When people share emotional stories, the intensity of their emotions doesn’t necessarily match the length or detail of their words. Recent research published in Pro Swan This suggests that the gap between what people express and how much they say is not a mistake, but an intentional communication strategy. These findings provide evidence that humans use a wide range of expressive styles that artificial intelligence currently struggles to reproduce.
Ryan Sang-baek Kim, founding director and principal investigator at the Ryan Institute in Paris, conducted the study to question common assumptions in psychology and computer science. Many experts believe that healthy communication requires people to perfectly match their inner feelings to their spoken and written words. Kim found that this discrepancy is usually ignored as an error.
“Affective science has long treated the gap between what people feel and what they say as measurement noise,” Kim told SciPost. “My research across psychology, emotional science, and AI ethics has led me to suspect that this gap is structure rather than noise.” People often regulate how much emotion becomes language, especially in the stories of relationships. “We wanted to test whether that regulation would leave a measurable impact on the data,” he said.
To map out these communication patterns, Kim analyzed exactly 351,734 English relationship stories. These articles were collected from public online advice forums and support communities between 2012 and 2023. To protect the privacy of the original authors, all personally identifying information was completely removed from the data. This vast collection provided a window into how real people discuss their relationships in a natural, unscripted environment.
Kim measured two key characteristics for every single story in the dataset. The first is narrative complexity, which is a structural measure of the text itself. This concept looks at the total length of posts, the variety of vocabulary used, and the density of sentence structure. Writing a very complex story requires a lot of mental effort.
The second feature was the linguistically inferred intensity of emotion. Emotion is a term used by psychologists to describe feelings and the experiences that underlie them. The researchers used special software to analyze the text and estimate the amount of emotion contained in the words. This tool measured how strong emotional language was, regardless of whether the overall sentiment was positive or negative.
By comparing these two measurements, Kim calculated the discrepancy in the story’s impact. This concept describes the precise mathematical gap between the complexity of a story and the emotional intensity it contains. He did not try to guess the hidden inner world of the writer. Instead, we simply measured how much linguistic effort people put into the page relative to the emotion they put into it.
“I was most surprised by the near-zero correlation,” Kim said. “We expected narrative complexity and emotional intensity to be at least weakly linked, but they were almost orthogonal,” he explained. In statistical terms, orthogonal variables are completely independent of each other. “Data shows that even if a story doesn’t sound emotionally intense, it can be psychologically complex,” Kim added.
“The main lesson is that emotionally complex experiences don’t necessarily sound emotional on the surface,” Kim said. “This challenges a common assumption in both emotion research and emotional AI: the idea that stronger or more difficult emotional states should manifest as stronger emotional language,” he noted. “Our data showed that people often described painful or psychologically difficult experiences in calm, restrained, or indirect terms rather than in highly emotional terms.”
“In other words, people who say, ‘I’m fine’ aren’t necessarily hiding their emotions badly,” Kim explained. “At times, restraint itself can be part of how humans communicate distress.” These findings provide evidence that humans use a wide range of expressive styles, rather than automatically matching complexity to emotions.
Kim identified four distinct patterns of emotional expression in the data. The majority of the stories, approximately 91.3 percent, fell into a category called conjunctive expressions. For this group, story complexity and emotional intensity were relatively balanced, with no extreme gaps. There were no serious signs of exaggeration or understatement of emotion in the writing.
The remaining stories were categorized into three specific discrepancy categories. Approximately 20,223 stories exhibited strategic understatement, with writers expressing intense emotions but using little narrative structure. Another 2,223 stories demonstrated strategic exaggeration, meaning that the author used highly complex language to express relatively low emotional intensity. This strategy indicates that people use broad vocabulary to create a protective cognitive distance from a topic.
The final group of 8,040 stories fell into a pattern that Kim calls collapse. These writers exhibited very strong emotional intensity, but lacked the structural expression to support it. This pattern tends to occur when your emotions are too strong to organize your thoughts into a cohesive story. The narrative structure effectively collapses under the weight of emotion.
After mapping out these human patterns, Kim tested the artificial intelligence system using a safety-focused language model. “An important finding for AI was that one language model aligned to RLHF occupied about 1.70 times smaller representational space than a human under the same measurement framework,” Kim said. This type of program is trained using reinforcement learning from human feedback, so it is polite and helpful. “This model was particularly absent in the emotional part of language, where humans speak indirectly, withhold, and shut down their emotions,” he pointed out.
“The clearest human signal is not always the loudest signal,” Kim added. “We were also surprised to see that the model contraction was uneven, especially in areas where humans communicate through strategic understatement and collapse of expression.”
“While the 1.70-fold contraction is statistically obvious, the real significance lies in where the contraction occurs,” Kim said. “When a tuned model occupies a narrow expression space, it can become less sensitive to people who communicate distress through understatement, confusion, silence, or fragmented words rather than direct emotional intensity. This is important for mental health tools, AI companions, and other systems that seek to interpret emotional language. Systems that only listen to intensity will miss people who speak suppressed.”
Readers may misunderstand the research by assuming that the software perfectly captures the author’s true inner feelings. “The most important caveat is that this study does not claim to directly measure subjective emotions,” Kim cautioned. “It measures the geometry of emotional expression, not the full inner experience underlying it, but what people put into their language.”
This study also has some limitations regarding scope and sample. “The data also comes from English-language public relations stories, so patterns may vary by language, culture, and environment,” Kim said. “Finally, AI comparisons involve one model under a fixed configuration, so they should be read as baseline results rather than judgments against all aligned models.” Beyond these three limitations, Kim suggests that the safest interpretation is that this study measures one stable asymmetry between the representational geometry of humans and aligned models, and not a verdict on emotional AI as a whole.
For future research, Kim plans to examine how these communication styles change over time. “This research is part of a broader research program on the geometry and emotional sovereignty that influences narrative,” Professor Kim said. “The empirical question is how people construct emotional meanings in language.” The governance question is what happens when artificial intelligence systems begin to interpret those meanings for humans, he noted.
He aims to study how long-term interactions with artificial intelligence affect human behavior. “My next step is longitudinal: whether repeated exposure to the conditioned model changes the way people express, control, and interpret their emotions over time,” Kim said. “The deeper question is whether AI will only respond to our emotional language, or will it change the way we talk about ourselves over time?”
Research resources are fully open to the public to encourage further research. “The dataset and analysis code are publicly available on Zenodo,” Kim said. “Open data and reproducible analysis are especially important here, as claims about AI and emotions are prone to speculation. My hope is that other researchers will test, challenge, and extend this framework.”
The study, “Narrative and Influence Discrepancies as Regulated Degrees of Freedom in 351,734 Relationship Narratives,” was authored by Ryan SangBaek Kim.
