At a glance
- What it is: Thoughtful Adoption of NLP for Civic Participation: Understanding Differences Among Policymakers
- Who: Jose A. Guridi, Cristóbal Cheyre and Qian Yang (Cornell University)
- Where: Proceedings of the ACM on Human-Computer Interaction (CSCW), April 2025. doi.org/10.1145/3711091
- Type: qualitative study, 20 interviews.
First reading: what it does and what it finds
It helps to start with what kind of study this is, because that organizes everything else. It is a qualitative work. The authors interviewed 20 people (7 appointed officials and 13 career civil servants) in five ministries in Chile and in Uruguay’s digital government agency, AGESIC, and analyzed those conversations. They do not measure whether NLP improves participation, nor do they evaluate any model. They explain how two different groups think about these tools and how they say they act.
The starting point comes from prior literature, not from this study: NLP can help organize large volumes of citizen comments, and yet governments barely use it. The paper’s own question is why. And its contribution begins with a distinction that the literature tends to flatten: instead of treating “decision-makers” as a single bloc, it separates politicians (the appointed officials who head institutions) from civil servants (the career staff who design and implement). Both look at the same tool and see different things, because they build their legitimacy in opposite places.
Politicians build it outward, in citizens’ trust and good press. That is why they are drawn to NLP that looks objective and modern, and that dispels the suspicion of human manipulation when opinions are counted. “The machine doesn’t interpret, it just processes data,” says one. The study names the risk this hides: the illusion of objectivity, believing that something is reliable just because no person touched it.
Civil servants build it inward, in front of their superiors, and they work drowning in volume. All of them talked about reducing their workload. But since they are the ones who operate the system, they see the concrete risks that politicians barely mention: classification errors, sensitive data, surveillance. They do not want AI to replace them. They want, in the words of one civil servant, a team member that systematizes and that they can then question.
The third finding is the one that gives the title. Neither group assumes who should drive adoption or take charge if it goes wrong. They blame each other: politicians say civil servants resist and do not know about technology; civil servants say politicians do not lead or give them the time or infrastructure to learn. Added to this is an obstacle both acknowledge: public procurement is not built for this. One interviewee spent more than a year trying to contract a mass email service. For the authors, what holds back adoption is not the quality of the algorithm, but that lack of clarity about who is responsible.
Second reading: from Latin America
Something the authors themselves stress: this is evidence from inside the region, with interviews in Spanish and in real ministries, not extrapolated from the Global North. The literature in their field barely studies these countries, so the value lies in taking a close look at something rarely examined.
Where you look from matters. Chile and Uruguay are two of the countries in the region with the longest track record in digital government: the Uruguayan agency AGESIC is a benchmark, and Chile has a National AI Policy. If the lack of clarity about who is responsible and the friction of public procurement show up even there, it is reasonable to think they weigh more heavily where the civil service is weaker and political turnover is higher. This last point is my own reading: the paper did not study those other countries, and I leave it as a hypothesis, not a finding.
Where the study connects with the present is in the change of era. The interviews portray a world where NLP was contracted from a vendor. Today a team can paste citizen comments into a chatbot without a tender or an audit. That does not make the paper obsolete, it makes it more relevant. Its findings on technical barriers are the ones LLMs are erasing. Its findings on who builds legitimacy and in which direction, and on the lack of clarity about who answers for the result, still stand. When using the tool is that easy, what keeps getting in the way is not the model’s capability, but who decides and who takes responsibility. The authors themselves point to this when they warn that LLMs can lower technical barriers and bring new risks.
From there come three ideas that the paper suggests and that translate directly to the region, bearing in mind that they come from interviews and not from an experiment. The first: do not offer AI for participation only as an efficiency gain, because politicians are moved by legitimacy. The second: what civil servants asked for, pilots, human validation and the ability to trace results back to the raw data, should be the floor and not a luxury. The third: someone should explicitly have the responsibility of deciding whether to adopt and also of putting on the brakes.
The question it leaves works for any ministry in the region. Now that using AI is as simple as pasting the comments into a chatbot, who answers for it when that is used to listen to citizens? If the answer is no one, the study suggests that is where the problem lies, and not in the algorithm.
The fine print
- It is a qualitative study: its value lies in explaining the why and the how of these decisions, not in measuring how many or proving that NLP works. The numbers in the tables say how many interviewees mentioned each thing, not how widespread it is.
- It talks about NLP at a time before LLMs became widespread. The authors themselves warn that this advance may change some details. The organizational side, by contrast, ages slowly, because it is about people and institutions.
- Transparency: the author of this blog is a coauthor of the paper.
Paper keywords: Artificial Intelligence, Public Participation, Stakeholders, eGovernment, Policymakers, Natural Language Processing
Automated reading. This text was generated by Claude, an Anthropic model, from the original source, without line-by-line human review. It may contain errors or debatable interpretations; to check any point, see the original source.