The question of emotions in human-machine conversations is a key challenge in communication studies. Picard (1997) states that a machine cannot convincingly appear intelligent if it lacks the ability to recognize and respond to human emotions. Yet, current AI models are often trained on datasets that rely on outdated conceptualizations of emotion and are annotated with overly simplified taxonomies that reduce emotional expression to only a handful of emotions and to binary states that are either on or off. Equally, many datasets annotate emotions at the single-sentence level, which compresses the complexity of emotional expression and take it out of its dialogic structure in which it naturally occurs. This stands in a stark contrast to the importance of emotions for communication. Emotions shape decision making processes (Picard 1997) and influence how messages are conveyed and received across various contexts (Bolls 2010). There is a clear need for interdisciplinary collaboration and the engagement of both the social science and the informatics community in the construction of training datasets for emotion modelling.

Many emotion-annotated corpora rely on categorical models such as the ones of Ekman (2003) or Plutchik (1980), that contain 6-8 emotion categories. Although more fine-grained and statistically validated frameworks have emerged in recent years, such as Cowen and Keltner’s 27 emotion categories (2017), these approaches still capture only part of the complexity of emotional expression. Emotions are not binary states that are simply present or absent; rather, they unfold like waves that rise and fall over time, without necessarily returning to a stable or “neutral” baseline (Plantin 2020). Instead of discrete, static entities, emotions are dynamic processes, not only as they unfold within the individual, as described by Scherer’s Component Process Model (2009), but also how they emerge in interaction. Emotions emerge dynamically across turns, sequences, and situational cues (Peräkylä / Sorjonen 2012). However, most emotion-related text-based corpora rely on isolated utterances, single sentences or tweets, thereby removing emotional expression from the dialogic structure in which it naturally unfolds. This dialogical structure is particularly relevant for conversational agents, whose primary mode of interaction is turn-based, mimicking human to human communication. Moreover, emotion models do not simply represent an underlying emotional reality, they actively produce the reality they claim to describe. The co-creation process with generative AI requires concessions and adaptations on the user’s side, meaning that we also adopt unconscious implicit cultural, normative, or stylistic assumptions of the models (Garmon et al. 2025). Social constructionist theories of emotion emphasize that socialization plays a significant role in how we understand, categorize, and express emotions (Kemper 1987; Turner 1999). In this sense, AI systems participate in a form of emotional socialization: they learn from the categories we provide, and users, in turn, adjust to the emotional repertoire they make available.

As conversational agents such as ChatGPT become increasingly integrated into communicative practices, not merely as intermediaries, but as communicators and conversational partners (Haqqu / Rohmah 2024), it becomes essential to examine how emotions emerge within these exchanges. Our aim in this study is to analyze emotional expression in human-machine interaction by situating emotions within their dialogic structure and the context in which they are produced. To do so, this paper addresses the research question: how can emotional expression be represented, annotated, and measured in a way that reflects its social and dynamic properties?

To this end, we collaborated with informatics researchers to better understand the shortcomings of current datasets as well as the specific needs and challenges of model training. This includes both the technical perspective, concerning data structure and model requirements, and the social-scientific perspective, which concerns how emotions are conceptualized and measured. Building on these insights, we developed a chatbot that is used to collect human-machine dialogues containing traces of emotional expression

The chatbot is built on the open-source model Ollama and runs on our own servers, ensuring full control over data and independence from external updates.

. Participants are invited to discuss a personal emotionally meaningful episode, with full freedom over what they share to minimize potential distress. The aim of this approach is to analyze emotional expression within the dialogic structure in which it occurs. We aim for a sample of 500 participants, diverse in age and socioeconomic background. The initial study will be in French, as French corpora remain underrepresented in AI training. However, the research protocol is designed to be transferable to other languages and publics.

For analysis, we are developing a codebook designed to measure both which emotions are expressed and how they are expressed. Following an extensive review of emotion theories and models, the study adopts a combined approach for identifying expressed emotions: an extensive categorical model based on the work of Cowen and Keltner (2017), complemented by the valence-arousal-dominance model by Russell and Mehrabian (1997) that captures emotions within a three-dimensional space. This combined approach allows us to compare the two frameworks and evaluate their convergences and divergences. To address how emotions are expressed, Scherer’s Component Process Model (2009) will be adapted for the analysis of textual data. The components outlined by Scherer, such as evaluation of a stimulus event, physiological response, motivation to act, will be supplemented with emotion-specific linguistic profiles derived from widely used emotion datasets by our informatics colleagues (Lecourt et al. manuscript submitted). We will not only analyze single utterances but take into account how emotions evolve across the dialogue, identifying broader discourse patterns in the interaction.

Through this multi-layered analysis, the project aims to provide a multidimensional annotated dataset that shows not only which emotions are expressed but also how, providing a holistic view of emotions and situating them within the dialogic and relational dynamics in which they emerge. As generative AI systems increasingly participate in social processes like communication, understanding how users engage with AI systems becomes critical, not only for designing more sensitive and responsible AI tools, but also for understanding how these technologies contribute to creating the social and cultural conditions in which emotions are expressed and how they might reshape communication.