How AI Writes Emotion: Naming, Bodily Clichés and Objects

AI stories mostly convey emotion through stock bodily reactions, not emotion names. The StoryScope finding, one scene in three versions, a corrected term.

Paylaş
How AI Writes Emotion: Naming, Bodily Clichés and Objects
Beyond 'Show, Don't Tell': AI Story Generation & Character Emotions

How AI Writes Emotion: Naming vs. Bodily Clichés

Bodily Clichés in AI Fiction: One Scene Written Three Ways

Bodily Clichés in AI Fiction: One Scene Written Three Ways

Summary and Conceptual Frame

A story can convey emotion in three ways: it names the emotion, it reports the character's bodily reaction, or it leaves the emotion to a behaviour, an object or a setting. The StoryScope study, published in 2026, reported that in stories written by five large language models, explicit emotion names were the dominant mode less often than in human stories, while bodily sensation was the dominant mode far more often. This article reports that finding, writes the same scene in three ways, and corrects the term delivery-mode drift, which this site had defined in one direction only, to a two-directional definition.

In AI-written stories emotion is mostly not named; it is reported through stock bodily reactions such as a pounding heart, a tight throat or shaking hands. This article calls those stock reactions bodily clichés for short. The writing advice "show, don't tell" asks the writer to take the emotion name out of the scene; it does not say what to put in its place. The finding below suggests that models fill that gap mostly with the body.

Three ways to convey emotion

Computational linguists have been annotating the channel through which emotion passes in a text for some time. Casel, Heindl and Klinger (2021) annotated emotion components in literature and in Twitter posts and wrote that authors of literature tend to show what characters do and leave the interpretation to the reader, while on Twitter emotion is mostly reported directly. Kim and Klinger (2019) studied non-verbal channels of emotion in fan fiction, such as facial expression, voice, gesture and posture.

This article separates three ways. The separation is a proposed classification; its borders are not sharp in every scene.

At a glance

WayExampleShare of stories where it is dominant in StoryScope (human / model)Place in Objective Projection
Naming"He was very sad."29% / 8%The Emotion Embargo leaves this out of the narrator's voice.
Bodily reaction"His heart was pounding."38% / 81%No rule forbids it; it may not pass the camera question.
Behaviour, object, setting"He began washing a clean glass."Not given in the section read for this article.The way the method mainly works.

What did StoryScope find?

Russell and colleagues collected 10,272 writing prompts. For each prompt there is one human story, and one story was requested from each of five different language models; the models refused 24 requests, which left 61,608 stories in total. The human stories were taken from published books. A language model extracted the narrative features of the stories; two human annotators separately labelled 240 items from 12 stories. Mean agreement between human and model was reported as κ 0.84, and agreement between the two humans as κ 0.74.

Data. On the question "How are characters' emotions most commonly conveyed?", explicit emotion labels were the dominant mode in 29% of human stories and 8% of model stories. Bodily sensation and bodily metaphor were the dominant mode in 81% of model stories and 38% of human stories. The narrator commented explicitly on the theme in 77% of model stories and 52% of human stories.

Interpretation. These figures do not fit the idea that "models tend to name emotions". The models in this study avoid emotion names more than humans do and use the body most in their place. The interpretation has limits: what is measured is the dominant mode per story, not a count of sentences; a model extracted the features; the human stories come from published books; the study does not compare older and newer model versions.

Sample scene: one moment, three versions

The three versions below tell the same moment: Tuncay comes home early on the day he is laid off. The scenes were written for this article.

Version A (naming): When Tuncay got home he was very sad. He had been laid off and did not know how to tell Aysel. He felt a deep shame.

Version B (bodily reaction): Tuncay came through the door. His heart was pounding. His throat tightened and his stomach clenched. He took a deep breath; his hands were shaking.

Version C (behaviour and object): Tuncay came through the door. He walked to the kitchen with his ID badge still around his neck. The phone alarm set for the shuttle home went off; he silenced it. He pushed the cardboard box onto the top of the fridge, behind the jars. "You're early," Aysel said. Tuncay turned on the tap and began washing a clean glass.

Version B contains no emotion name; even so, the scene does not belong to Tuncay. The same four sentences could be written for a student leaving an exam or an actor about to go on stage. In version C the badge, the alarm, the cardboard box and the clean glass belong to this day only. What a reader takes from version C is not known: some readers may ask what is in the box and find no answer.

Delivery-mode drift: the corrected definition

In September 2026 this site defined delivery-mode drift as a model's tendency to name emotions instead of encoding them in physical detail. The StoryScope figures do not support that direction. The definition has been corrected as follows:

Delivery-mode drift: a systematic shift, relative to human writing, in how a language model conveys emotion. The shift may run toward naming the emotion or toward stock bodily reactions; its direction and size are measured per model, prompt and genre.

The new definition claims no direction. Which direction holds under which condition is the question of the open designs in the research agenda, and none of them has been run yet.

From the viewpoint of Objective Projection

The Emotion Embargo leaves out one thing only: the narrator naming the emotion in the scene. It does not forbid bodily reactions. Version B therefore follows the rule; yet it is not the scene the method asks for. The camera question from the imagery article helps here: a camera sees the shaking hand; it does not see the tight throat or the clenched stomach. A body sentence the camera cannot see comes close to an emotion statement with the name left off. Filter words such as "felt" and "noticed" set up the same distance.

Proposal (untested): a quick check for a body sentence is to move it to another character and another situation. If the sentence moves without changing, it does not belong to the scene. This runs in the same direction as the idea in the subtext article: what is left unsaid carries if it leaves a trace in the scene.

Data. In the same study the score for "setting as a mirror of inner state" was given as 4.07 for model stories and 3.58 for human stories; the range of the scale was not verified for this article.

Interpretation. Making the setting a mirror of the character's inner state also appears to be more marked in model stories. So leaving emotion to objects and setting does not, on its own, set a text apart from model writing. The difference in stating the theme points the same way as the proposal "the theme is not told to the reader" in the theme vs. topic article; that proposal is untested as well.

Three questions when revising a draft

  1. Can a camera film this sentence? If it cannot, the sentence reports an inner state; the distinction between objective and subjective description is useful here.
  2. Can the sentence move to another scene unchanged? If it can, an object or action specific to this scene can be looked for.
  3. How many body sentences follow one another on a page? One shaking hand can do its work; when four body sentences stack up, the reader reads a list, not a scene.

These questions are a checklist, not a rule. The Scene Checker flags emotion names and similes; it does not flag bodily clichés. Why it matters whether a model or a human labelled a feature is covered in the article on inter-annotator agreement; whether readers can tell model stories apart is covered in a separate article.

What the doctrine cannot say here

  • It cannot say that emotion conveyed through objects and actions is built more strongly or more accurately by readers than emotion conveyed through bodily reactions; this has not been tested.
  • It cannot say that bodily reactions are bad writing; they are the dominant mode in 38% of human stories too.
  • It cannot say that models have changed over time, or why they write this way; the study reported here does not compare versions and gives no cause.
  • It cannot say that these figures hold for Turkish stories or for short scenes.
  • It cannot say that a text can be identified as model-written from its body sentences.

Limitations

This article rests on three figures from a single study, and only the main text of that study was read, not its appendices. The three-way separation is the author's proposal; whether two independent people would apply it in the same way has not been measured. The sample scenes were written by one author and were not tested with any reader. "Bodily cliché" is used here as a descriptive shorthand; it is not a proposal for a new term. The relations of the Emotion Embargo and delivery-mode drift can be seen in the knowledge graph of Levent Bulut; all work on Objective Projection is listed on the corpus page of Levent Bulut.

Frequently Asked Questions

How does AI write emotion?

According to the 2026 StoryScope study, in the stories of the five language models examined, emotion is conveyed mostly through bodily sensation and bodily metaphor. Explicit emotion names are the dominant mode less often in model stories than in human stories.

What is a bodily cliché?

A stock sentence of bodily reaction, such as a racing heart, a tight throat, a clenched stomach or shaking hands, that fits almost any character and any situation unchanged. It is used in this article as a descriptive shorthand.

Is it wrong to write bodily reactions?

No. Bodily reaction is a common way in human stories too. The question is whether the sentence is specific to this scene and how many of them follow one another.

What is delivery-mode drift?

A systematic shift, relative to human writing, in how a language model conveys emotion. The shift may run toward naming the emotion or toward stock bodily reactions; its direction and size are measured per model, prompt and genre.

Does this finding make "show, don't tell" invalid?

No. The finding suggests that removing the emotion name is not enough on its own. What takes the name's place is a separate question, and the advice does not answer it.

References

  • Russell, J., Rajendhran, R., Pham, C. M., Iyyer, M. and Wieting, J. (2026). StoryScope: Investigating idiosyncrasies in AI fiction. COLM 2026. arXiv:2604.03136.
  • Casel, F., Heindl, A. and Klinger, R. (2021). Emotion Recognition under Consideration of the Emotion Component Process Model. arXiv:2107.12895.
  • Kim, E. and Klinger, R. (2019). An Analysis of Emotion Communication Channels in Fan-Fiction: Towards Emotional Storytelling. Proceedings of the Second Workshop on Storytelling, 56–64. doi:10.18653/v1/W19-3406
  • Bulut, L. (2026). A Research Agenda for Objective Projection: Testable Hypotheses, Power Analyses, and Open Designs (v0.2). Zenodo. doi:10.5281/zenodo.24115699

Citation

@misc{bulut2026bodilycliches,
  author       = {Bulut, Levent},
  title        = {How AI Writes Emotion: Naming, Bodily Clichés and Objects},
  year         = {2026},
  howpublished = {\url{https://leventbulut.com/how-ai-writes-emotion-bodily-cliches/}},
  note         = {leventbulut.com}
}
G-Verified: Levent Bulut