Microsoft has unveiled a generative AI model based on GPT-4, specifically developed for US intelligence agencies, which operates without internet connection. This marks Microsoft’s first deployment of a large language model in a secure environment. It allows intelligence services to analyze top-secret information without the risk of internet leaks and securely communicate with a chatbot resembling ChatGPT and Microsoft Copilot. However, improper usage of the model could potentially mislead agency personnel due to inherent limitations of AI language models.
The new artificial intelligence model (the name of which is currently undisclosed) responds to the growing interest of intelligence services in using generative AI for processing classified information, while reducing the risks of data leaks or hacking attacks. Microsoft holds a license to use GPT, having invested $10 billion in OpenAI. While ChatGPT operates on Microsoft’s cloud servers, which could lead to data leaks and interception.
Last year, the CIA announced plans to create its own service similar to ChatGPT, but according to reports, the Microsoft project is a separate initiative.
William Chappell, Microsoft’s director of emerging projects and technologies, told Bloomberg that the development of the new system took 18 months. Engineers modified a supercomputer with artificial intelligence located in Iowa. Based on GPT-4, the model can process files provided by users but does not have access to the open internet. Chappell stated, “This is the first instance where we have an isolated version, meaning it doesn’t have an internet connection. It operates in a special network accessible only to the US government.”
The new service was activated last week and is now available for testing to approximately 10,000 members of the intelligence community. According to Chappell, the chatbot “answers questions” at the moment.
One of the main drawbacks of using GPT-4 for analyzing important data is that it can generate inaccurate answers, draw incorrect conclusions, or provide users with incorrect information. Neural networks are not databases and operate based on statistical probabilities. To mitigate this issue, retrieval augmented generation (RAG) methods can be used, allowing AI to access information from external sources.
Given this limitation, it is entirely possible that GPT-4 could misinform or mislead US intelligence services if improperly used. There is currently no information on how the system will be controlled, what restrictions will be imposed on its use, and how its accuracy will be verified.
