Government Finds Way to Use AI Without Exposing Sensitive Data
A Pennsylvania county demonstrated a narrow but practical use of artificial intelligence that sidesteps the privacy concerns keeping many government agencies cautious about AI adoption. At the National Association of State CIOs Midyear Conference in Philadelphia, Crawford County showed how AI can process requests without ever accessing the underlying data.
The core problem government faces with AI is straightforward: if an AI system learns from inputs, how can agencies feed it sensitive information without risking exposure? Many state and local governments have responded by purchasing enterprise versions of AI tools that don't use their data for training. California agencies have gone further, instructing staff not to run anything subject to public records requests through generative AI at all.
Crawford County's approach eliminates the problem entirely. The county added an AI chat feature to its parcel viewer - the online tool citizens use to look up property information. When a user types a natural language question, the AI translates it into SQL code, then hands off that code to the database system. The AI never sees the actual data.
Scott Drzyzga, executive director of Pennsylvania's State Geospatial Coordinating Board, demonstrated the system during the conference. He typed: "find properties within a half-mile that sold for more than $1 million in the last three years." The system returned results in a table with a map showing the locations.
"The AI agent never touches the data," Drzyzga said. "All it does is interpret the plain language question into SQL. It never remembers it, it never sees the owner information, it never sees the address, it never sees the value."
This approach uses AI as an interface layer rather than a data processor. It lets people without SQL knowledge access complex information through conversational language. The parcel viewer itself does the same work it always did - the AI simply makes it accessible to more users.
The application works because the use case is narrow and defined. The AI translates, then exits. It doesn't analyze, learn, or store anything.
For government workers evaluating AI tools, this model offers a template: identify tasks where AI can operate at the interface level without touching sensitive information. Understanding how generative AI and large language models work helps identify where these opportunities exist. Learning about AI applications in government provides additional context on how other agencies are implementing similar solutions.
Your membership also unlocks:
