How an AI System Fact-Checks Content with Atomic Claims
掘金 · 人工智能 describes an AI content-verification system that breaks text into atomic claims, retrieves evidence by authority level, and has models classify relationships while rules determine final labels. It uses a multistage small-model pipeline for large-scale processing and has reportedly been deployed in five business categories, including knowledge encyclopedias and content Q&A. The report does not provide independent validation of the system’s accuracy or cost reductions.
The account offers a practical architecture for large-scale AI fact-checking, separating semantic judgment from deterministic rules and narrow tasks to enable smaller models.
导读
AIGC 内容里的事实错误越来越隐蔽：前几段都对，最后一句张冠李戴；检索结果里对的错的都有，模型挑了不权威的那条。我们建设了一套内容 AI 核查系统，核心做法是把内容拆成最小的可验证主张，以权威证据为准逐条判断，让模型只判断证据与主张的关系，由规则给出结论。在此基础上，我们用多阶段小模型流水线把核查成本降下来，覆盖海量批量数据。系统已在知识百科、内容问答等五类业务落地。
01 核查难在哪
先看几个线上真实出现过的错误：
这些错误有个共同点：单看每一句都”有出处”，错误藏在局部、藏在限定条件里、藏在多个数字之间的关系里。人工抽查覆盖不了这么大的量，简单规则也识别不出来。