AI Can Help Governments Hear More Citizens Without Deciding Which Voices Count

Software architect Oleksandr Kostyrka argues that civic AI should operate inside an auditable, government-controlled data pipeline. The technology can help public employees review large volumes of feedback as long as original submissions stay authoritative, models stay replaceable, and humans stay responsible for every conclusion. 

Artificial intelligence could make public participation easier and simultaneously make the resulting feedback harder to review. 

The scale of that problem is already visible. In 2025, the White House received more than 10,000 comments during a seven-week consultation on its AI Action Plan. A local planning department may receive fewer comments, but it also operates with a much smaller review team. 

Oleksandr Kostyrka, founder of Crumina and a senior software developer at Salt Lake City-based Penna Powers, has spent more than a decade designing commercial software and public-facing digital systems. His work has included platforms serving transportation, higher education, economic development, public information, and community engagement in Utah. 

He believes one of AI’s most valuable civic applications will be helping public employees navigate feedback without replacing the judgement of the people responsible for reviewing it. 

“The goal is not to let AI decide what the public thinks,” Kostyrka says. “The goal is to give staff a map of what was submitted, with a clear path back to every source.” 

Public Comments As Part of the Administrative Record  

A public comment may become part of an administrative record used by planners, elected officials, attorneys, residents, and future reviewers. It can contain a policy argument, a description of personal circumstances, or evidence about how a proposal may affect a particular property or neighbourhood. 

That context changes how AI should be introduced. 

“A public comment is part of a democratic process,” Oleksandr says. “An AI-generated summary can assist with review. The original statement remains the authoritative record.” 

Under the architecture he is developing, the agency would preserve every submission unchanged. AI would work on a controlled copy and attach analytical results to the source record. Themes, summaries, and similarity groups would function as navigational aids rather than replacements for what residents submitted. 

AI As a Review Tool 

Oleksandr Kostyrka sees several useful tasks for an AI-assisted review system. It could identify recurring subjects, group substantially similar passages, surface low-frequency concerns, and help staff move through a large collection of comments. 

Each task needs limits. 

Clustering similar language, for example, does not prove that a campaign is deceptive or improperly coordinated. Residents often participate through neighbourhood groups, advocacy organisations, unions, professional associations, or shared letter-writing campaigns. Those are legitimate forms of civic participation. 

“Similarity is a fact about language,” Oleksandr Kostyrka says. “The system can show reviewers that 300 submissions contain substantially similar text. What it should never do is decide that those 300 people are less legitimate.” 

A single comment about wheelchair access to a proposed facility, for example, may be more consequential than a large cluster of general statements expressing support. An effective review system must help staff find both the dominant subjects and the exceptions that could materially affect a decision. 

This is why Oleksandr prefers the term “review assistant” to “decision engine.” The system can reduce the time required to locate and compare information, but public employees must interpret that information within the legal and policy context of the consultation. 

Privacy Depends On the Entire Data Path 

Public comments often include names, email addresses, phone numbers, property addresses, employment details, and descriptions of medical or financial circumstances. Even when a comment is subject to public-records laws, sending it to an AI provider creates a separate processing event with its own retention, security, and governance questions. 

A March 2026 U.S. Government Accountability Office report identified continuing gaps in federal guidance addressing privacy risks from government AI use. Among the issues raised were separating sensitive data, evaluating privacy impacts, informing the public about the use of personally identifiable information, and understanding the trade-offs between privacy protection and model performance. 

Oleksandr Kostyrka argues that agencies need control over the complete data route. That includes deciding which information reaches a model, where processing occurs, how long temporary copies remain, whether data can be used for training, who may access the output, and when derived records must be deleted. 

“Privacy is a set of decisions about how information moves through the entire system,” he says. “Treating it as a checkbox attached to a model misses the point.”  

Government-Controlled Civic AI 

Crumina is exploring an optional AI-assisted review capability for Crumina Civic, its open-source platform for local government projects, documents, meetings, feedback, and constituent management. The capability is in the design stage and is not currently available as a released feature. 

Kostyrka’s proposed architecture would not require every agency to use the same language model or cloud provider. Instead, it could support several deployment patterns based on an organisation’s infrastructure, resources, and risk requirements. 

A government with sufficient technical capacity could operate a private model inside its own environment. Comment data would remain within infrastructure controlled by the agency, although the agency would also assume responsibility for securing, evaluating, and maintaining the model. 

These deployment patterns give agencies explicit control over where risk is accepted and how that risk is monitored. 

The Public Record Must Outlast the Model 

Language models are changing faster than most government procurement cycles. A model selected today may become too expensive, fall behind competing systems, or no longer meet an agency’s privacy requirements. 

Oleksandr believes civic platforms should treat the model as a replaceable component rather than the foundation of the system. 

“The public record and the rules governing it will outlast any model,” he says. “That is exactly why the architecture should make every model a swappable component.” 

This approach aligns with the NIST AI Risk Management Framework, which treats governance, measurement, documentation, and monitoring as continuing responsibilities throughout the life of an AI system. 

Before deployment, Oleksandr Kostyrka says agencies should test models against representative historical comments. Evaluation should measure whether the system preserves minority concerns, invents unsupported themes, exposes personal information, or changes its conclusions when comments are presented in a different order. 

Trust Requires Visible Limits 

The role of AI in this context is specific: help reviewers find evidence, compare submissions, identify questions that need closer attention, and document how they reached their conclusions. Deleting comments, deciding which submissions are legitimate, or making policy decisions on behalf of elected officials falls outside that role. 

For Oleksandr, this is the standard by which civic AI should be judged. 

“The real measure of civic AI is whether the government can hear more people, preserve what they actually said, and remain accountable for every decision that follows,” he says. “Speed and headcount savings are secondary to that standard.”  

AI can help public institutions process more information. Turning that processing into genuine listening requires context, traceability, and responsibility — obligations that belong to the government, regardless of which model performs the initial analysis.  

Disclosure: Oleksandr Kostyrka is the founder of Crumina and leads the development of Crumina Civic. The AI-assisted feedback capability described in this article is planned and is not currently available as a released feature.