Rethinking Responsible AI in Public Service: Your Staff Already Started Without You

The views expressed are those of the author and do not necessarily reflect the views of ASPA as an organization.

By Thuong “Annie” Bui
July 10, 2026

Most conversations about responsible AI in government begin in the wrong place. They begin with policy: draft guidelines, form a committee, publish a framework and responsibility will follow. It is a comforting sequence, but in my research, it is almost never how things actually happen.

Over the past year, I interviewed 30 municipal IT leaders across various cities in Southern California about how their organizations are approaching generative AI (GenAI). One pattern emerged consistently: employees began using these tools long before any official policy existed. By the time leadership convened its first AI working group, the technology was already embedded in daily work.

I call this the “adoption sequence inversion.” We assume organizations decide to adopt AI and then employees follow. That was the past. For GenAI in local government, the order runs backward. Individual use comes first, and governance is built afterward, often in response.

Once we recognize this reality, the question for public sector leaders changes. It is no longer “should we allow GenAI?” That decision was already made quietly, one prompt at a time, by the workforce. The real question is whether employees have the judgment to use these tools responsibly while formal rules catch up.

You may ask: “Why does judgment matter more than rules?” Here is the part that many AI training programs overlook. These AI systems do not know anything in the way a person knows something. They learn patterns from enormous amounts of data and reproduce the patterns that appear most often. They have no internal understanding of whether a pattern is accurate, fair or appropriate for a specific community. They only know that it is common.

Think of new employees who learn the job entirely from an outdated procedures binder. They will apply those procedures confidently, politely and in the correct format. Nothing about their work will immediately appear wrong. The problem only surfaces when a resident is denied something they were entitled to because the rules changed three years ago and the binder was never updated.

AI outputs carry a similar risk, with one added danger: they arrive looking polished and neutral. A fluent, well-organized answer can feel objective. However, the assumptions underneath it come from the data the system learned from, the way the prompt was written, the context that was excluded and how the reader interprets the result. Bias can enter at any of those points, and none of them are visible in the final output.

This is why I have come to believe that responsible AI in public service is less a document problem than a literacy problem. Not technical literacy. Staff do not need to understand how neural networks adjust their weights. They need something simpler and more difficult: the habit of treating AI output as a draft from a confident stranger rather than a verdict from a neutral expert, especially when that output influences decisions affecting residents.

So what does this look like in practice? In my study, the cities making real progress were those practicing what I describe as “governed learning”: allowing organizations to learn the technology while building guardrails at the same pace.

Three approaches stood out, and none require a large budget:

1. Allow staff to use approved AI tools in a safe workplace environment, such as a sandbox, beginning with low-risk tasks. Employees experimenting openly are easier to guide than employees experimenting secretly through personal accounts.
2. Create rules around the riskiest moments rather than attempting to regulate every possible use. The leaders I interviewed identified three priorities: human review of resident-facing content, protection of sensitive data and a clear process for stopping problematic uses.
3. Provide effective training using real examples by showing employees AI-generated content that appears credible but contains subtle errors. Then allow them to identify those errors. That experience changes behavior in ways generic training often cannot.

Here is the test I would put to any public sector leader: If an AI-assisted process served some residents worse than others, would anyone recognize the problem or would it simply appear as another polished output?

Responsible AI in public service is not a binder on a shelf. It is a workforce that uses a powerful tool with awareness, within boundaries that leadership actively maintains. The technology is already in your organization. The remaining question is whether the judgment to use it responsibly is there as well.

---

Author: Annie Bui is a Doctoral Candidate in Public Administration major, where her research focus on the integration of GenAI in Local Governments. She is the current President of Student Public Administration Association (SPAA) at University of La Verne. At the same time, she works for Small Business Development Center (SBDC), a non-profit organization funded by Small Business Administration, provides zero cost one-on-one consulting services to entrepreneurs for their new and existing businesses. As both a researcher and a practitioner, she committed not just to discussing theories but to actively seeking comprehensive resources that enhance our understanding of how to leverage this technology effectively in the public sector – a sector traditionally slower to adopt technology than the private sector. Her aim is to explore practical solutions that help public leaders to catch up with this fast-changing technology by being ready and proactive in adopting it. Her monthly article series will cover various topics related to GenAI in the current local government settings. Each article is designed to give government agencies the essential knowledge and tools to prepare for an effective and responsible GenAI adoption.

(2 votes, average: 5.00 out of 5)
Loading...