Uncovering the Risks of AI: LLMs, Watermarks, and OpenAI's Cloud

Researchers from the University of North Carolina, Chapel Hill have published a preprint paper highlighting the challenges of removing sensitive data from large language models (LLMs) like OpenAI's ChatGPT and Google's Bard. While it is possible to delete information from LLMs, verifying its removal is equally difficult. LLMs are pretrained on databases and fine-tuned, making it impossible to delete specific files to prevent related outputs. The researchers found that even state-of-the-art model editing methods fail to fully delete factual information from LLMs. Defense methods are constantly playing catch-up to new attack methods, making the problem of deleting sensitive information an ongoing challenge.