AI use among Centralia city employees is now regulated under an ordinance its board of aldermen unanimously passed Monday evening. Now, when city staff use AI, they must be able to verify the accuracy of any generated content.
Only a third of public sector leaders have a defined AI policy or framework in place, according to a national survey published by software provider Tyler Technologies this year.
City Administrator Tara Strain said generative AI has already helped Centralia staff save time on paperwork and Facebook posts. But because AI tools can hallucinate, or fabricate, information, the ordinance requires staff to verify any generated content for accuracy. And the potential for fabrication isn’t the only limiting factor for AI.
“It’s never going to replace manual labor, and that’s where we’re kind of hurting right now is on our police force,” Strain said.
An additional officer was appointed during Monday’s aldermen meeting, but Strain said the police department currently has a few people out with injuries. That night, the board also considered raising volunteer firefighters’ compensation to improve availability and response. Centralia’s aldermen paused that conversation, citing uncertainty about the number of calls that volunteers would need to be compensated for.
“AI will never replace those positions,” Strain said.
The ordinance and Centralia’s AI policy recognize that using AI at all in city government comes with a measure of risk. City supervisors will decide the extent to which their staff can use AI to ease their workload, giving departments flexibility depending on the type of work they do.
WashU political science professor Jacob Montgomery warns that AI, when used in particular ways, can undermine a government’s legitimacy.
“You would not want it to write messages out to citizens in emergency situations where humans haven’t had a chance to review the content,” Montgomery said.
As part of the ordinance, Centralia staff can face disciplinary action or be fired if they violate authorized uses of generative AI set by their supervisors. AI-generated videos, images, and audio created for city business will also require written permission by supervisors.
“I think small governments are also disadvantaged because creating customized AI infrastructure that, for instance, runs on local compute infrastructure so that data isn’t transferred to third parties can be quite expensive and require specialized staff that they may not have.”
When users type information into an AI tool, the model will train itself on any inputted data. Centralia’s ordinance outlines additional protocols to prevent staff from entering confidential information.
Uncertainty about AI’s place in government
MU public affairs Professor Lael Keiser, who studies how the public responds to governments using AI, explained that models designed for specific city functions can be a risk for residents.
“In that case, the one trade-off is that when individuals are making mistakes, they’re typically isolated; they're not completely widespread,” Keiser said. “But if an AI model makes a mistake, it can affect a lot of people really quickly.”
Keiser’s survey work has found that people are generally more likely to prefer AI systems where humans dominate decision-making or are in the loop.
Montgomery said one of the biggest questions about the implementation of generative AI in governance is whether the public will accept its use, regardless of if it helps city staff.
“So, it comes with a pack of issues that are far beyond just the mere technical issues about whether it works, but how it fits into our governing system,” Montgomery said.
Almost half of the respondents in the Tyler Technology survey reported experimenting with AI tools. As this experimentation by public sector leaders continues, Keiser said she thinks governments’ internal regulation of AI will become more common.
“I think most public organizations are trying to figure out how they can manage that,” Keiser said. “Because there are risks involved, especially just at the very basic level of employees putting in data that’s actually supposed to be private data into one of those systems, not understanding that by doing that, that data is no longer secure.”
