Washington is consumed with a fight over whether to slow artificial intelligence down or speed it up. Prominent voices have called for a pause; the administration insists that slowing down hands the future to China. But both sides are overlooking where America is already falling behind: putting the technology we have today to work in government.
Think of it like Olympic swimming. It doesn’t matter if you have the best gear, the best pools and the most credentialed coaches if the country never gets in the water. You’ll lose to a nation with mediocre equipment and modest coaches that actually got its people to show up and practice. That’s roughly where the U.S. stands on AI. We have the best labs, the best chips, the most capable frontier models. But we have a public sector that has barely started swimming.
Adoption is the race we’re not showing up for. The case for urgency points somewhere few in Washington are looking — not the next model release, but the DMV, the school district, the benefits agency. It is within public institutions where the technology becomes real for the people it serves. Beijing’s own policy makes this focus explicit: Its “AI Plus" initiative targets more than 70 percent adoption of AI devices and agents by next year, as part of a broader push to integrate AI into everything from hospitals and classrooms to government offices. That’s a country racing for strength by making sure everyone is using the systems that already exist.
Our own two years of research at RethinkAI found mostly avoidance and uncoordinated experimentation. State lawmakers have proposed more than 1,600 pieces of AI-related legislation since 2019, but 77 percent of those bills focus on restricting or auditing AI use, reflecting how much more attention lawmakers have devoted to regulating the technology than to helping their own governments put it to work. Code for America’s 2026 assessment of all 50 states and the District of Columbia found the same pattern: While experimentation has increased, most deployments “remain confined to narrow use cases” and few states have established ways to evaluate the value of AI deployments.
There’s a reason we keep making this mistake: Every incentive in the system rewards speeding up AI development and punishes the unglamorous work of adoption. Money and headlines flow to the next model release; almost none flow to a state agency’s data-modernization project or a school district’s staff-training budget.
Adoption is also a trust problem. A resident won’t accept an AI-assisted decision about their benefits, their building permit or their child’s school placement because the underlying model is powerful. They’ll accept it once institutions show, repeatedly and visibly, that the system is accountable and fair. That’s management and policy work, not an engineering challenge, and it’s what gets skipped when attention is fixated on the next AI release cycle.
What would taking adoption seriously look like? Governments need clear guidance about where AI can be used and where it shouldn’t be; safe environments for employees to test applications; opportunities for residents to shape how the technology is deployed; and practical training for managers and frontline workers. Just as importantly, agencies need to measure whether AI is actually improving services, not simply saving staff time. None of this necessitates big expenditures, but it does take time and staff resources.
Some governments, particularly at the state level, are beginning to show what building this institutional capacity looks like. Pennsylvania spent a year testing generative AI with state employees, establishing safeguards, evaluating results and learning how the technology changed their work. It has since expanded AI use to more than 3,000 employees across 35 agencies. Similarly, other states, including Colorado, New Jersey and Utah, have created more centralized AI capacity and supervised environments to securely deploy and assess AI use.
You also see encouraging signs in the broad commitment to staff training. At InnovateUS, a public-sector workforce program I help lead, we’ve partnered with more than 80 state and city governments and reached 275,000 learners with training in AI and other digital skills. But even that reach, combined with the state examples above, represents only a fraction of the public workforce nationwide. That gap, not the speed of the frontier, is what’s holding the country back.
We have spent years building the machine that supplies AI. We have barely begun building the institutional capacity to make good use of it. That’s the race that will actually decide who benefits from this technology, and it isn’t the one being run in the labs.
Governing's opinion columns reflect the views of their authors and not necessarily those of Governing's editors or management.
A founder of RethinkAI, a professor of practice at Northeastern University’s Department of Public Health and a senior fellow at its Burnes Center for Social Change