Why language models fail to assess human experiences.
A new study refutes the common belief that strong feelings must be expressed in long and detailed texts.
On the contrary, the gap between narrative complexity and emotional richness is not a mistake, but a conscious communicative strategy. Artificial intelligence, it turns out, poorly captures such nuances: its expressive range is almost 1.7 times narrower than that of humans.
Researchers analyzed 351,734 English-language relationship stories from open online forums (2012–2023), reports PsyPost. They compared two parameters:
Narrative complexity — volume, vocabulary diversity, sentence density.
Linguistic affective intensity — the strength of emotional load in words.
No expected correlation was found between them. Complex experiences can be expressed in a restrained manner, while a strong surge of emotions can be conveyed briefly and simply.
Four styles of expression were also discovered:
Connected expression (91.3%) — complexity and emotions are balanced.
Strategic understatement — strong emotions with minimal narrative structure.
Strategic exaggeration — very complex language with low emotional intensity, creating a protective distance.
Collapse — emotions overwhelm the author so much that a coherent narrative breaks down.
The Problem for AI
A comparison with a language model tuned for safety showed that its expressive range is about 1.7 times narrower than that of humans.
The model particularly struggles to recognize strategic understatement and expressive collapse — that is, situations where a person communicates suffering through restraint, fragmentary expression, or indirectness. In other words, the system, which only 'hears' intensity, misses those who speak quietly. It remains to be seen whether AI can improve its understanding of human emotions in the future.
Leave a comment