Overview
We ask AI to compile industry data. The answer reads airtight and confidently professional—until we check and find real data, outdated data, and invented data mixed in the same paragraph, with no visible difference on the surface. Behind this is a worsening asymmetry: AI has driven the cost of "generation" near zero while pushing the cost of "trust" to a historic high. Checking a 500-word answer often takes longer than writing it. Over time people form two bad habits: blind trust, or trusting nothing. Neither helps AI. We did not want to choose only between "reads smoothly but cannot be trusted" and "line-by-line human verification until you give up"—so we built HallucC. Before building, we surveyed what already exists and found a gap. Three judgments shaped the product. Judgment one: evaluation frameworks do not solve runtime problems. Existing eval tools target developers' pre-launch batch tests—run prompts, run datasets. But "which sentence in this answer is true" is a runtime, sentence-by-sentence question that non-technical colleagues often need to see. No product had turned "sentence-level verification + evidence + correction" into something ordinary people would actually use. Judgment two: false positives are the top killer of verification tools. A tool that keeps shouting "this sentence is made up" and then wrongly blames the AI will be abandoned by users within two weeks.
Connected profiles
More General-purpose agents products
Sources · 0Sources & provenance
English text is maintained separately from the original record. Consult the original source before citing a fact.
No source links are available for this profile.
Original source text
Names may use phonetic renderings when no English name is supplied. The original wording is retained below.
我们让 AI 整理一份行业数据，回答读起来严丝合缝、专业自信，核对时才发现——真数据、过时数据、凭空编的数据，混在同一段话里，外表完全看不出区别。这件事背后是一个正在恶化的不对称：**AI 把「生成」的成本降到了接近零，却把「信任」的成本推到了历史最高。**核对一段 500 字的回答，常常比写它还费时间。时间久了，大家养成两种坏习惯：要么盲目信任，要么一律不信。这两种态度，对 AI 都没有好处。我们不想在「读起来流畅但不敢信」和「逐句人工查证累到放弃」之间二选一——所以做了 HallucC。动手之前我们看了一圈市面上已有的东西，发现一个空档。三个判断，决定了这个产品的形状：判断一评测框架解决不了运行时问题现有 eval 工具面向开发者上线前的批量测试：跑 prompt、跑数据集。但「这条回答里哪一句是真的」是运行时、逐句、还经常要给非技术同事看的问题。没有产品把「逐句核验 + 给出证据 + 给出修正」做成普通人敢直接用的东西。判断二误报是核验工具的第一死因一个整天喊「这句是编的」、结果一查冤枉了 AI 的工具，两周之内就会被用户抛弃。