What 9,050 AI-Written Scripts Taught Me About Slop

A recent analysis of 9,050 AI-generated scripts reveals that while 'slop'—repetitive, generic AI phrasing—is a common issue, it can be effectively mitigated through detailed rubrics and specific prompting. The author, who leads a team at Towards AI, tested 181 model configurations to determine which could best replicate a human voice. Key findings suggest that 'max effort' settings often yield diminishing returns, with smaller, cheaper open-source models capable of producing high-quality results when paired with a robust judging pipeline. The study emphasizes that the real challenge lies in the selection process rather than generation. By using a panel of diverse LLM judges to evaluate and merge drafts, users can achieve superior writing quality at a fraction of the cost. The author concludes that while AI can significantly streamline content creation, human oversight remains essential for final polish and maintaining a unique, authentic tone.