Cheap "System One" models rival frontier LLMs in content moderation accuracy

A new research paper from arXiv explores the effectiveness of "System One" models for content moderation, finding them to be accuracy-competitive with frontier LLMs like GPT 5.x, Claude, Gemini-3.*, and Grok, while being significantly cheaper. The study highlights that these cheaper models can achieve an AUROC of 0.83, comparable to many advanced models, and are well-calibrated. However, the research also identifies a critical distinction between human disagreement and decision risk, suggesting that a single confidence score from these models may not adequately capture both uncertainties, especially when dealing with nuanced content like satire or coded language. AI

IMPACT Suggests cost-effective AI solutions for content moderation could be viable alternatives to expensive frontier models.

RANK_REASON Research paper published on arXiv detailing findings about AI model performance. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• .claude
• DagsHub
• Gemini-3.*
• Gotit.pub
• GPT 5.x
• Grok
• Hugging Face
• ScienceCast
• System One

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →