Scientists from the University of North Carolina have published new research, casting light on the complexities surrounding the deletion of sensitive information from large language models like OpenAI's ChatGPT and Google (NASDAQ:GOOGL)'s Bard.
The paper underlines a significant potential for these Large Language Models (LLMs) to output sensitive information such as personally identifiable information or financial records.
According to the research, the process of deleting information from these LLMs is possible, but it’s just as difficult to verify the information has been removed as it is to actually remove it.
AI black-box
Large language models are pretrained on extensive databases and later fine-tuned to generate coherent text.
Because of this, it is not straightforward for developers to go back into the database to delete specific files or data.
The very nature of these models renders the information in their architecture almost undefinable without generating some form of output, contributing to what is known in the industry as the 'AI black box'.
Risks
The risks arise from the model's inability to selectively forget information.
According to the researchers, while it is technically possible to remove sensitive data, confirming its complete erasure is equally challenging.
This research resonates particularly in an era where AI is increasingly becoming integrated into various sectors, including finance, healthcare, and technology.
It raises essential questions about data privacy and the ethical responsibilities that come with deploying these large-scale models.
The findings are expected to fuel ongoing discussions in the field about the feasibility of "forgetting" within machine learning models and may influence future developments aimed at enhancing data privacy measures.