Memorandum
- From
- Derek Weston via Fast Company
- Date
- Filed
- Society·5 min to read
- Re
Government workers remain unsure how to use AI, InnovateUS survey finds
ReGovernment workers remain unsure how to use AI, InnovateUS survey finds
A study of more than 250,000 public-sector workers trained by the nonprofit InnovateUS finds that many government employees are unsure whether generative AI applies to their jobs or whether they are permitted to use it.
Many government employees are still unsure how generative AI applies to their jobs, how to use it, or whether they are allowed to use it at all, according to research from InnovateUS, a nonprofit learning initiative that has trained more than 250,000 public workers in AI and digital skills. The findings suggest that the gap between the enthusiasm surrounding generative AI and everyday working life in government remains wide, even as the technology could help improve public services.
InnovateUS works with state and local government employees across all 50 states and the District of Columbia, in roles ranging from public safety and benefits administration to legal services and executive leadership. Its no-cost workshops, held across the country and online, give public servants a chance to compare experiences about what works and what does not. Participants have explored using artificial intelligence to improve benefits delivery with New York City's health commissioner and have heard from current and former state leaders about how to introduce workers to using, and even building, AI tools.
The organization's first report from its Observatory of Public Sector AI draws on 106,300 precourse survey responses and 67,002 postcourse responses from participants in its introductory generative AI course, collected between June 2024 and May 2026. Among public-sector respondents entering the course, 51 percent had never tried the technology, and another 14 percent had experimented with it personally but were not yet using it at work.
Among respondents who were not using generative AI, 28 percent said they did not think it applied to their work, 23 percent did not know how to use it, and 19 percent did not know whether they were allowed to use it on the job. Managers reported similar uncertainty: among managers who were not using AI, 24 percent did not know how, and 22 percent did not know whether it was permitted. The authors note that employees often look to supervisors for guidance, but supervisors may need support of their own.
The researchers argue that these findings complicate the common stereotype of government workers as resistant to change. They write that providing access to a tool or publishing a policy leaves much of the work undone. Employees need relevant examples, chances to practice, and managers prepared to support responsible use. The report points to concrete service problems where AI could help, such as reducing wait times for people calling about health or unemployment insurance, making it easier and cheaper to involve the public in policymaking, and helping agencies regulate and buy AI in ways that reduce dependence on any single vendor.
One of the more encouraging results concerns workers who start with the least familiarity. Learners who entered the course with no reported AI knowledge showed the largest gains in reported knowledge, and those who began with no confidence also reported particularly substantial improvements in confidence. The researchers note a parallel with private-sector research by academics at Stanford and MIT, which found that less experienced and lower-skilled customer-support workers gained the most from an AI assistant. That study measured workplace performance, while the InnovateUS data measures reported learning. Together, the authors suggest, the findings indicate that organizations should not concentrate support only on enthusiastic early adopters.
The data also show a link between confidence and the uses people can imagine for generative AI tools. After training, highly confident respondents planned to apply AI to an average of 3.8 tasks, compared with 1.5 among those with low confidence. The researchers caution that this association does not establish that confidence causes broader use. They also warn that confidence is an incomplete measure of success, since a person can feel more capable without becoming more accurate. In their view, training must build the judgment to question an answer, verify a source, protect sensitive information, and recognize when AI is inappropriate.
4
