good LLMs are at unmasking anonymous authors. The results are concerning. Researchers tested a range of LLMs, including commercial models like GPT-4 and open-source models like Qwen and Baichuan, on various text types, from emails and blogs to research papers. They discovered that LLMs can correctly guess authorship far more often than random chance, raising alarm bells for anonymous systems like peer review platforms. In one test, GPT-4 showed impressive accuracy in identifying authors even across different topics, hinting at the potential for misuse. The study also highlighted a key limitation: many LLMs struggle with long texts, like academic papers. To address this, the researchers developed a Retrieval-Augmented Generation (RAG) method. This method helps narrow down the possibilities by first finding semantically similar texts and then feeding them to the LLM. They found that this approach significantly improves the accuracy of author identification. The implications of this research are significant. While the ability to identify authors could be helpful in some situations (like detecting plagiarism), it also poses a substantial risk to privacy. Imagine a scenario where an LLM is used to de-anonymize whistleblowers or reveal the identities of reviewers in a sensitive peer review process. This research underscores the need for measures to safeguard anonymity in the age of powerful LLMs. Future research could explore techniques to make texts more resistant to author identification, providing a much-needed layer of privacy protection in our increasingly digital world.