Language models are being explored to approximate human use of sensory language, but they differ significantly from human usage.
In a study using a corpus of parallel human and model responses to short story prompts, it was found that models generate stories that differ from human usage of sensory language.
Gemini models use significantly more sensory language than humans, while models from the other families use significantly less.
Linear probes suggest that models are capable of identifying sensory language, but instruction tuning may discourage its usage.