Why AI Writing Still Reads Like Slop — Louis-François Bouchard, Towards AI
Towards AI releases Tone-Bench, a benchmark measuring LLM creative writing quality and 'slop' avoidance
“reading something that looks like generated text is truly terrifying”
Louis-François Bouchard from Towards AI defines AI 'slop' across six dimensions—overused words, excessive generality, empty content, repetition, hallucination, and tone mismatch—and launches Tone-Bench, a new benchmark for measuring creative writing quality in LLMs. The talk argues that better models alone won't solve the slop problem, as low-quality AI text continues to scale even with model improvements. The benchmark is a niche but practical tool for the growing conversation around AI content quality, though it comes from a non-major research lab.