"Want to Keep AI From Sharing Secrets? Train It Yourself"

Artificial Intelligence (AI) is subject to the same privacy regulations as other technologies. In March 2023, there was a security incident in which ChatGPT users were able to view the chat histories of other users, prompting Italy to temporarily ban ChatGPT. This problem could impact any system that stores user data. Italy lifted its ban after OpenAI added features that give users greater control over their data storage and use. However, there are other unique challenges faced by AI. Generative AI models are not designed to reproduce training data and generally cannot do so in any given instance, but it is not impossible. A paper titled "Extracting Training Data from Diffusion Models," which was published in January 2023, describes how Stable Diffusion can generate images that are similar to those in the training data. The Doe v. GitHub lawsuit has examples of code generated by GitHub Copilot, a tool powered by a Large Language Model (LLM) from OpenAI, that matches code in training data. This raises concerns that generative AI controlled by a third-party may inadvertently expose sensitive data. Some generative AI tools, such as ChatGPT, exacerbate this concern by incorporating user data into their training set. Companies concerned with data privacy have little choice but to prohibit its use. This article continues to discuss the privacy issues associated with AI and private AI as a potential solution.  

IEEE Spectrum reports "Want to Keep AI From Sharing Secrets? Train It Yourself"