Technical Summary: Evaluating Human and LLM-Generated Thematic Analysis in HRI for Vulnerable Populations

Problem Statement

Thematic Analysis (TA) is traditionally viewed as an inherently human, reflexive, and interpretive process requiring contextual judgment and ethical sensibility. While Large Language Models (LLMs) have shown promise in automating coding and theme generation, their applicability in sensitive domains—specifically Human-Robot Interaction (HRI) research involving vulnerable populations (e.g., people with disabilities)—remains largely unexamined. Existing comparative studies primarily rely on aggregate inter-rater agreement metrics, failing to systematically investigate whether discrepancies between human and LLM analyses are random or reflect systematic interpretive patterns with ethical consequences. There is a critical gap in understanding whether LLM-generated TA risks marginalizing or misrepresenting the lived experiences of vulnerable participants, particularly regarding the directionality of disagreements and their ethical significance.

Methodology

This study presents a comparative analysis of human- and LLM-generated TA applied to interview data from a within-subjects HRI user study involving 31 university students with diverse disabilities (including Autism, Specific Learning Differences, and Mental Health conditions). The study employed a 2×2 design crossing support roles (Signposting vs. Sounding Board) with agent embodiments (Pepper robot vs. Amazon Echo Dot).

The methodology involved three distinct analytical phases:

1. Human Thematic Analysis: A researcher with expertise in disability and neurodivergence conducted a reflexive TA following Braun and Clarke's six-phase framework. The process involved familiarization, inductive coding of salient segments, and the construction of candidate themes, which were refined to reflect participant perspectives.
2. LLM Thematic Analysis: The same six-phase framework was adapted for an LLM (Claude Sonnet 4.6 with extended thinking). The model was provided with domain-specific context (study design, participant demographics, analytic focus) and processed the data in batches to manage token loads. Codes from all batches were pooled and normalized to generate a consolidated set of themes.
3. Comparative Evaluation:
   • Clustering Consistency (RQ1): Adjusted Mutual Information (AMI) was used to quantify the agreement between human and LLM groupings of utterances at the theme and subtheme levels, correcting for chance.
   • Semantic Similarity (RQ2): Pretrained sentence embeddings (all-mpnet-base-v2) were used to compute cosine similarity between human and LLM labels for codes, subthemes, and themes.
   • Qualitative Ethical Analysis (RQ3): A qualitative evaluation of selected utterances (highest and lowest semantic similarity) was conducted using three criteria: Interpretive Grounding (depth of engagement vs. paraphrase), Representational Fidelity (faithfulness to participant voice), and Bias and Marginalisation Risk (potential for silencing or misrepresentation).

Key Results

Quantitative Findings

• Clustering Consistency: The LLM produced clusterings that aligned with human clusters significantly better than chance. The observed AMI was 0.231 at the theme level and 0.225 at the subtheme level. However, alignment was uneven; while "sensitivity to disability needs" and "effort requirements" showed strong correspondence with specific LLM themes, the human theme "power dynamics" was distributed across multiple LLM categories, indicating weaker mapping for conceptually diffuse themes.
• Semantic Similarity: Semantic alignment varied across levels and themes. At the code level, the highest consistency was observed for utterances associated with "sensitivity to disability needs." At the subtheme level, the highest consistency was observed for "effort requirements," with closely comparable levels found in "fluency of interaction" and "social attributes." At the theme level, the strongest alignment was found for "sensitivity to disability needs" and "fluency of interaction." In contrast, the lowest semantic similarity was observed for "power dynamics" and "effort requirements." Notably, subthemes sometimes performed worse than themes, likely reflecting the model's difficulty at an intermediate level of abstraction where it must coherently group codes without the flexibility of broader themes.

Qualitative Findings

The qualitative analysis revealed that disagreements were systematic rather than random, often involving shifts in interpretation and abstraction:

• Interpretive Grounding: LLM codes frequently functioned as paraphrases or extractions of lexically salient elements rather than genuine interpretive labels. For example, in cases of low similarity, the LLM failed to abstract the overall meaning, instead focusing on discrete phrases.
• Representational Fidelity and Bias: The study identified risks of "semantic sanitization" and generalization. In one instance, the LLM replaced the participant's specific term "ableism" with the broader "discrimination," flattening a disability-specific account of harm. In another, the LLM reframed a reflection on relational experience as a functional utility statement, prioritizing task-oriented framings over affective ones.
• Marginalisation Risk: These systematic shifts pose ethical risks by potentially silencing specific participant experiences or imposing hegemonic framings that generalize individual lived experiences into population-level categories.

Key Contributions

• First Comparative Analysis in HRI/Vulnerable Contexts: This is the first study to investigate the feasibility and appropriateness of LLM-based TA specifically within HRI research involving vulnerable populations.
• Beyond Agreement Metrics: The paper moves beyond simple inter-rater agreement to analyze the character of disagreements, demonstrating that systematic divergences carry ethical weight regarding whose experiences are centered.
• Ethical Evaluation Framework: The study introduces a qualitative evaluation framework (Interpretive Grounding, Representational Fidelity, Bias/Marginalisation Risk) to assess the ethical implications of LLM-assisted analysis in sensitive domains.
• Empirical Evidence of Limitations: The findings provide empirical evidence that while LLMs can support lower-level analytical tasks (e.g., initial coding, familiarization), their reliability decreases as the level of abstraction and interpretive ambiguity increases, particularly for themes involving identity, power, and lived experience.

Significance and Claims

The paper claims that LLM-assisted thematic analysis offers a bounded role in HRI research. While LLMs can effectively support early-stage analysis, the study argues that human-in-the-loop mechanisms remain indispensable for higher-level interpretations, particularly when dealing with vulnerable populations. The authors assert that the risks of misrepresentation, flattening of lived experience, and amplification of normative biases are heightened in these contexts. Consequently, researchers employing LLM-assisted TA must exercise reflexive oversight to ensure that findings do not systematically marginalize the experiences of the participants they aim to study. The study concludes that the observed patterns likely extend to other vulnerable populations, suggesting a need for hybrid human-AI workflows where human researchers review and refine LLM-generated outputs to mitigate ethical risks.