Interpreting Pangram: Pangram AI Slop Detector Reads Structure, Not Wording
博客要点 (中文)
Armin Ronacher 2026-09-14 实测稿。先让 Claude Opus 5 按 8 段 prompt 写一段 David Sacks 风格的「pacing the frontier」推文，Pangram 判 100% AI；接着他自己一句句改写，结构基本沿用、只换词换句，最后相似度检测说 50% 重合、没有任何一句原句保留，Pangram 还是判 100% AI slop。结论：Pangram 检测的不是词面，是「LLM 帮过的结构痕迹」——只要你第一步用过 LLM 给出提纲，结构本身就被打上 slop 标签，后面的人肉重写救不回来。Ronacher 反思写两年 AI-assisted 博客之后，自己既更依赖 LLM、又被 LLM 反向改写得很厉害。对任何在 X / 公众号用 AI 起草、又希望过 AI 检测器的内容工作者，这是当下最诚实的一份现场报告。
Key claims (English)
Armin Ronacher's 2026-09-14 empirical post: he has Claude Opus 5 write a David-Sacks-style 'pacing the frontier' tweet following 8 prompts, Pangram returns 100% AI; he then rewrites it sentence by sentence, keeping the structure but swapping words and phrasings, similarity detection reports 50% overlap with no original sentence preserved — and Pangram still calls it 100% AI slop. The conclusion: Pangram doesn't detect wording, it detects 'the structural traces of LLM assistance'. Once an LLM has produced the outline, the structure itself is tagged slop and subsequent human rewriting can't recover it. Ronacher reflects on writing two years of AI-assisted blog and finding himself both more LLM-dependent and more LLM-rewritten. For anyone using AI to draft on X / WeChat while hoping to pass AI detectors, this is the most honest in-the-field report so far.
Obsidian 证据摘录
「结论是 Pangram 检测的不是词面，是『LLM 帮过的结构痕迹』——只要你第一步用过 LLM 给出提纲，结构本身就被打上 slop 标签，后面的人肉重写救不回来。Ronacher 反思自己写了两年 AI-assisted 博客之后，发现自己既更依赖 LLM、又被 LLM 反向改写得很厉害。」——OpenClaw定时任务/AK-RSS-Digest（89源精选）/2026-09-15-AK-RSS-Digest（89源精选）.md L49
链接
- 原文：https://lucumr.pocoo.org/2026/9/14/interpreting-pangram/
- AK-RSS-Digest：OpenClaw定时任务/AK-RSS-Digest（89源精选）/2026-09-15-AK-RSS-Digest（89源精选）.md