后训练护栏使LLM文本可检测，Pangram CTO称基础模型更多样
原标题：LLMs could write like humans but post-training guardrails make their text detectable
AI 摘要
AI文本检测公司Pangram的CTO Bradley Emi在博客文章中表示，大型语言模型理论上可以像人类一样多样化写作，但后训练和安全护栏使其文本可被检测。系统如ChatGPT、Claude或Gemini学习行为规则以避免危险输出或审查某些政治言论，这大大缩小了它们的表达范围，称为“模式崩溃”。基础模型和专门微调的模型写作更多样，因此Pangram的检测不会标记它们，但水印仍然有效。
正文节选
LLMs could write like humans but post-training guardrails make their text detectable LLMs could theoretically write as diversely as humans, but they don't. Post-training and safety guardrails keep their text detectable, argues Bradley Emi, CTO of AI text detector Pangram, in a blog post. Systems like ChatGPT, Claude, or Gemini learn behavioral rules to avoid dangerous outputs or censor certain political statements. This sharply narrows their expressive range, an effect called "mode collapse." So