Half of Government Workers Have Never Tried AI, Survey of 250,000 Finds

Fifty-one percent of public-sector workers entering a foundational generative AI course had never used the technology. That figure comes from the first report of InnovateUS's Observatory of Public Sector AI, published by the nonprofit that has trained more than 250,000 government workers across all 50 states and the District of Columbia.

The dataset is large: 106,300 precourse survey responses and 67,002 postcourse responses from participants in the introductory generative AI course, collected between June 2024 and May 2026. The findings, the authors say, reveal a substantial gap between AI headlines and everyday working life in government.

What keeps public workers from using AI?

The numbers complicate the stereotype of government employees as resistant to change. Among non-users, 28% said AI didn't apply to their work, 23% didn't know how to use it, and 19% didn't know whether they were allowed to use it at all. Another 14% had experimented with AI personally but not at work.

Managers fare no better. Among managers not using AI, 24% didn't know how, and 22% didn't know whether it was permitted. Employees look to supervisors for guidance, the report notes — supervisors who may need help themselves.

The takeaway in plain words: handing workers a tool or publishing a policy leaves most of the job undone.

Do beginners benefit most?

Yes, according to the matched-response data. Learners who entered with no reported AI knowledge showed the largest gains in reported knowledge. Those starting with no confidence improved most in confidence.

This parallels private-sector findings. A study of customer-support workers by academics at Stanford and MIT found less experienced, lower-skilled workers gained the most from an AI assistant — though that study measured workplace performance, while InnovateUS measures self-reported learning. The comparison is suggestive, not equivalent.

Confidence also correlates with ambition. After training, highly confident respondents planned to apply AI to an average of 3.8 tasks, versus 1.5 among those with low confidence. The authors are careful: the association does not establish that confidence causes broader use. And confidence alone is an incomplete measure — someone can feel more capable without becoming more accurate.

The one-sentence takeaway: training helps most those who start with least, but judgment must be taught alongside confidence.

Does faster work mean better services?

Not automatically. McKinsey's 2026 global survey found 80% of respondents said AI improved their own productivity, yet only 37% reported a positive impact on their organization's earnings. InnovateUS sees the same pattern: participants described tasks — drafting reports, creating presentations, summarizing documents — but fewer connected those tasks to outcomes like more accessible services or better public participation.

The authors concede their own questions partly explain this; they asked about specific tasks. They have begun collecting voice-based feedback and adding questions probing whether learning improves services.

For democratic institutions, the test is whether time saved shortens benefits queues, helps more people complete applications, or improves decisions from public feedback. Those links remain unproven.

What are the study's limits?

Plenty, by the authors' own account. The surveys are voluntary; participants are a large but not representative sample of the government workforce. Results measure reported knowledge, confidence and intended use immediately before and after a course — not retention, consistent application, fewer mistakes or better services. There is no comparison group, so the effect of training cannot be isolated from other influences.

One workshop anecdote illustrates the promise without proving it: a state leader in Colorado reported using generative AI for performance reviews, with better feedback, greater worker satisfaction, reduced turnover and time savings. That is a self-report, not an audited outcome.

What comes next?

Washington, D.C., which has adopted a whole-of-workforce training approach, offers a chance to study how individual skills interact with managerial support, tools and agency rules. The authors' next step is connecting learning data with administrative measures — wait times, case-resolution accuracy, application errors — under privacy protections.

Even that will require careful design. A shorter queue could reflect staffing changes, new software, or lower demand, not training. Until those studies exist, the claim that AI training improves public services rests on reported confidence, not measured outcomes. The authors' own standard makes the point: success will not be how many employees use AI or how much time they save, but whether residents experience better services and regain trust in government.