The Federal Government Is Too Broke and Stupid for the AI Revolution

Love it or hate it, AI is blooming in offices across the world. Unless those offices are full of employees working for the U.S. federal government. A new report from Fedscoop casts doubt on Washington’s ability to keep up with the times. According to its analysis of multiple federal agencies, D.C. doesn’t have the funds or the talent to train to keep up with AI.

Last year, President Joe Biden watched Mission: Impossible—Dead Reckoning Part One and got freaked out about AI. It scared him so bad that he signed an executive order that called on tech companies to develop the tech responsibly. The White House also called on federal agencies to issue reports that would detail how they plan to use AI, their plans to mitigate risks against humanity, and to elaborate on what barriers stood in the way of mass adoption of AI.

The White House wanted these reports posted publicly by September. Most of the federal agencies replied and Fedscoop has collected them all in one place and found a common theme. According to Fedscoop, twenty-nine agencies submitted reports. A dozen of them mentioned data hurdles, six mentioned a lack of AI-trained employees, and six said that a lack of funding was hurting their AI initiatives.

The Department of Energy, which is in charge of the nation’s nuclear arsenal, complained that understandable security concerns around cloud services are preventing it from ramping up its adoption of AI. It also doesn’t have enough graphics cards.

“The IT infrastructure barrier extends beyond the serverless CSP services to the availability and timeliness of securing virtual machines with the requisite Graphics Processing Unit (GPU) hardware to develop, train, manage, and deploy advanced AI models,” a DOE report on its adoption of AI said. “This challenge is industry-wide; however, it will impact the rollout and adoption of more advanced customized use cases that require dedicated GPU hardware.”

The recurring data problem is a big one. Many of these federal agencies have been around for decades. People drift in and out, some sticking around for years and others cycling in and out with each new presidential administration. Many of the technological systems at these agencies are ad-hoc in nature. Things get replaced when they absolutely have to be, but not before.

A great example of this is America’s nuclear weapons systems. The Air Force used enormous eight-inch floppy discs to run the software governing nuclear command and control until 2019. When Colin Powell became Secretary of State under George W. Bush in 2001, he discovered that his office was full of pre-internet era computers designed by a company that had gone bankrupt in 1992.

The mass adoption of AI is forcing D.C. to face similar technology challenges all at once across all its departments. The data problem, repeated over a dozen reports, is a reflection of this ad-hoc build-up. Training internal LLMs for government use requires data to be centralized and secure. In many of these agencies, data is distributed in hundreds of different places and few off-the-shelf solutions are secure enough for government work.

Another theme was a lack of understanding of AI in the workforce and an outright fear of it. The Nuclear Regulatory Commission, which deals with nuclear power plants and radioactive materials, said its workforce was interested but “it has also expressed trepidation as well as a general lack of knowledge of AI capabilities. To address this, the agency must continue to enable effective change management to enable the workforce to take full advantage of AI capabilities as they are introduced.”

Many of the agencies also complained about funding. The NRC said it was “only able to assess, test, implement, and maintain new capabilities where resources have been made available to do so.”

“AI use cases compete for funding and staffing with other important priorities at the Bank including non-IT investments in core EXIM capabilities, cyber security, and other use cases in our modernization agenda,” the Export-Import Bank of the U.S. said in its report.

It’s going to be expensive and time-consuming for these agencies to catch up with the rest of the country. Back in 2001, Powell had to buy everyone in his office a computer, 44,000 machines to hear him tell it. “Don’t ask me how I got the money, because I won’t tell you,” he said at an IT symposium in 2019.