The General Services Administration’s (GSA) newly launched generative AI tool is reserved for internal use for now, but GSA Chief AI Officer (CAIO) Zach Whitman said today that the tool has the potential to be applied across government to empower the entire Federal workforce.
GSA launched the tool – known as GSAi – in March to increase internal productivity. Whitman said the tool features a chatbot, an application programming interface (API), and “an administration console” – which he said can evaluate different models for different purposes.
“I’m really excited about that project because I think it has general application across other agencies, and it focuses on the quality aspect,” Whitman said on Tuesday at GovCIO’s AI FedLab event in Reston, Va.
“Instead of focusing on only safety concerns … we are focusing on the quality of the performance of each model based on the specific function areas of the agency,” said Whitman, who also serves as GSA’s chief data scientist.
For example, Whitman said GSAi is able to evaluate how well a specific AI model can answer questions about procurement.
GSA can evaluate the model’s ability to answer procurement-related questions over time, detect model drift, and determine if the model is fit for purpose – or if it needs interventions like fine-tuning or retrieval-augmented generation (RAG).
“We don’t want to empower our workforce with a tool that’s just giving random advice,” Whitman said. “At least we’re having that open conversation: this is how good it is, and this is what it’s suitable for – versus kind of operating blind and just saying, ‘buyer beware, you better double check your facts.’”
“Having a quality measurement of its performance and then being able to understand what to do with that is really critical,” he added.
Whitman also pointed to a surprising secondary benefit of GSAi’s rollout: it has become a catalyst for long-overdue conversations about data quality and metadata.
“We want to use AI as that kind of Trojan horse to backdoor metadata,” Whitman explained. “There’s no better way to demonstrate the quality, or the lack thereof, of your data stores by putting an AI engine in front of it and letting it hit its head against the wall.”
“It’s just nice to be able to have that conversation and just see the energy in the A suite about, like, ‘Okay, we need to talk about metadata,’” he added. “So, it’s been a really nice little happy accident.”