‘A private ChatGPT-like system’: Sandia Labs deploys AI chatbot for employees aimed at improving workflow

Companies across the globe are searching for ways to adapt and use artificial intelligence to their advantage. Sandia National Laboratories is no exception.

Sandia is the first facility within the nuclear security enterprise to develop a personalized AI chat, called SandiaAI Chat, which allows employees to ask sensitive questions and work more efficiently.

During the rise of OpenAI’s ChatGPT, Michael Vigil, a Sandia project manager, fielded requests from lab workers to implement AI into their work process. But concerns arose about training AI models on sensitive data.

“Our leadership heard this, and they asked us to work with Microsoft and deploy a private ChatGPT-like system,” Vigil said. “The beauty about our system is no one has access to the information, the data, except Sandians themselves.”

In October 2023, Sandia officials began the task of creating an AI application that was safe to use and protects classified information. It took the company less than 27 days to develop an early version of the application and six months of testing before it was introduced to lab workers in May 2024. It was the fastest Vigil had ever seen an application develop at Sandia, he said.

With a new AI model built using the underlying large language model that OpenAI developed, Sandia was able to create a fire-walled version of ChatGPT that allows workers to use the application without security concerns.

With a separate network of servers — Azure Cloud — questions asked by software users stay within the framework of the servers, meaning only Sandia has access to all the data used. Microsoft and OpenAI are not able to access any of the data stored.

“The model is a standalone model,” said Brian Sims, a Sandia solutions architect who was the technical lead for the project. “The foundations of what made all of this work was that we are guaranteed that no third party would have access to the data that Sandians put in and interact with the models with it.”

Sandia also used the coding it created for the AI to make further improvements. During the early stages of SandiaAI Chat, coders would ask the system to train itself, Sims said.

Vigil said the AI also features content filters built into the language model that can identify and flag questionable topic searches, like how to build a dirty bomb, to further protect information.

Since its creation, more than 10,000 workers — nearly all of the lab’s employees — have used SandiaAI at least once, Sims said.

“We need to embrace AI as a valuable partner,” Vigil said. “We need to learn and explore how leveraging AI can help us ... maintain our competitive edge in delivering on Sandia’s national security mission.”