How the Government Publishing Office is using AI to enhance operations
In an interview with Nextgov/FCW, GPO Chief Information Officer Sam Musa said the agency’s AI journey first began in earnest in 2019, after he attended a class on the emerging capabilities at the Massachusetts Institute of Technology. Musa, who was named the agency’s CIO in 2018, said he was interested in learning more about the technology and how it could be used to support GPO’s mission.
This ultimately led to the creation of an AI-specific policy for the agency, which Musa said was meant to address some basic questions: “What is AI? How are we going to provide training to both the developers and the trainers — train the trainers, if you will — and employees? What areas can we leverage AI? What risks [are there] with AI?”
The process of onboarding these new capabilities, however, was not fast and involved working with employees across GPO’s legal and IT units to determine appropriate safeguards.
Even though GPO is a legislative branch agency and does not explicitly have to follow executive branch guidance, for instance, Musa said the agency still adheres to the National Institute of Standards and Technology’s AI risk management framework. He added that, while Microsoft Azure OpenAI “is the engine of our AI initiative,” it took GPO roughly two years to train its developers on how to use the capabilities.
Musa said one of the agency’s early lessons was that “AI is not cheap, and AI takes time to master,” adding that “you will do a lot of training [and] you do a lot of security assessments.” And security still remains top of mind for GPO.
“One thing we learned too — because they're brand new, and everybody's jumping on them — [AI tools] are still not considered secure by design,” he said. “So that's a main concern that you’re always going to have to watch out for. Some of those tools are not FedRAMP certified. For us, that's a showstopper. If a cloud tool is not FedRAMP certified, sorry, out of the picture.”
GPO initially started with several AI pilots, Musa said, including an internal chatbot to help employees get clarifications on existing rules and policies. And while GPO is now also leveraging AI for document classification and summarization purposes, the agency is expanding its uses of these capabilities — including in novel ways.
AI-generated podcasts
Although GPO and many federal entities have rolled out chatbots to help employees answer basic work questions, the agency is likely unique in its use of internal AI-generated podcasts to summarize official documents in a more engaging format.
Musa said GPO has been using Google NotebookLM to create the podcasts, which allows creators to also determine the audio lengths — typically around four to five minutes, he added.
In a podcast snippet played for Nextgov/FCW, the tone is conversational and realistic sounding. The male and female speakers even use vocalized pauses as they set up the conversation.
“Imagine getting an insider look, you know, sitting in on their top executive meetings, getting the real insights without all the dense jargon or those endless slide decks,” the woman says.
“Of cutting through the noise,” the man offers.
“Well, today, that's kind of what we're doing. We're diving deep into the GPO’s executive team off-site report,” the woman continues.
Musa said the podcast in question summarized a long report that originated from an agency strategy meeting.
“No one is going to read the entire 60-page document,” he added. “So I said, ‘Okay, let's go ahead and turn this into a podcast.’ Google puts this for you already with a male and female talking and just interacting back and forth about the points highlighted in that meeting. And they add all the AI stuff in it. Again, that's what I love about it. They make it engaging, and I was really surprised by the good quality and accuracy.”
Musa said the benefits of the AI-generated podcasts also extend beyond simply making policy documents and internal guidance more engaging, noting that they can also provide visually-impaired employees with a new way to digest information. And he added that the agency is already moving to turn more of its guidance into AI-generated podcasts.
“The people that manage the policy directives are converting all the directives [to podcasts],” he said.
“So for those who do not enjoy reading, you just go to the policy page and you will see the actual LinkedIn directive [for example], and next to it you will see the audio file, and it's an MP4,” Musa added, noting that “you just click on it and it plays” but that “you have to be internal to the network” to use it.
Exploring other uses of AI to bolster operations
Musa said AI can be useful for a host of activities across GPO, from streamlining the agency’s hiring and human capital activities to making congressional documents and services more readily accessible.
He said, for instance, that some of GPO’s most basic AI uses can include leveraging the capabilities to come up with interview questions for job candidates, as well as crafting position descriptions for openings. GPO’s acquisition team, he added, is also looking at how AI can help with reviewing its requests for proposals.
Another potential area where AI can prove beneficial, Musa said, is when it comes to fraud detection — something other federal agencies have also been exploring. The Centers for Medicare and Medicaid Services, for instance, launched a challenge to identify emerging tech solutions that can be used to detect fraud in the Medicare program. Musa said GPO’s inspector general asked if the agency could develop a specific AI use case for fraud detection.
Musa also envisions AI improving some of GPO’s more public-facing responsibilities, such as printing Americans’ passport books and publishing official congressional publications. He said, for instance, that AI can potentially help with more precisely measuring the sheets of paper inside passports, thereby helping to reduce waste.
And Musa said that using more precise, AI-powered proofreading tools also has the ability to help with reviewing government documents — including congressional legislation — although he cautioned that it “is a very complex process” and “a lot harder than you would expect.”
“One letter could make a significant impact,” Musa said. “So we tend to hire folks to visually inspect documents and proofread them physically, word-by-word. We want to rely more on technology to do that. But again, the accuracy level is not 100% yet. When it comes to those documents and bills, you have to be 100% accurate.”
]]>