Enterprises are rapidly adopting large language models (LLMs), yet many are unable to build foundational models due to the staggering investments required. As a result, organizations often rely on off-the-shelf solutions from providers like OpenAI. However, these models come with inherent vulnerabilities. One well-documented concern is the ability of LLMs to memorize and inadvertently expose sensitive information from their training data during inference. This includes private details such as email addresses, phone numbers, physical addresses, and even copyrighted content (as highlighted in cases like NY Times vs. Microsoft and Authors vs. Meta).

For organizations leveraging LLMs, understanding the risks around data privacy and compliance is essential. Join Sriram as he explores the latest research like machine unlearning and inference time redaction on this topic. Beyond ethical considerations, the exposure of confidential or copyrighted information by an LLM integrated into your services could lead to significant legal and financial consequences.

By the end of this talk, you will gain insights into the ethical and legal risks associated with LLMs, the current advancements in mitigating these issues, and the limitations of existing solutions.