---
title: "什么是 Hallucination？"
canonical: https://wavect.io/zh/glossary/hallucination/
language: zh
description: "当 LLM 产出流畅、自信却纯属虚假的输出时，这是模型工作方式的直接后果，无法被完全消除，只能被减少。"
image: "https://wavect.io/img/general/bak/open_graph_preview.jpg"
---

技术

# Hallucination

当 LLM 产出流畅、自信却纯属虚假的输出时，这是模型工作方式的直接后果，无法被完全消除，只能被减少。

最近审阅: 2026-06-02 审阅人 [Kevin Riedl](/zh/team/kevin-riedl/) [wiki ↗](https://www.wikidata.org/wiki/Q139796365)

幻觉是指模型生成了听起来正确而实际错误的东西：一个编造的引用、一个虚构的 API、一个貌似可信却虚假的事实。它不是你能打补丁修掉的故障。LLM 预测的是可能的文本，而一个流畅、听起来自信的回答在统计上就是可能的，无论它碰巧是否为真。模型没有「我不知道」的内在感觉，所以它用某个符合模式的东西来填补空白。

因为它是结构性的，诚实的定位是降低风险，而非消除。最大的杠杆是接地：在运行时通过检索（ [RAG](/zh/glossary/rag/) ）把真实事实给模型，让它从真实的源文本作答，而不是从它训练时刻的猜测作答。要注意 [微调](/zh/glossary/fine-tuning/) 在这里是用错了杠杆，因为它塑造的是行为，而不是提供事实。约束输出格式，让即兴发挥的空间更小。让模型引用你能核查的来源。而且至关重要的是要评估，构建一个真实问题的测试集，衡量系统出错的频率，并把这个数字当作一项像其他指标一样要追踪的质量指标。

这就是 AI 与 QA 相遇之处，而大多数团队会跳过它。在没有评估框架的情况下上线一个 LLM 功能，等于上线未经测试的代码却把它叫做完成。你需要在你的用户替你发现之前，就知道你的错误率。我们把这一点视为在 [Software Quality Assurance](/zh/services/software-quality-assurance/) 下不可协商的事项。

为什么评估装置不可协商，举个实际例子：一个团队手动测了十几个问题、看上去都很好，就把一个法律文档助手上线了。到了生产环境，它自信地引用了一条上传合同里根本不存在的条款，一位用户据此采取了行动，于是出现了一项真实的隐患。本能能抓住这个问题的那套装置并不光鲜：几百个带已知正确答案的真实问题，在每次改动时都跑一遍，产出一个单一的数字，即系统答错的百分比。没有它，你就不知道自己的错误率，也就意味着由你的用户替你发现，一次一个错误答案地发现。有了它，你就能在上线之前判断，对照风险，这个错误率是否可以接受。

在受监管或高风险领域，信任这一面就是全部的游戏。一个推荐电影的聊天机器人里的幻觉只是耸耸肩的事。同样的幻觉出现在金融、法律或医疗输出中就是一项隐患。让护栏与出错代价相匹配，永远不要让一个流畅的回答取代一个经过核实的回答。

// FAQ

## 常见问题

### LLM 为什么会产生幻觉？

因为它们预测的是可能的文本，而不是经过核实的真相。一个流畅、自信的回答在统计上就是可能的，无论是否正确，而模型没有内置的’我不知道’信号，所以它会填补空白。

### 幻觉能被彻底修复吗？

不能。它们是 LLM 工作方式的结构性产物。你可以用接地（检索）、受约束的输出、来源引用和评估来减少它们，但总会残留一定比例。要管理它，不要假设它已经消失。

### 在生产环境中如何减少幻觉？

通过 RAG 把答案接地到真实数据，约束输出格式，要求可核查的来源，并运行一个在真实问题上衡量你错误率的评估框架。让护栏的强度与出错代价相匹配。

相关术语

- [LLM](/zh/glossary/llm/)
- [RAG](/zh/glossary/rag/)
- [Fine-Tuning](/zh/glossary/fine-tuning/)

相关服务

- [Software Quality Assurance](/zh/services/software-quality-assurance/)
- [人工智能](/zh/services/artificial-intelligence/)

相关案例研究

- [Twinsoft AI](/zh/case-studies/twinsoft-ai/)

相关指南

- [如何为 MVP 选择技术栈](/zh/software-development-guide/how-to-choose-a-tech-stack-for-mvp/)

聊聊

## Structured Data

```json
{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@id": "https://wavect.io/#organization",
      "@type": [
        "Organization",
        "ProfessionalService",
        "LocalBusiness"
      ],
      "employee": [
        {
          "@id": "https://wavect.io/team/kevin-riedl/#person",
          "@type": "Person",
          "jobTitle": "Managing Director",
          "name": "Kevin Riedl",
          "url": "https://wavect.io/team/kevin-riedl/",
          "worksFor": {
            "@id": "https://wavect.io/#organization",
            "@type": [
              "Organization",
              "ProfessionalService",
              "LocalBusiness"
            ]
          }
        },
        {
          "@id": "https://wavect.io/team/christof-jori/#person",
          "@type": "Person",
          "jobTitle": "Managing Director",
          "name": "Christof Jori",
          "url": "https://wavect.io/team/christof-jori/",
          "worksFor": {
            "@id": "https://wavect.io/#organization",
            "@type": [
              "Organization",
              "ProfessionalService",
              "LocalBusiness"
            ]
          }
        }
      ],
      "founder": [
        {
          "@id": "https://wavect.io/team/kevin-riedl/#person",
          "@type": "Person",
          "jobTitle": "Managing Director",
          "name": "Kevin Riedl",
          "url": "https://wavect.io/team/kevin-riedl/",
          "worksFor": {
            "@id": "https://wavect.io/#organization",
            "@type": [
              "Organization",
              "ProfessionalService",
              "LocalBusiness"
            ]
          }
        },
        {
          "@id": "https://wavect.io/team/christof-jori/#person",
          "@type": "Person",
          "jobTitle": "Managing Director",
          "name": "Christof Jori",
          "url": "https://wavect.io/team/christof-jori/",
          "worksFor": {
            "@id": "https://wavect.io/#organization",
            "@type": [
              "Organization",
              "ProfessionalService",
              "LocalBusiness"
            ]
          }
        }
      ],
      "legalRepresentative": [
        {
          "@id": "https://wavect.io/team/kevin-riedl/#person",
          "@type": "Person",
          "jobTitle": "Managing Director",
          "name": "Kevin Riedl",
          "url": "https://wavect.io/team/kevin-riedl/",
          "worksFor": {
            "@id": "https://wavect.io/#organization",
            "@type": [
              "Organization",
              "ProfessionalService",
              "LocalBusiness"
            ]
          }
        },
        {
          "@id": "https://wavect.io/team/christof-jori/#person",
          "@type": "Person",
          "jobTitle": "Managing Director",
          "name": "Christof Jori",
          "url": "https://wavect.io/team/christof-jori/",
          "worksFor": {
            "@id": "https://wavect.io/#organization",
            "@type": [
              "Organization",
              "ProfessionalService",
              "LocalBusiness"
            ]
          }
        }
      ],
      "name": "Wavect GmbH",
      "subjectOf": {
        "@id": "https://wavect.io/verified-claims.json#dataset",
        "@type": "Dataset",
        "creator": {
          "@id": "https://wavect.io/#organization",
          "@type": [
            "Organization",
            "ProfessionalService",
            "LocalBusiness"
          ]
        },
        "description": "A machine-readable registry of quantitative and qualitative claims published by Wavect, with review dates, localized page appearances and public third-party citations where available.",
        "inLanguage": "en",
        "isAccessibleForFree": true,
        "license": "https://creativecommons.org/licenses/by/4.0/",
        "name": "Wavect verified publication claims",
        "url": "https://wavect.io/verified-claims.json"
      },
      "url": "https://wavect.io/"
    },
    {
      "@id": "https://wavect.io/team/kevin-riedl/#person",
      "@type": "Person",
      "jobTitle": "Managing Director",
      "name": "Kevin Riedl",
      "sameAs": [
        "https://www.wikidata.org/wiki/Q139796365",
        "https://www.linkedin.com/in/wsdt",
        "https://github.com/wsdt"
      ],
      "url": "https://wavect.io/team/kevin-riedl/",
      "worksFor": {
        "@id": "https://wavect.io/#organization",
        "@type": [
          "Organization",
          "ProfessionalService",
          "LocalBusiness"
        ]
      }
    },
    {
      "@id": "https://wavect.io/team/christof-jori/#person",
      "@type": "Person",
      "jobTitle": "Managing Director",
      "name": "Christof Jori",
      "sameAs": [
        "https://www.wikidata.org/wiki/Q139796367",
        "https://www.linkedin.com/in/jocr77/",
        "https://github.com/jo-chris"
      ],
      "url": "https://wavect.io/team/christof-jori/",
      "worksFor": {
        "@id": "https://wavect.io/#organization",
        "@type": [
          "Organization",
          "ProfessionalService",
          "LocalBusiness"
        ]
      }
    },
    {
      "@id": "https://wavect.io/#website",
      "@type": "WebSite",
      "inLanguage": [
        "en",
        "de",
        "es",
        "zh"
      ],
      "name": "Wavect",
      "potentialAction": {
        "@type": "SearchAction",
        "query-input": "required name=search_term_string",
        "target": {
          "@type": "EntryPoint",
          "urlTemplate": "https://wavect.io/search/?q={search_term_string}"
        }
      },
      "publisher": {
        "@id": "https://wavect.io/#organization",
        "@type": [
          "Organization",
          "ProfessionalService",
          "LocalBusiness"
        ]
      },
      "url": "https://wavect.io/"
    }
  ]
}
```

```json
{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@id": "https://wavect.io/zh/glossary/hallucination/#term",
      "@type": "DefinedTerm",
      "alternateName": [],
      "description": "当 LLM 产出流畅、自信却纯属虚假的输出时，这是模型工作方式的直接后果，无法被完全消除，只能被减少。",
      "inDefinedTermSet": "https://wavect.io/zh/glossary/#termset",
      "name": "Hallucination",
      "termCode": "hallucination",
      "url": "https://wavect.io/zh/glossary/hallucination/"
    },
    {
      "@id": "https://wavect.io/zh/glossary/hallucination/#webpage",
      "@type": "WebPage",
      "about": {
        "@id": "https://wavect.io/#organization",
        "@type": [
          "Organization",
          "ProfessionalService",
          "LocalBusiness"
        ]
      },
      "author": {
        "@id": "https://wavect.io/team/kevin-riedl/#person",
        "@type": "Person",
        "name": "Kevin Riedl",
        "url": "https://wavect.io/team/kevin-riedl/"
      },
      "dateModified": "2026-06-02",
      "description": "当 LLM 产出流畅、自信却纯属虚假的输出时，这是模型工作方式的直接后果，无法被完全消除，只能被减少。",
      "headline": "Hallucination",
      "inLanguage": "zh",
      "isPartOf": {
        "@id": "https://wavect.io/#website",
        "@type": "WebSite"
      },
      "lastReviewed": "2026-06-02",
      "mainEntity": {
        "@id": "https://wavect.io/zh/glossary/hallucination/#term"
      },
      "mentions": [
        {
          "@id": "https://wavect.io/#organization",
          "@type": [
            "Organization",
            "ProfessionalService",
            "LocalBusiness"
          ]
        },
        {
          "@type": "Service",
          "name": "Software Quality Assurance",
          "url": "https://wavect.io/zh/services/software-quality-assurance/"
        },
        {
          "@type": "Service",
          "name": "人工智能",
          "url": "https://wavect.io/zh/services/artificial-intelligence/"
        }
      ],
      "name": "Hallucination",
      "reviewedBy": {
        "@id": "https://wavect.io/team/christof-jori/#person",
        "@type": "Person",
        "name": "Christof Jori",
        "url": "https://wavect.io/team/christof-jori/"
      },
      "speakable": {
        "@type": "SpeakableSpecification",
        "cssSelector": [
          ".gls-single__h1",
          ".gls-single__tldr"
        ]
      },
      "url": "https://wavect.io/zh/glossary/hallucination/"
    },
    {
      "@type": "BreadcrumbList",
      "itemListElement": [
        {
          "@type": "ListItem",
          "item": "https://wavect.io/zh/",
          "name": "首页",
          "position": 1
        },
        {
          "@type": "ListItem",
          "item": "https://wavect.io/zh/glossary/",
          "name": "技术、产品与合作模式术语表",
          "position": 2
        },
        {
          "@type": "ListItem",
          "item": "https://wavect.io/zh/glossary/hallucination/",
          "name": "Hallucination",
          "position": 3
        }
      ]
    }
  ]
}
```

```json
{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "因为它们预测的是可能的文本，而不是经过核实的真相。一个流畅、自信的回答在统计上就是可能的，无论是否正确，而模型没有内置的'我不知道'信号，所以它会填补空白。"
      },
      "name": "LLM 为什么会产生幻觉？"
    },
    {
      "@type": "Question",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "不能。它们是 LLM 工作方式的结构性产物。你可以用接地（检索）、受约束的输出、来源引用和评估来减少它们，但总会残留一定比例。要管理它，不要假设它已经消失。"
      },
      "name": "幻觉能被彻底修复吗？"
    },
    {
      "@type": "Question",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "通过 RAG 把答案接地到真实数据，约束输出格式，要求可核查的来源，并运行一个在真实问题上衡量你错误率的评估框架。让护栏的强度与出错代价相匹配。"
      },
      "name": "在生产环境中如何减少幻觉？"
    }
  ],
  "speakable": {
    "@type": "SpeakableSpecification",
    "cssSelector": [
      ".faq-question",
      ".faq-answer"
    ]
  }
}
```