aceEmerging evidence
More
In a 73-person tax test, a GPT assistant sometimes gave wrong answers that sounded right
Study title: Conversational-agent-mediated public encounters: an analysis of citizen communication in public service delivery
- Evidence type
- Interviews or qualitative study
- Sample
- n = 73
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2573cite or check it
Study title: Conversational-agent-mediated public encounters: an analysis of citizen communication in public service delivery
A qualitative study analysed conversations between 73 participants and a GPT-based conversational agent that helped them fill in a fictitious tax return. The study found that the agent occasionally gave incorrect but plausible answers, known as hallucinations, despite efforts to avoid them.
Read the study abstract
Purpose Conversational agents are often used in public organizations as an entry-level application of artificial intelligence and constitute a form of digitally mediated public encounter. However, little is known about how citizens communicate in conversational-agent-mediated public service interactions. By analyzing conversations between citizens and a conversational agent, this article formulates three propositions that contribute to understanding the communicative dynamics of such encounters. Design/methodolo…
The problem
In a qualitative study of 73 participants filling in a fictitious tax return, a GPT-based conversational agent occasionally gave incorrect but plausible answers, even though the researchers tried to prevent it.
In brief
The findings matter because conversational agents are often an entry-level use of artificial intelligence in public organizations, and they form a digitally mediated encounter between citizens and the state. The authors describe vulnerabilities that arise from the agent's role as an information intermediary.
Citizens in the experiment asked simple, concise questions and did not use elaborate prompt engineering. They also expected the agent to know a lot about the topic, which raises the stakes when it is wrong. The study used a fictitious tax return, so real-world consequences were not measured.
Main points
- The experiment had 73 participants who filled in a fictitious tax return with help from a GPT-based conversational agent.
- Researchers analysed the conversation logs qualitatively and drew three propositions about how citizens communicate with such agents.
- Citizens communicated in a highly task-oriented way, using simple and concise questions rather than elaborate prompt engineering.
- Participants expected the agent to have a lot of knowledge about the topic in question.
- In rare cases the agent gave incorrect but plausible answers (hallucination), even though this was tried to be avoided.
- The authors describe the study as among the first to analyse conversation log data from citizen interactions with a GPT-based agent in a public-service setting.
The full story
The data came from an experiment in which participants had to fill in a fictitious tax return while assisted by a GPT-based conversational agent. The researchers analysed the resulting conversations qualitatively, with 73 participants (N = 73).
From the logs they formulated three propositions about conversational-agent-mediated public encounters: strong expectations about information provision, limited engagement with prompting strategies, and vulnerabilities arising from the agent's role as an information intermediary.
Participants sought answers through short, direct questions instead of crafting elaborate prompts. The agent was expected to be knowledgeable about the topic, while in rare cases it produced answers that were wrong yet plausible.
Sources
1 on file- International Journal of Public Sector Managementdoi:10.1108/ijpsm-06-2025-0236
Talk it over
Who carries this cost, and what would have to change to prevent it?
In a 73-person tax test, a GPT assistant sometimes gave wrong answers that sounded righthttps://truvace.com/item/conversational-agent-mediated-public-encounters-an-analysis-of-citizen-communica
Who carries this cost, and what would have to change to prevent it?
The problem In a qualitative study of 73 participants filling in a fictitious tax return, a GPT-based conversational agent occasionally gave incorrect but plausible answers, even though the researchers tried to prevent it.
Answer in the threadHow do we know? What can this kind of evidence show, and what can’t it?
The evidence This rests on interviews or another qualitative study with a sample of 73, from International Journal of Public Sector Management.
Answer in the threadWhat does it leave open? What else would you want to know before acting on it?
The record’s own caveat The study analysed a fictitious tax return in an experiment with 73 participants, and the hallucinations were reported as rare, so the findings may not reflect real filings or real public-service use.
Answer in the threadHave you seen AI used this way where you study, work or live? Did it go the same way?
Answer in the thread
The questions are the same on every record. The details under them are taken from this record as written.
What readers say