Emotion words may be easier to understand as points on a map than as isolated labels, according to a new study that used GPT-4 to reconstruct the semantic landscape of human feelings. Researchers at the Laboratory of Neurophysiology, National Institute for Basic Biology, asked the language model to arrange emotion terms according to their meanings and then compared its results with judgments made by human volunteers. The experiment suggests that a widely used artificial intelligence system can reproduce important features of how people organize emotional concepts—not because it experiences emotions, but because it has learned the ways humans connect emotion-related words through language.

The question is deceptively simple: How close are “joy” and “surprise,” and how far apart are “joy” and “sadness”? Answering it systematically becomes increasingly difficult as more words are added. If researchers want to compare 99 emotion terms pair by pair, they would need an enormous number of judgments, creating a heavy burden for human participants. To reduce that burden, the Japanese research team designed a spatial mapping task. Three emotion terms appeared simultaneously on a digital grid, and the participant—or GPT-4—placed them so that semantically similar words were near one another while dissimilar words were separated. A single arrangement captured three relationships at once, allowing the researchers to estimate a much larger semantic structure efficiently.

The first experiment focused on six basic emotion terms: joy, surprise, anger, fear, disgust, and sadness. Eighty-nine human participants completed the grid-based task, while GPT-4 was instructed to perform the same operation through natural-language prompting. The researchers also examined conventional word embeddings, mathematical representations in which words are converted into high-dimensional numerical vectors. In an embedding space, words that occur in similar linguistic contexts tend to occupy nearby locations. These models have been widely used to study meaning, but their geometric relationships do not always reproduce the way humans interpret concepts, especially when words have complex emotional or cultural associations.

GPT-4 and the human participants produced strikingly similar arrangements. Both placed joy and surprise in one group, while anger, fear, disgust, and sadness formed another. The distances between individual emotion terms generated by people and by GPT-4 were also strongly correlated. Human responses were consistent with one another, suggesting that the task was measuring broadly shared semantic relationships rather than highly personal emotional associations. Conventional word embeddings, however, generated a different pattern: they grouped joy and anger together, separating the words in a way that did not match the human arrangement. The result highlights a distinction between statistical word similarity and the richer conceptual structure people may infer from language.

That distinction is important because GPT-4 is not being presented as a machine that feels happiness, fear, or grief. The model has no evidence of subjective emotional experience in this experiment. Instead, it appears to use the relationships among emotion concepts encoded in its training data to produce a human-like arrangement. When prompted to treat words as objects in a semantic space, GPT-4 can draw on countless linguistic patterns, associations, descriptions, and contrasts found in human communication. In this sense, the model functions less like an emotional subject and more like a measurement instrument for the structure of concepts expressed through language.

The researchers then expanded the vocabulary from six terms to 99, a scale at which conventional human testing would become impractical. GPT-4 was prompted repeatedly to place different combinations of three emotion words on a grid, generating enough data to reconstruct a broader map. The resulting organization revealed two major patterns associated with pleasure and dominance. Pleasure refers to whether an emotion is experienced or described as pleasant or unpleasant, while dominance reflects whether a person feels in control of a situation or overwhelmed by it. The separation between the major clusters was clearer in the GPT-4 arrangements than in the corresponding word-embedding analysis, indicating that task-specific prompting may capture relationships that generic vector representations blur.

A finer layer of organization emerged inside those broad clusters. The 99-word map showed variation related to arousal, the degree of activation or excitement associated with an emotion. High-arousal terms such as excitement and rage were separated from lower-arousal terms such as serenity and sadness, even when the words shared the same general pleasantness or unpleasantness. This pattern suggests that arousal may not always serve as the main axis dividing the entire emotional vocabulary. Instead, it can act as a secondary dimension, organizing related emotions according to how activated, energized, or subdued they are. The distinction was difficult to detect in the six-word experiment but became visible when the vocabulary included many more nuanced terms.

The finding offers a possible explanation for why different theories of emotion sometimes appear to disagree. Studies using a small number of representative categories may identify only the largest contrasts, such as pleasant versus unpleasant or positive versus negative. When the vocabulary is expanded, additional dimensions can emerge, including activation, control, and subtle distinctions between closely related emotional states. In mathematical terms, the observed structure depends on both the underlying semantic space and the sampling of words used to probe it. A limited vocabulary may project a multidimensional system onto a simpler pattern, while a broader set of terms provides enough resolution to reveal relationships hidden between the major categories.

The study also points toward a practical role for large language models in psychology and linguistics. Researchers can ask GPT-4 to perform a controlled semantic task using ordinary instructions, then analyze the resulting arrangements with techniques such as multidimensional scaling or correlation-based comparisons. This may be more accessible than training specialized embedding models, assembling large datasets, or developing machine-learning pipelines. However, the approach still requires careful validation, because language models can reflect biases, uneven cultural representation, and patterns inherited from their training material. The Japanese team plans to investigate whether similar emotional maps appear in other languages, including Japanese, and whether cultural differences alter the distances among emotion terms.

The researchers stress that the map describes the semantic and conceptual organization of emotion words, not the architecture of emotional experience in the brain. A person may use the word “sadness” in a position that resembles a model’s placement of the term, while the biological, personal, and situational experience behind that word remains far more complex. Even so, the results demonstrate how AI can help expose hidden structure in language. By increasing the number of emotion terms that can be compared, GPT-4 revealed a nuanced organization in which broad emotional dimensions coexist with finer distinctions of arousal and control. The study, published in Scientific Reports, suggests that the future of emotion research may involve not only asking people how feelings relate, but also using language models to map the conceptual terrain surrounding those feelings.

Subject of Research: Semantic and conceptual organization of emotion terms using GPT-4, human judgments, and word embeddings

Article Title: Mapping 99 emotion terms with GPT4 prompting reveals nuanced semantic conceptual structure.

News Publication Date: 10-Jul-2026

Web References: https://doi.org/10.1038/s41598-026-60536-4

References: Scientific Reports, DOI: 10.1038/s41598-026-60536-4

Image Credits: Laboratory of Neurophysiology, NIBB

Keywords: GPT-4, artificial intelligence, emotion, semantic mapping, psychology, language models, word embeddings, arousal, pleasure, dominance, emotion concepts

Tags: AI semantic landscape of feelingsAI versus human emotion judgmentsAI-driven visualization of human feelingscomputational modeling of human emotionsemotion similarity spatial mappingemotion word proximity analysisGPT-4 emotion word organizationhuman emotion mappinglanguage models understanding emotionslarge-scale emotion comparison methodsmapping human emotional concepts with AIneurophysiology of emotional concepts