Where AI Slop Comes From and How to Turn It Down in a Model

Summary

The article examines whether AI slop is represented as a distinct direction inside a language model. Using the open 4-billion-parameter Qwen3-4B assistant, the researchers generated the same 100 pieces in three styles: the model’s normal voice, an exaggerated AI-assistant voice, and a plain-spoken human voice. The average difference between the exaggerated and plain versions was treated as a direction in the model’s vector representation. That direction ranked the sloppier version higher in all 64 hand-written test pairs. On the model’s everyday outputs, the direction’s score tracked the article’s slop score with a correlation of up to 0.82. The direction can also be inspected through the words it encourages, suggesting a way to measure or reduce this style within the model.