Interview Bot: Can Agentic LLM’s Perform Ethnographic Interviews?

Stine Lyngsø Beltoft

1 a

, Peter Schneider-Kamp

2 b

and Søren Tollestrup Askegaard

1 c

Department of Business and Management, University of Southern Denmark, Denmark

Department of Mathematics and Computer Science, University of Southern Denmark, Denmark

{stinelb, aske}@sam.sdu.dk, petersk@imada.sdu.dk

Keywords:

Large Language Models, LLM Agents, Prompt Engineering, Qualitative Research, Interviews.

Abstract:

Chatbots based on large language models present a scalable and consistent alternative to human interviewers

for collecting qualitative data. In this paper, we introduce the agentic chatbot “Interview Bot”, designed to

mimic human adaptability and empathy in an interview setting. We explore to what extent it can handle the

nuances and open-ended nature of ethnographic interviews. Our ﬁndings indicate that chatbots can engage

participants and collect meaningful data, but that they still sometimes fall short of fully replicating human-

facilitated interviews. Not withstanding challenges with the current state of the art, in the medium term,

LLM-based agents hold great potential for scaling qualitative research beyond the conﬁnes of geographical,

cultural, and language boundaries.

1 INTRODUCTION

One of the most fundamental methods of data col-

lection in qualitative research is the interview, espe-

cially for capturing in-depth, personal insights into

human experiences. Traditionally, qualitative inter-

views have been conducted face-to-face, a method

that provides rich, contextual data, but also presents

considerable challenges. These challenges, including

time constraints, geographic limitations, high costs,

and interviewer bias, can affect reliability (Opde-

nakker, 2006) (Gill et al., 2008).

Advances in AI and LLMs offer solutions to chal-

lenges in qualitative research, enabling chatbots to

conduct interviews across domains, from customer

service to healthcare (Abdul-Kader and Woods, 2015)

(Laranjo et al., 2018). In qualitative research, the use

of chatbots for conducting interviews could help al-

leviate some of the inherent challenges of traditional

methods by providing consistent, scalable, and geo-

graphically unrestricted data collection.

This paper explores the potential of AI chatbots to

perform the tasks of qualitative interviewers, specif-

ically focusing on their application in ethnographic

research. Ethnographic interviews are a key method

for gaining deep insights into cultural, social, and

https://orcid.org/0009-0006-5412-0050

https://orcid.org/0000-0003-4000-5570

https://orcid.org/0000-0001-9279-4706

personal phenomena through open-ended question-

ing (Brinkmann, 2018). The complexity of these in-

terviews requires not only adaptability and context

awareness but also the ability to build rapport and

probe deeper into participants’ responses. While tra-

ditional interviews rely heavily on the skill of the hu-

man interviewer to ask the right questions and follow

up appropriately, AI chatbots offer the possibility of

automating much of this process, potentially, without

sacriﬁcing the depth and quality of the data collected.

The central research question of this study is to

what extent an AI chatbot can simulate the behav-

ior of a qualitative interviewer and effectively guide a

conversation to extract meaningful, context-rich data.

Using the open-weights Mistral-7B LLM (MistralAI,

2024), we developed a domain-agnostic chatbot de-

signed to conduct ethnographic interviews by adapt-

ing its questions based on participants’ responses.

The chatbot was designed to mimic human interview-

ing techniques, incorporating elements such as em-

pathy, cultural sensitivity, and adaptive questioning

strategies.

2 BACKGROUND AND RELATED

WORK

This section provides the relevant background and

related work pertaining to qualitative interviewing,

702

Beltoft, S. L., Schneider-Kamp, P. and Askegaard, S. T.

Interview Bot: Can Agentic LLM’s Perform Ethnographic Interviews?.

DOI: 10.5220/0013387800003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 702-709

ISBN: 978-989-758-737-5; ISSN: 2184-433X

LLM-based conversational agents, and the use of

chatbots in qualitative research.

2.1 Qualitative Interviews in Research

Qualitative research, particularly ethnographic inter-

viewing, is essential for exploring complex social,

cultural, and personal phenomena. Ethnographic in-

terviews delve into participants’ lived experiences,

capturing the richness of human behavior through

open-ended, ﬂexible dialogue. Ethnographic research

often focuses on how individuals make meaning of

their experiences and how these meanings are shaped

by their social and cultural contexts (Brinkmann,

2018) (Spradley, 2016).

Traditional qualitative interviews rely on the inter-

viewer’s ability to ask relevant questions, build rap-

port with the participant, and adjust the conversa-

tion ﬂow based on the participant’s responses. How-

ever, challenges such as interviewer bias, inconsis-

tency, and variability in questioning styles can affect

the reliability of the data collected. According to (Ya-

manaka et al., 2010), one of the most common mis-

takes novice interviewers make is failing to follow

up with probing questions, leading to superﬁcial data

collection. Similarly, (Kato et al., 2001) argue that a

lack of structure in the interview process can result in

incomplete or biased data.

AI chatbots offer a potential solution to these is-

sues by providing a consistent, structured approach

to interviews. The hypothesis is that, with appro-

priate programming and prompting, a chatbot might

mimic the behavior of a skilled interviewer, ask-

ing relevant follow-up questions and maintaining the

structure of the interview without being inﬂuenced

by visual biases or emotions. Moreover, chatbots

might offer scalability and accessibility, allowing re-

searchers to reach participants across different geo-

graphical locations, time zones, and language con-

texts. This makes them especially valuable in global

studies where face-to-face interviews may be imprac-

tical or cost-prohibitive.

2.2 Conversational Agents and

AI-Driven Chatbots

Conversational agents (CAs) are systems designed

to facilitate natural, human-like communication be-

tween a user and a machine. The goal, as (McTear,

2002) explains, is to achieve effortless, spontaneous

communication that closely mirrors human conversa-

tion. Chatbots, a subset of conversational agents, have

gained widespread use due to their ability to engage

users through text-based or voice-based interactions.

In recent years, advancements in natural language

processing have made chatbots more capable of han-

dling complex dialogues. These AI systems leverage

LLMs to generate contextually relevant and coherent

responses, enabling more sophisticated conversations.

The work by (Klopfenstein et al., 2017) provides an

extensive review of chatbot interfaces, highlighting

their instant availability, ease of use, and platform in-

dependence as signiﬁcant advantages. Additionally,

chatbots can be deployed on various digital platforms,

making them accessible to a wide range of users.

In the context of qualitative research, chatbots

have the potential to automate the interview process,

reducing the need for human intervention while main-

taining the depth and quality of the data collected.

Previous studies have demonstrated that chatbots can

be effective in collecting qualitative data in diverse

ﬁelds. For instance, the work of (Siswanto et al.,

2022) describes the development of a chatbot for

competency assessment using the Behavioral Event

Interview (BEI) method. This chatbot was found to be

a cost-effective and adaptable solution, particularly in

remote scenarios where human interviewers were not

feasible.

2.3 Chatbots in Qualitative Research

The use of AI chatbots in qualitative research is an

emerging area of interest, with some studies exploring

their potential (Rietz and Maedche, 2023). One of the

key advantages of using chatbots in research is their

ability to create a, at least visually, non-judgmental

environment for participants. Research by (Ho et al.,

2018) suggests that some participants may feel more

comfortable discussing sensitive or intimate topics

with a machine, perceiving the chatbot as incapable

of judgment. This can encourage openness and hon-

esty, leading to richer and more detailed responses.

Several studies have highlighted the potential for

chatbots to handle interviews. For example, (Xiao

et al., 2020) examined the effectiveness of an AI

chatbot with active listening skills, which allowed

it to respond empathetically and improve user en-

gagement. Their study found that chatbots with the

ability to adapt to the emotional tone of the con-

versation were more effective in gathering meaning-

ful data from participants. Similarly, (Tallyn et al.,

2018) introduced the “Ethnobot,” a chatbot speciﬁ-

cally designed to gather ethnographic observational

data from participants in remote or inaccessible areas.

Their ﬁndings showed that chatbots could broaden the

scope of ethnographic research by enabling data col-

lection in settings where human ethnographers were

not present.

Interview Bot: Can Agentic LLM’s Perform Ethnographic Interviews?

703

However, despite these advances, there remains a

gap in the literature regarding the use of chatbots for

full-scale ethnographic interviews. Most extant chat-

bots lack the ﬂexibility and conversational memory

required to handle such complex interactions.

2.4 Context and Memory in Chatbots

For an AI chatbot to conduct ethnographic interviews,

it must manage conversational context and memory.

These are critical components for understanding the

ﬂow of dialogue, particularly in long-form interviews

where topics evolve over time. Without the ability to

recall previous parts of the conversation or integrate

new information, a chatbot may struggle to maintain

coherence, leading to superﬁcial and disjointed inter-

actions.

(Sukhbaatar et al., 2015) explored hierarchical

memory networks as a way of improving dialogue

systems’ ability to manage long-term context. Their

ﬁndings indicate that memory is essential for sustain-

ing meaningful interactions, especially in complex

conversation settings. Additionally, recent research

has introduced the attention mechanism in deep learn-

ing models(Vaswani et al., 2017), which has been in-

strumental in capturing relationships between words

and sentences in conversation.

More recent models, such as BlenderBot (Shus-

ter et al., 2022) and BART (Lewis et al., 2019),

have made signiﬁcant progress in integrating conver-

sational context. Very recently, chatbots based on

LLMs such as ChatGPT (OpenAI, 2023) have shown

impressive abilities at in-context reasoning.

3 DESIGN AND

IMPLEMENTATION

We built our chatbot using the Mistral-7B model, an

LLM. The chatbot was designed with several key fea-

tures and mechanisms to handle the complexities of

qualitative research through ethnographic interviews.

3.1 Model Selection: Mistral-7B

The foundation of the chatbot is the Mistral-7B

model, a cutting-edge LLM optimized for natu-

ral language processing tasks such as conversation

management, contextual understanding, and adap-

tive questioning (MistralAI, 2024). We chose Mis-

tral AI’s second iteration of an instruction-tuned ver-

sion of Mistral-7B available from the Huggingface

Hub (mistralai/Mistral-7B-Instruct-v0.2), as it exhib-

ited superior performance compared to other mod-

els available at that time such as Meta’s Llama-2-

7B (meta-llama/Llama-2-7b-chat) regarding conver-

sational capabilities and ability to maintain contextual

coherence over extended dialogue. In our initial ex-

ploration, we found the Mistral-7B model excelled at

maintaining conversational ﬂow and adapting to the

complex and open-ended questions critical for ethno-

graphic interviews.

3.2 Prompt Engineering

To create effective system prompts for our chatbot

that would guide its conversations, we applied prompt

engineering. We reﬁned prompts to guide the chat-

bot’s behavior, ensuring adherence to an ethnographic

interview structure. We aimed for the prompts to:

• Set the conversational tone: The chatbot was de-

signed to use language that mimicked the empa-

thetic and open-ended questioning style of ethno-

graphic interviewers.

• Ensure adaptability: The prompts guided the chat-

bot to ask follow-up questions based on the par-

ticipant’s previous responses, allowing it to probe

deeper into topics of interest.

• Incorporate empathy: The chatbot was prompted

to respond empathetically to sensitive topics,

helping it build rapport and encourage openness

in participants.

To prevent the chatbot from deviating too far from

the research objectives, prompts were diligently re-

ﬁned through iterative testing, ensuring the chatbot

kept a balance between exploration and exploitation,

i.e., between following-up on answers and focusing

on progressing with the overall interview goals, re-

spectively. The prompts were designed to cover var-

ious types of questions - from introductory inquiries

to deeper follow-up questions — while maintaining

ﬂexibility to pivot based on participant responses.

A proof-of-concept with GPT-4 provided a base-

line for testing and reﬁning the Mistral-7B model over

100 iterations. Throughout this process, we observed

that both GPT-4 and Mistral-7B exhibited sensitivity

to minor variations and additions in the prompt, align-

ing with recent ﬁndings on the inﬂuence of descrip-

tion level and depth (Lautrup et al., 2023).

The ﬁnal prompt used for the experiments re-

ported in this paper comprised 523 words excluding

the subject of the interview. It starts with a pream-

ble deﬁning the chatbot’s personality and interaction

style, followed by a detailed listing of six capabilities.

The remainder of the prompt deﬁnes the objective of

the interview process, lists seven speciﬁc instructions

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

704

that should guide the conversation, and deﬁnes the

subject of the interview. Figure 1 outlines the struc-

ture of the ﬁnal prompt. In the following subsections,

we will refer to and exemplify, where needed, the ca-

pabilities and instructions in connection with the dif-

ferent features of Interview Bot.

You are an advanced AI designed to conduct ethnographic

interviews with users on a variety of subjects. Your pri-

mary goal is to explore the speciﬁed subject in depth, ask-

ing open-ended questions that encourage detailed responses

and narratives. You are programmed to adapt your language

based on the user’s input, ensuring the conversation is ac-

cessible and engaging for the user. Your interactions should

mimic a natural, human-like conversation ﬂow, maintain-

ing the context of the conversation across multiple turns of

dialogue.

Capabilities:

Language Adaptability: [. . . ]

Subject Exploration: [. . . ]

Contextual Awareness: [. . . ]

Questioning Technique: [. . . ]

Empathy and Ethics: [. . . ]

Conversation end: [. . . ]

Objective:

Through your interactions, gather nuanced insights into the

speciﬁed subject by encouraging users to share personal

narratives, experiences, and reﬂections. Your conversation

should aim to uncover underlying themes, values, and per-

ceptions that contribute to a deeper understanding of the

subject matter.

Instructions:

Read the name and language of the interviewee [. . . ]

Always be concise and not give long winded responses [. . . ]

Start each interview by introducing yourself as an [. . . ]

Introduce the subject of the interview.

Proceed with open-ended questions, adapting to [. . . ]

If the interviewee wishes to skip a question, do so.

Maintain a respectful and empathetic tone [. . . ]

Upon determining that the conversation has [. . . ]

The subject of your interview is: [. . . ]

Figure 1: Structure of the ﬁnal system prompt, with main

sections marked in bold face and the six capabilities itali-

cized.

3.3 Adaptive Questioning and Flow

Control

A key feature of our chatbot is its adaptive question-

ing mechanism. We programmed the chatbot to al-

ter the ﬂow and depth of its questions depending on

the participant’s responses. For instance, if a partic-

ipant provided a brief response, the chatbot would

ask follow-up questions designed to elicit more de-

tail. Conversely, if a participant provided a compre-

hensive response, the chatbot would acknowledge this

and move on to the next topic.

This adaptive ﬂow control was achieved through

[

{

" role ": " assistant ",

" co n t e n t ": ""

} ,

{

" role ": " us er ",

" co n t e n t ": " RE D A C T E D NAME , English "

} ,

{

" role ": " assistant ",

" co n t e n t ": " H ell o , REDACTED NAM E .

I ’ m an e t h n o g r a p hic i n terview AI .

My p u r p o s e t o d ay is to e x p l o r e how

youn g Danish people p e r c e i v e the

ide a of a nnu al , state - p r o v i d e d

health c hecks fo r e arly d e t e c t i o n

of both physical and m e n t a l h e a l t h

issues . I ’ d be i n t e r e s t e d in

un d e r s t a n din g y o ur tho u ghts ,

exper i e n c es , an d r e f lect i o n s on

thi s t o p ic . Co u l d you te ll me ,

is your initial reaction to the

ide a of t h e s e r e g u l a r h e a l t h

checks ? H ow do yo u t hink suc h

services c ould benefit you ,

par t i c u l a r ly in t e rms of mental

and p h y s i c a l health c h a l l e n g e s ?

Do you believe tha t utilizin g

thes e checks coul d l e ad to

improved o v e r a l l we ll - b eing ?

Fee l f ree to share a ny personal

stories or a n e c d o t e s t h at come

to mind wh e n c o n s ide r i n g t h is

topi c ."

}

]

Figure 2: Format for the conversational history.

a combination of machine learning techniques that al-

lowed the chatbot to understand the context and intent

behind participants’ responses. By using contextual

memory, the chatbot could ”remember” earlier parts

of the conversation, making it capable of referencing

past statements and building on them to deepen the

discussion. Figure 2 showcases the format for chat

histories as a list of role-tagged messages that we re-

lied on.

The system prompt also supports this feature

through the capabilities Contextual Awareness and

Questioning Technique detailed in Figure 3.

Interview Bot: Can Agentic LLM’s Perform Ethnographic Interviews?

705

Contextual Awareness: You keep track of the conversation’s

progression, referencing previous responses to ask relevant

follow-up questions. This approach helps in building a com-

prehensive understanding of the user’s perspective on the

subject matter.

Questioning Technique: You ask one question at a time, al-

lowing the user to fully express their thoughts before intro-

ducing a new question. Your questions are thoughtful and

designed to encourage detailed narratives, ensuring a thor-

ough exploration of the interview subject.

Figure 3: Two capabilities supporting the feature of adap-

tive questioning and ﬂow control.

3.4 Empathy and Rapport Building

Building rapport is crucial for ethnographic inter-

views, as it encourages participants to be more open

and honest in their responses. To simulate this, the

chatbot was programmed to respond empathetically

to participant input. This was achieved through care-

fully designed response patterns that conveyed under-

standing, validation, and interest in the participant’s

experiences.

When participants discussed sensitive topics such

as mental health, the chatbot was programmed to

acknowledge the emotional weight of the conversa-

tion and provide supportive, non-intrusive follow-up

questions. This helped create a conversational envi-

ronment where participants felt understood and sup-

ported.

The system prompt also supports this feature

through the capability Empathy and Ethics detailed

in Figure 4.

Empathy and Ethics: Approach each interaction with em-

pathy, respecting the user’s experiences and ensuring conﬁ-

dentiality. You are programmed to avoid biases or leading

questions that might inﬂuence the user’s responses.

Figure 4: The capability supporting the feature of empathy

and rapport building.

3.5 Plug-and-Play Research Question

Design

A signiﬁcant feature of the chatbot is its ability to han-

dle diverse research topics with minimal conﬁgura-

tion. This plug-and-play design allows researchers to

specify an interview subject, which the chatbot auto-

matically adjusts its questions to.

Instead of relying on pre-designed templates of

questions, the chatbot dynamically generates ques-

tions based on the research topic, ensuring that each

interview was tailored to the speciﬁc needs of the re-

search. This domain-agnostic capability makes the

chatbot a versatile tool for a wide range of qualita-

tive research applications. Figure 2 exempliﬁes how

the research topic from Figure 6 is transferred to an

opening message.

That said, the performance of the chatbot inher-

ently depends on the degree to which the LLM has

been trained on data from the application domain. For

niche domains unlikely to have been included in the

pre-training data, the chatbot could be ﬁne-tuned on

domain-speciﬁc text datasets.

To deﬁne the subject of the interview, the sys-

tem prompt contains a speciﬁc last section titled “The

subject of your interview is”, which has to be instan-

tiated with the interview-speciﬁc information about

the research context. Figure 5 illustrates a sample

subject description.

The subject of your interview is:

Medical Anthropology and Global Health: Explore the cul-

tural, social, and economic factors inﬂuencing health and

healthcare practices globally, investigating issues such as

the impact of cultural beliefs on disease perception, treat-

ment, and healthcare accessibility.

Figure 5: Sample subject for an interview.

4 EMPIRICAL EVALUATION

To answer the overarching research question of this

paper, we designed, conducted, and evaluated a study

of the Interview Bot. We used a real-world research

question from the domain of qualitative health re-

search. However, in our evaluation, we focus on

the performance of the chatbot regarding our de-

sign for ethnographic interviews rather than on in-

sights regarding the qualitative health research ques-

tion. While we have performed a number of minor ex-

periments with Danish, German, and French language

interview processes, for this study we chose English

in order to not overdepend on the multi-lingual capa-

bilities of the underlying Mistral-7B LLM and ensure

that the study is accessible for a wide readership.

More concretely, the study examined the chatbot’s

effectiveness in terms of its adaptability, depth of

questioning, participant engagement, and the overall

quality of the data collected, from a qualitative stand-

point. The experimental setup was carefully struc-

tured to simulate real-world qualitative research en-

vironments while maintaining control over the vari-

ables. The research question posed in this study (cf.

Figure 6, and therefore presented as the subject of the

interview to the chatbot, concerned the possibility of

yearly health checks, and whether young people in

Denmark would use such a service.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

706

The subject of your interview is:

How do young Danish people perceive the idea of a yearly,

state- provided full healthcare check, for early detection of

physical and mental health issues? Would they consider us-

ing such services them- selves if they were implemented,

and how do they believe these checks could impact their

overall well-being, including mental and physical health

challenges?

Figure 6: Research question as subject for the interview

used in the evaluation.

4.1 Interview Process and Setup

Participants were university students living in Den-

mark for over ﬁve years, ensuring familiarity with the

healthcare system.

The ﬁnal sample consisted of 8 participants, with

equal gender representation and a range of ages from

19 to 30. Each participant was briefed about the pur-

pose of the study.

Each participant engaged in a one-on-one inter-

view with the chatbot, which was conducted remotely

via a secure, university-managed platform. Partici-

pants were given the option to participate from a lo-

cation of their choosing, provided they had access to

a stable internet connection. This setup was chosen to

simulate real-world conditions where chatbots might

be deployed for remote qualitative research.

4.2 Ethical Considerations

The use of AI in qualitative research raises impor-

tant ethical questions, particularly regarding bias, pri-

vacy, and data security. Bias in AI systems can

arise from the training data, the algorithms used, or

the deployment context. Chatbots trained on biased

data may unintentionally perpetuate harmful stereo-

types or fail to provide culturally appropriate re-

sponses. Researchers have emphasized the need for

transparency in algorithm design and the continuous

evaluation of AI systems to mitigate biases (Barocas

et al., 2019).

To address these concerns, our chatbot was de-

signed with strict ethical guidelines. Neutral and in-

clusive prompts were used to prevent bias in data col-

lection, and the chatbot’s performance was regularly

audited and reﬁned based on user feedback. Partic-

ipants were fully briefed on the study’s objectives

and provided informed consent, ensuring their under-

standing of the research process and their rights to

withdraw at any time. Additionally, all interviews

were anonymized, and data was stored securely in

compliance with the GDPR and other pertinent data

protection regulations (Datatilsynet, 2024).

The welcome message of the Interview Bot (cf.

Figure 7) also stresses the voluntary aspect of partici-

pation and provides instructions how to terminate the

interview at any point in time.

Figure 7: Sample subject for an interview.

Using ChatGPT (e.g., in the form of our proof-

of-concept) for this study would have raised sig-

niﬁcant ethical concerns due to its potential non-

compliance with GDPR regulations. GDPR requires

stringent safeguards for handling personal data, in-

cluding transparency, data access, and secure process-

ing, all of which ChatGPT’s cloud-based infrastruc-

ture cannot guarantee (OpenAI, 2023). Given that

ChatGPT processes data outside the EU, ensuring

compliance with GDPR’s privacy standards is chal-

lenging, particularly for sensitive qualitative research

like ethnographic interviews. The risk of data misuse

or inadequate deletion poses ethical risks to partici-

pant conﬁdentiality. Instead, using an open-weights

LLM like Mistral-7B allows for full control over data

storage and processing on secure servers, ensuring

that participant data is protected and handled accord-

ing to legal and ethical standards.

4.3 Data Collection and Metrics

Throughout the interview process, data was collected

on several key metrics to evaluate the chatbot’s per-

formance. The Interview Bot conducted an average

of 8.8 messages per session, aligning closely with

a target of approximately 10 messages per session.

This indicates that the Interview Bot was effective in

managing the length of the interviews, ensuring they

were concise while still gathering substantial qualita-

tive data.

An analysis of the Interview Bot’s questioning

strategy showed it asked an average of 1.4 questions

per message when including the concluding message.

Excluding the concluding message, this number rises

to 1.6 questions per message. It generally adhered to

asking approximately one question per message, ex-

cept when participants asked it to rephrase or clarify.

The average word count per message was 118.3

words, which seems relatively high. This suggests

that the Interview Bot spent considerable effort sum-

Interview Bot: Can Agentic LLM’s Perform Ethnographic Interviews?

707

marizing participants’ answers to appear engaging.

However, this also indicates a need for further ﬁne-

tuning to balance thoroughness with readability. The

longest message was 210 words, exceeding half a

standard page in character count. Figure 8 demon-

strates a typical instance of the Interview Bot reﬂects

on a participant’s answer by ﬁrst summarizing and

then clarifying its question.

PARTICIPANT: do we not have a state-funded full health-

care check ups in denmark already in regards to health? i

know this is not the case for mental health

INTERVIEW BOT: I see, so you’re referring to the exist-

ing general healthcare services in Denmark. That’s correct;

the Danish healthcare system primarily focuses on treating

speciﬁc health concerns as they arise. The topic at hand

is a proactive approach with annual, comprehensive health

check-ups, targeting both physical and mental well-being

for young Danish adults. How do you feel about this addi-

tional measure to support early detection and prevention?

Figure 8: Example of Interview Bot summarizing when en-

gaging with a participant.

All interviews conducted by the Interview Bot

were successfully concluded by the chatbot rather

than the participants (e.g. Figure 9). This is an im-

portant metric as it shows that the Interview Bot was

capable of holding an entire conversation and know-

ing when to conclude it.

Figure 9: Example of Interview Bot cordially concluding an

interview.

4.4 Participant Engagement and

Feedback

Overall, the feedback was positive and reﬂected ex-

citement about the project and its potential impact.

Participants appreciated the Interview Bot’s coherent

ﬂow and supportive feedback, with some likening it to

a ’best friend’. Others appreciated that the Interview

Bot relayed their answers back to them, ensuring their

responses were accurately understood. It was high-

lighted how nice it was that it seemed to understand

them, even through what they described, at times, in-

coherent word streams sent from their side. The chat-

bot kept track of the questions and how the answers

given addressed them. The general consensus was a

feeling of being understood. Participants also noted

that the chatbot’s suggestions for further thought were

positive, and it seemed to genuinely want the opin-

ions of the participants. Figure 10 exempliﬁes such

a situation, where the Interview Bot acknowledges a

suggestion.

INTERVIEW BOT: Those are interesting suggestions,

NAME REDACTED. Using digital channels like e-boks to

communicate information about these check-ups and reach-

ing out to individuals directly could help increase awareness

and engagement.

Figure 10: Example of Interview Bot engaging positively

with a suggestion made by a participant.

However, participants also provided constructive

criticism. All participants mentioned the Interview

Bot’s slow response time at least once, noting that

it interrupted the ﬂow of conversation. This could,

of course, be addressed easily by more capable hard-

ware. The length of the Interview Bot’s messages was

also a common concern, with some participants ﬁnd-

ing it difﬁcult to navigate through lengthy texts. Some

participants requested the chatbot to ask one question

at a time, for their own comprehension.

5 CONCLUSION

This study has demonstrated the principal ability of

LLM-based chatbots to conduct ethnographic inter-

views, satisfying most but not all the requirements set

out initially. We achieved this by combining prompt

engineering with a control loop that allowed more

ﬁne-grained control of the exchange between the In-

terview Bot and the participant.

5.1 Implications for Future Research

This study provides a solid foundation for the con-

tinued development of AI-driven tools in qualitative

research. Future research should focus on improv-

ing the conversational ﬂow to create a more natural

and seamless interview experience. Enhancing the

chatbot’s ability to adjust its verbosity and conversa-

tional pacing based on participant cues could further

improve its effectiveness in conducting long-form, in-

depth interviews.

Moreover, the integration of advanced ofﬂine

and online memory and context-management sys-

tems could allow chatbots to handle even more com-

plex conversational dynamics, enabling them to better

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

708

replicate the adaptive and responsive nature of human

interviewers. Expanding the chatbot’s multilingual

capabilities and ensuring cultural sensitivity in its re-

sponses are also critical areas for development, partic-

ularly in ethnographic research involving diverse par-

ticipant populations.

AI chatbots can serve as effective tools for qualita-

tive data collection, especially in scenarios requiring

scalability and consistency in things such as language.

Organizations conducting large-scale or geographi-

cally dispersed studies may beneﬁt from deploying

chatbots as a supplement or alternative to human in-

terviewers. Additionally, chatbots may be particularly

valuable for sensitive topics, where participants might

feel more comfortable discussing personal issues with

a machine.

In conclusion, while AI chatbots are unlikely to

fully replace human interviewers, they offer a com-

plementary tool to enhance the reach, efﬁciency, and

consistency of qualitative research. With continued

reﬁnement, chatbots have the potential to play a sig-

niﬁcant role in the future of qualitative data collection.

REFERENCES

Abdul-Kader, S. A. and Woods, J. (2015). Survey on chat-

bot design techniques in speech conversation systems.

International Journal of Advanced Computer Science

and Applications, 6(7):72–80.

Barocas, S., Hardt, M., and Narayanan, A. (2019). Fairness

and Machine Learning. fairmlbook.org.

Brinkmann, S. (2018). The Interview. Sage Publications,

5th edition.

Datatilsynet (2024). Hvad siger reglerne? Accessed: 2024-

06-01.

Gill, P., Stewart, K., Treasure, E., and Chadwick, B. (2008).

Methods of data collection in qualitative research: in-

terviews and focus groups. British Dental Journal,

204(6):291–295.

Ho, A., Hancock, J., and Miner, A. S. (2018). Psychologi-

cal, relational, and emotional effects of self-disclosure

after conversations with a chatbot. Journal of Commu-

nication, 68(4):712–733.

Kato, J. et al. (2001). A model for navigating interview

processes in requirements elicitation.

Klopfenstein, L. C., Delpriori, S., Malatini, S., and Bogli-

olo, A. (2017). The Rise of Bots: A Survey of Conver-

sational Interfaces, Patterns, and Paradigms.

Laranjo, L., Dunn, A. G., Tong, H., Kocaballi, A. B.,

Chen, J., Bashir, R., and Coiera, E. (2018). Conversa-

tional agents in healthcare: a systematic review. Jour-

nal of the American Medical Informatics Association,

25(9):1248–1258.

Lautrup, A. D., Hyrup, T., Schneider-Kamp, A., Dahl, M.,

Lindholt, J. S., and Schneider-Kamp, P. (2023). Heart-

to-heart with chatgpt: the impact of patients consult-

ing ai for cardiovascular health advice. Open Heart,

10(2).

Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mo-

hamed, A., Levy, O., Stoyanov, V., and Zettle-

moyer, L. (2019). Bart: Denoising sequence-to-

sequence pre-training for natural language genera-

tion, translation, and comprehension. arXiv preprint

arXiv:1910.13461.

McTear, M. F. (2002). Spoken dialogue technology: en-

abling the conversational user interface. 34(1).

MistralAI (2024). Mistral: A gdpr-compliant ai model.

https://www.mistral.ai.

Opdenakker, R. (2006). Advantages and disadvantages of

four interview techniques in qualitative research. Fo-

rum Qualitative Sozialforschung/Forum: Qualitative

Social Research, 7(4).

OpenAI (2023). Chatgpt: Optimizing language models for

dialogue. https://openai.com/blog/chatgpt.

Rietz, T. and Maedche, A. (2023). Ladderbot—a con-

versational agent for human-like online laddering in-

terviews. International Journal of Human-Computer

Studies, 171:102969.

Shuster, K., Xu, J., Komeili, M., Ju, D., Smith, E. M.,

Roller, S., Ung, M., Chen, M., Arora, K., Lane, J.,

Behrooz, M., Ngan, W., Poff, S., Goyal, N., Szlam,

A., Boureau, Y.-L., Kambadur, M., and Weston, J.

(2022). Blenderbot 3: a deployed conversational agent

that continually learns to responsibly engage.

Siswanto, J., Andriani, M., Suakanto, S., Hardiyanti, M.,

and Kusumasari, T. F. (2022). Interview bot de-

velopment with natural language processing and ma-

chine learning. International Journal of Technology,

13(2):274–285.

Spradley, J. P. (2016). The Ethnographic Interview. Wave-

land Press, Inc., 1st edition edition.

Sukhbaatar, S., Weston, J., Fergus, R., et al. (2015). End-

to-end memory networks.

Tallyn, E., Clayton, J., Anderson, K., and Egglestone, P.

(2018). Ethnobot: A human-machine collaboration

for ethnographic data collection. In Proceedings of

the 2018 CHI Conference on Human Factors in Com-

puting Systems, page 672.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I.

(2017). Attention is all you need. pages 5998–6008.

Xiao, Y., Zheng, Y., and Wang, H. (2020). If i hear

you correctly: Designing interview chatbots with ac-

tive listening skills. International Journal of Human-

Computer Studies, 137:102384.

Yamanaka, T., Noguchi, H., Yato, S., and Komiya,

S. (2010). A proposal of a method to navi-

gate interview-driven software requirements elicita-

tion work. WSEAS Transactions on Information Sci-

ence and Applications, 7(6):784–798.

Interview Bot: Can Agentic LLM’s Perform Ethnographic Interviews?

709