Artificial intelligence overuses ancient rhetorical figures in text generation

Federico Boggia, a teacher and trainer from Italy, identified that large language models frequently use a rhetorical device called epanorthosis. This technique involves a writer making a statement and immediately replacing it with a grander, more emphatic version, such as saying a product is not a course but a journey.

Boggia argues that this habit likely stems from models being trained on promotional web text and then being refined through human feedback that rewards confident, punchy phrasing.

Boggia measured these patterns by comparing model outputs against human writing across several genres. In oratorical speeches, the models produced these emphatic corrections at more than double the rate of human speakers. Conversely, in informal question-and-answer writing, the models used the figure at less than one-fifth the rate of human authors, who tend to use more spontaneous qualifications.

The researcher found that while models use this device for dramatic effect, they often fail to improve their actual reasoning when they use self-correction phrases like "let me reconsider." To address this, Boggia tested several ways to reduce the habit without retraining the entire model. He developed a removable adapter that could act as a stylistic dial, successfully lowering the frequency of these corrections to match human rates in some genres.

He also found that simple prompt instructions or automatically selecting the least repetitive response among multiple candidates could significantly curb the habit. Boggia suggests these patterns might reflect deep structural tendencies in language that both ancient orators and modern machines have discovered.