InnovateUS 
Read the report: From Curiosity to Capability
P(doom), the estimated probability that advanced artificial intelligence will kill us all, has dominated headlines lately. The dramatic resignation of Anthropic researcher Jacob Coxon, over concerns about “out of control” AI systems, sparked a global wave of panic.
The New York Times wonders, “A.I. Could Possibly End Humanity. How Are Humans Supposed to Process That?” Quick reaction surveys of the public find that nearly two-thirds of Americans say there is at least a “moderate” risk that advanced AI could destroy humanity.
At the same time, the White House launched America.gov, a search tool that promises to make information from 29,000 government websites available through a single front door. An accompanying executive order rebranded AI as “super intelligence.”
AI-powered search using large language models (LLMs) is hardly superintelligence.
A more reasonable argument is that AI is normal technology, especially given how it works in everyday use. But ordinary improvements can matter. Residents shouldn’t need to know which agency handles their problem just to find the right form. Making information easier to find is a long-overdue step in the right direction. The next challenge is improving what happens once they walk through that front door.
AI-powered search is one of many unglamorous ways the government could use AI to help public agencies work better. We spend so much energy debating a possible future catastrophe that we overlook the practical uses already sitting in front of us.
We could make faster progress on these opportunities by teaching government workers how to use AI well to serve the public. That’s the hypothesis behind InnovateUS. We’ve trained more than 250,000 public professionals through courses and workshops in AI, human-centered design, and data analysis to help them improve how government works.
Why we built an observatory
When workers are trained, they report benefits. When Pennsylvania trained 175 state employees in responsible ChatGPT use, they reported saving an average of 95 minutes on days they used it for drafting, research, and summarizing dense documents.
Failing to train can also create risks. In Victoria, Australia, a child protection worker with no AI training used ChatGPT to help draft a report for the Children’s Court. The worker pasted confidential family details and never noticed the chatbot downplayed danger signals.
At InnovateUS, we organized our courses and training to help governments close the learning gap, and early evidence suggests we are already succeeding. At the state level, InnovateUS has trained 32,981 workers in Washington, DC, 29,000 in Indiana, and another 28,000 in Ohio. But training by itself doesn’t tell us whether learning turns into judgment, confidence, and permission to use these tools well.
We are wrestling with questions such as: does the AI tool an agency buys change how a caseworker or a budget analyst does the job? Does learning how to summarize, synthesize, and conduct research with AI translate into better outcomes for residents?
We know far more about governing AI than about governing with it.
Having trained public professionals in all 50 states and DC, we saw an unusual opportunity to help answer these questions by asking public servants how they use AI, what they use it for, and what stops them from trying new things.
To turn that opportunity into an ongoing research effort, we created the Observatory of Public Sector AI.
It’s meant to work like a telescope trained on the workforce, watching how people learn about AI, what they try, and what stops them, in something close to real time.
Academic studies of AI at work often follow a few hundred consultants or software engineers. The Observatory follows hundreds of thousands of public servants.
Every course invites learners to answer short, anonymous surveys before they start and after they finish. In our workshops, which have reached 40,000 people, we can ask questions before, during, and after. A question or two concerns the training itself.
Most of what we ask relates to how public professionals use AI, what barriers they face in using AI on the job, and what they would potentially add to their work tasks once they understand AI.
Today we’re releasing our Observatory’s first report. It’s a window into how the public sector is using AI and, more broadly, how workers respond when AI arrives in their workplace.
What we’re learning
The first surprise is how early in the adoption curve we still are.
Among people who took our most basic course, 51% had never tried generative AI before. Another 14% had tried it at home but never at work.
Nearly two-thirds of the learners we encounter have no workplace experience with generative AI, while the national conversation has moved on to autonomous agents.
The starting point also varies sharply by role: 64% of public safety learners had never tried generative AI, compared with 23% of digital services staff.
Figure 1. GenAI experience before the course, by role. Based on 94,744 pre-course respondents; selected job groups shown.
The second surprise is what’s in the way. It isn’t worker resistance. When we asked non-users why they don’t use AI:
28% said they didn’t think it applied to their work
23% said they didn’t know how to use it
19% didn’t know whether they were allowed.
Managers aren’t much further along. Among managers who weren’t using AI:
24% said they didn’t know how to use it and 22% didn’t know whether they were allowed.
That matters, because the research consistently finds that management support shapes whether employees see a new technology as useful and easy to use. Uncertainty at the top cascades. When the boss can’t tell good use from bad, the safest move is to do nothing.
Among those who do use AI, the work is pragmatic rather than futuristic.
Editing (70%), creating (70%), summarizing (57%), and researching (51%) lead the list.
But the mix of AI use depends heavily on the job. Educators lean heavily on knowledge-related tasks; IT staff write code; and public safety officers organize information and translate.
There’s no single story of AI in government. There are dozens, and they vary by role.
The most hopeful finding is about confidence.
The share reporting a lot of knowledge about AI was 13% before the course and 45% afterward, although those figures come from different groups of respondents.
Those who were not confident planned to use AI for an average of 1.5 different tasks, compared with 3.8 among very confident learners.
Confidence, it seems, is linked to imagining more ways a tool could help. But when we asked public servants about their tasks, they told us how they use AI to write reports and summarize files. Yet AI’s potential to benefit constituents far surpasses these uses.
As the Observatory grows, we want to understand how those new ways of working do or do not translate into quicker decisions, clearer letters, better services, and more public engagement with residents.
Figure 2. How everyday AI tasks could connect to public outcomes. These examples are illustrative, not measured improvements in services.
Building the evidence
The core of the Observatory’s value is scale. This first report draws on more than 106,000 pre-course surveys and 67,000 post-course surveys from a single introductory course, collected between mid-2024 and May 2026. For 22,303 learners, we can match the same person’s answers before and after, which lets us see change rather than guess at it.
These learners aren’t a random, representative sample of public servants. Some take the course because their employer requires it, but most choose to participate. Volunteers start out a bit more knowledgeable and optimistic about AI, so our findings may overstate familiarity and optimism. The surveys capture self-reported knowledge and confidence, along with intentions at the end of the course. They do not yet establish lasting changes in workplace behavior or better outcomes for residents.
Surveys also only get you so far. When we asked open-ended questions at the end of a course about where civil servants imagine AI could take them, the average answer was six words. You can’t learn how someone’s job changed from six words. So we built Public Voice, a tool that lets learners answer our questions out loud, then uses AI to ask follow-up questions like a good interviewer.
Instead of getting a respondent’s curt, “It was helpful,” we’re starting to hear what someone did on a Tuesday with what they learned. Answers are already longer and more specific, though we’re only at the beginning.
We’re also going deeper with partners. Washington, DC, required AI training for its entire workforce, and 32,981 employees, about 86% of the District’s staff, have enrolled. That gives us something rare: a near-census of one government.
Of those we taught, 41% had never tried generative AI, and a third of non-users said it didn’t apply to their work. The share reporting that they were very confident using generative AI was 30% before the course and 49% afterward. At the end of the course, 7% reported not using it at all. These DC results are preliminary.
We will continue to share DC insights with our 75+ government partners, thanks to their commitment to transparency.
What comes next: the AI driver’s license
When DC made AI training mandatory, its Chief Technology Officer, Stephen Miller, explained it simply: “We don’t give someone access to our fleet systems without making sure they have a responsible driving record.”
We come away from this first large-scale study with some strong recommendations about how governments should consider working with AI.
Investing in people alongside platforms runs through everything we found.
First, train the whole workforce. Policies that spell out which tools people can and can’t use are necessary, but they aren’t enough. Without training, we miss opportunities and multiply risks. For us, that means more short videos and more workshops at times people can attend, at their desks.
Second, train managers early and alongside their staff. Managers report skill and permission gaps. We know staff take their cues from the top. Leaders need to get trained and be seen getting trained. For us, that means in-person sessions for managers with our partners, plus more conversations about why their teams need this.
Third, say clearly which uses are permitted. About one in five non-users told us they don’t know what they are and aren’t allowed to do with AI. Vague guidance gets read as a no, or, just as bad, as permission to do the wrong thing. For us, that means sharing policies across partners and measuring how different policies relate to outcomes.
Fourth, teach AI in the context of the job. Use varies widely by occupation, even this early. Explaining what AI is, or what can go wrong, isn’t enough. People need to see the many ways it fits their role. That’s why we’ve launched workshops for lawyers, procurement professionals, and public engagement staff. We’ll soon offer workshops for government communications, human resources offices, and policy-related staff.
Finally, measure outcomes, not just uptake or satisfaction. Counting licenses or course completions tells you about the activities of government. Yet it doesn’t tell you whether residents are getting better service or whether they feel their voices matter, the essential elements of democracy. For the Observatory, that means pairing survey data of staffers with their usage data and working with individual jurisdictions on tailored research to see if beneficial public impacts are being felt.
We spend too much time debating AI policy and buying AI platforms, and too little understanding the public servants who are supposed to use them.
The front door is open. The question is whether the people inside are ready to help whoever walks through it.
Our first report, From Curiosity to Capability, lays out the full findings and recommendations. The companion literature review places them alongside what researchers know about AI and work and shows where government doesn’t fit the private-sector story.
Explore the Observatory of Public Sector AI.