'Simulacrum of Stories': Examining Large Language Models as Qualitative Research Participants

1 INTRODUCTION

• Proposals to replace human research participation with synthetic data cite goals such as speed, scale, lower costs, diversity, and participant protection.A key assumption is that models trained on vast datasets can closely mimic human-generated data.Source: §1 · p. 1
• Qualitative inquiry develops understandings of lived experience through contextual meaning-making, trust, rapport, and attention to power relations.It can surface subjective meanings excluded or silenced from public discourse.Source: §1 · p. 1
• The study used in-depth, semi-structured interviews with 19 qualitative researchers and a probe that let them recreate interviews from their own projects.Familiar projects were chosen to elicit nuanced reflections across HCI research topics.Source: §1 · p. 2
• Researchers identified six fundamental limitations, including absent palpability, ambiguous epistemic position, foreclosed consent and agency, erased community perspectives, and risks to qualitative ways of knowing.They initially found similar narratives in LLM outputs, but differences emerged over several conversational turns.Source: §1 · p. 2
• LLMs used as participant proxies can displace community voices with algorithmic approximations, raising ethical and epistemological concerns beyond technical model limitations.The paper frames this substitution through the surrogate effect described by Atanasoski and Vora.Source: §1 · p. 2

2. RELATED WORK

• LLMs have been explored for behavioral simulation, persona-based agents, research tasks, and qualitative data analysis, including brainstorming, annotation, surveys, and experiments.Believability is often assessed through natural interaction and consistency with designed character traits.Source: §2.1 · p. 2, §2.2 · p. 3

2.1 LLMs as proxies for human behavior

• Prior studies tested LLMs as proxies in controlled settings, whereas this work examines how qualitative researchers perceive generated data within qualitative inquiry.Argyle et al. [5] describe conditioning responses with sociodemographic backstories as achieving “algorithmic fidelity” in “silicon samples.”Source: §2.1 · p. 2

2.2 LLM use for research-related tasks

• Studies report mixed results: accuracy varies by socioeconomic status, political affiliation, and persona demographic identity, prediction performance diminishes without prior examples, and advanced LLMs produce unnaturally precise responses [63] [4].Source: §2.2 · p. 3
• Prior work reports that LLMs can misrepresent groups and erase within-group differences, while interviews require exchanges that support reflection, clarification, and elaboration.Wang et al. [109] argue these representational issues are unlikely to be resolved simply by newer model generations.Source: §2.2 · p. 3

2.3 Risks of harm posed by LLMs

• Risks discussed in prior work include biased or toxic training data, stereotyped outputs, misrepresentation of viewpoints, misinformation, anthropomorphism, appropriation, and environmental harms.LLMs may also exacerbate epistemic injustice by excluding some voices from knowledge production.Source: §2.3 · pp. 3–4

3 METHOD

• From March to June 2024, the researchers interviewed 19 qualitative researchers and used a hands-on probe to compare LLM-generated data with human-generated data.Participants brought anonymized transcripts from recent interview studies for reference.Source: §3 · p. 4
• The probe let researchers define an LLM participant through a system prompt, conduct simulated interviews, and review conversation histories; the study used GPT-4-turbo.The prompt template asked the model to respond as a participant and share lived experiences and anecdotes when appropriate; GPT-4-turbo was selected for performance as of February 2024 [120].Source: §3 · p. 4
• Each hands-on session lasted 30–45 minutes, with researchers adapting interview questions, thinking aloud, and comparing model responses with a human transcript.Interviews lasted up to 90 minutes overall.Source: §3 · p. 4
• Participants were experienced qualitative researchers recruited through social networks, email, institutional forums, and campus flyers; 16 worked in academia, two in industry, and one in nonprofits.All participants were located in the United States.Source: §3 · p. 5
• Researchers followed reflexive thematic analysis, coding interview transcripts and iteratively refining categories through team meetings; they synthesized five described categories into three top-level categories presented in Findings.They transcribed 1,611 minutes of video recordings and obtained 638 first-level codes; they did not apply inter-rater reliability measures.Source: §3 · p. 5
• The researchers obtained institutional review approval, protected participant identities, and described how their interpretivist orientation and team backgrounds shaped analysis.The sample was primarily academic, covered a limited range of topics, and used only GPT-4, limiting generalizability to other contexts and models.Source: §3 · pp. 5–6

4.1 Researcher perceptions of and interactions with LLMs

• Initially skeptical, researchers found LLM responses plausible and detailed, then saw how persona descriptions could prompt assumptions and essentialize identities.Some researchers viewed LLMs as tools for studying model behavior rather than human behavior; detailed outputs could offer volume without depth.Source: §4.1 · pp. 6–7

4.2 Fundamental Limitations of LLMs as Simulated Subjects in Qualitative Research

• LLM responses lacked the concrete context, spontaneity, and emotional depth that researchers draw on to interpret lived experience.Prompts could elicit stories, but researchers questioned their validity; this limitation was especially apparent for sensitive or personal topics.Source: §4.2 · pp. 7–8
• Persona and prompt control can amplify researcher bias, while researchers’ positionality shapes how they assess simulations and make meaning with communities.Researchers noted that outsiders may miss stereotypes, whereas in-group contextual knowledge can help identify harmful misrepresentation.Source: §4.2 · p. 8
• Simulated experiences that resemble researchers’ communities but contain uncomfortable inaccuracies could cause affective discomfort or harm, including microaggressions.Source: §4.2 · p. 9
• LLMs’ epistemic position is ambiguous: they can aggregate conflicting perspectives, shift with prompt wording, and reflect opaque, temporally limited training data.These conditions make it difficult to establish whose situated knowledge a response represents or evaluate its validity.Source: §4.2 · p. 9
• LLMs risk erasing community perspectives through stereotyped or sanitized responses that mimic identity without grounding claims in lived experience.Researchers noted that models could capture broad community contours while missing the nuanced realities of niche groups.Source: §4.2 · pp. 9–10
• Simulated participants rarely challenge researchers’ premises or introduce their own agendas, while generated responses can disclose sensitive material without consent.Researchers also questioned whether training data drawn from people’s stories had been obtained with their permission.Source: §4.2 · p. 10
• Researchers warned that using LLMs could further delegitimize qualitative research by suggesting contextual work can be replicated faster and encouraging data extraction over collaboration.They described qualitative inquiry as iterative and relational, including continued engagement with participants beyond data collection.Source: §4.2 · p. 10
• Replacing community perspectives with LLM outputs could erode trust, especially among communities already wary of being misrepresented.Researchers connected this risk to a history of extractive research and said algorithmic outputs can undermine the value of lived experience.Source: §4.2 · pp. 10–11
• LLMs may sound fluent and relevant, but researchers said they lack the depth, nuance, and authenticity of data grounded in lived realities.They also warned that technology-distorted perspectives could leave communities unheard by decision-makers.Source: §4.2 · p. 11

4.3 Imagining Use Cases for LLMs in Qualitative Research: Seeing Possibilities and Even More Caveats

• Researchers named possible LLM use cases but reached no consensus, and for each application they also identified potential harms or ineffectiveness.Most expressed discomfort with synthetic research data, and some said they would distrust findings based on LLM participants.Source: §4.3 · p. 11
• Some saw participant simulation as a lower-stakes teaching aid, but warned it could teach poor interviewing habits and cannot reproduce emotional management.LLM interaction may differ from human interviews; one researcher needed three prompts to elicit useful context, while risks include failing to build rapport or probe deeply.Source: §4.3 · p. 11
• LLMs could help test interview protocols when pilot recruitment is difficult, but may steer researchers toward questions or directions that real participants would not raise.Researchers suggested using simulations to gauge likely responses or refine questions, while cautioning that this can shift a project’s focus.Source: §4.3 · p. 11
• For sensitive topics, simulation might spare participants difficult retellings or help prepare for unsafe groups, yet risks erasing human experience and misrepresenting complex lives.Researchers suggested community collaboration to build expertise and prepare for difficult conversations rather than turning to an LLM.Source: §4.3 · p. 11

5 DISCUSSION

• The surrogate effect frames LLM participation as replacing community voices with algorithmic approximations while obscuring the human labor sustaining these systems.The discussion links that obscured labor to exploitation and unevenly distributed benefits in the global technology industry.Source: §5 · pp. 11–12
• LLM surrogacy can reshape knowledge and power: model biases may reinforce hierarchies while technology gains authority and vulnerable communities lose participation.The discussion presents this as a co-production of knowledge and social order.Source: §5 · p. 12

5.1 LLMs are incongruent with qualitative epistemologies

• Qualitative knowledge is co-constructed through researcher-participant engagement and relationships beyond a single interview, making LLM interviewing misaligned with these methods.The discussion describes interviews as intentional meaning-making rather than neutral extraction.Source: §5.1 · p. 12
• Qualitative accounts seek thick, culturally and materially situated description, whereas LLMs generalize patterns and flatten differences into decontextualized outputs.The discussion connects this tension to feminist epistemologies and a preference for efficiency and scalability over contextual engagement.Source: §5.1 · p. 12

5.2 LLM use undermines consent and autonomy of data subjects

• Using LLMs built from others’ work to generate research data risks undermining data subjects’ autonomy and exposing sensitive community accounts without explicit consent.The discussion extends consideration of voluntary consent to people whose data was used to produce the models.Source: §5.2 · p. 13

5.3 What does it mean to simulate the user in HCI?

• Researchers identified possible LLM roles in trialing materials, teaching, or unsafe contexts, but flagged misrepresentation and misleading training as risks, not endorsements.The authors say these considerations are not guidance to justify LLMs as simulated participants.Source: §5.3 · p. 13
• Researchers urge thoughtful model selection, transparency, and rigorous validation, while warning that disclosure may be performative and validation cannot resolve deeper epistemological questions.Western-centric training may fail to simulate non-Western experiences; member checking can demand time, energy, and emotional investment from community members.Source: §5.3 · p. 13

• Because research papers and products are already exploring LLMs as participant replacements, the authors urge caution, often a critical stance, in research on lived experiences.Source: §6 · pp. 13–14
• LLM stand-ins can foreclose participant consent and agency and erase community perspectives.Source: §6 · p. 14
• Their responses may lack palpability, obscure the model’s epistemic position, amplify researcher positionality, and risk undermining qualitative methods of knowing.Source: §6 · p. 14