Glossary · Core Concepts

Hallucination

Core Concepts 新手

30-Second Version · For the impatient

When AI generates plausible-sounding but factually wrong or invented information — stated with full confidence but simply not true.

Full Explanation +

01 · What is this?

Hallucination refers to a language model generating content that appears plausible and is expressed with confidence, but is actually wrong or entirely fabricated. The key point: AI doesn't know it's hallucinating when it does. It genuinely believes what it's saying is correct — because a language model's training objective is to generate the next plausible token, not to verify facts before outputting. When the model has insufficient training data on a topic, it doesn't say I don't know; instead it fills the answer with the statistically most plausible token sequence. You cannot judge whether an answer is correct from Claude's expressed confidence: the more certain it sounds doesn't mean it's more likely to be true. The danger of hallucination is that it's indistinguishable in form from a correct answer — only verification can reveal it.

02 · Why does it exist?

Hallucination is a direct byproduct of language model training objectives. Models are trained to minimize the loss function for predicting the next token — this makes them excellent at generating linguistically fluent, logically coherent text. But this training objective itself doesn't require accuracy — a grammatically perfect but factually wrong sentence contributes identically to the loss function as a factually correct one. Language models have no internal I don't know state. Humans say I don't know because they're aware of their own knowledge boundaries; language models have only one mechanism — generating tokens. When asked about a topic for which they have no reliable data, they don't fall silent — they continue generating plausible-sounding content, even if fabricated.

03 · How does it affect your decisions?

Knowing hallucination exists, you need to adjust three work habits: Treat Claude's output as a draft, not an answer. Numbers, statistics, academic citations, legal text — these have the highest hallucination rates and must be independently verified before use. Use RAG to reduce hallucination rates. Use RAG architecture to inject real data into the prompt, letting Claude answer based on data you provide rather than training memory. Ask Claude to flag uncertainty. Add to your System Prompt: If you're uncertain about a fact, explicitly state that you're uncertain rather than guessing.

04 · What should you do?

Three things you can do immediately: Build a needs-verification list. Identify which types of Claude output must be verified and which can be trusted more. Conceptual explanations have low hallucination rates; specific numbers, citations, regulations, and attributed quotes have high hallucination rates and must be verified. Use multi-round confirmation on high-risk tasks. For important factual queries, have Claude give an answer first, then follow up with: How confident are you? Which parts are you uncertain about? Treat citations as search leads, not direct sources. Let Claude provide citations, but treat them as search leads rather than directly usable references. Use the author names and years to search Google Scholar for verification.

Real-World Example +

Scenario: Writing an analysis report on AI regulation that requires citing academic research. Wrong workflow: Ask Claude what research supports strict AI regulation? Claude provides 5 seemingly complete citations with author names, journal names, years, and abstracts. You copy-paste them directly into the report; after publication, readers discover 3 of the citations don't exist. Why this happens: Claude has no access to academic databases. When asked for specific citations, it generates a set of plausible-sounding citations using statistical patterns from its training data. Author names may come from real AI researchers, journal names may be real journals, but the combination is fabricated. Correct workflow (four steps): 1. Ask Claude to summarize the main academic arguments in favor of strict AI regulation — no specific citations needed → get an argument structure framework 2. Use these arguments as search leads to find real papers in Google Scholar 3. Once you find real papers, have Claude help you understand and summarize their content 4. Only cite sources you've personally confirmed exist

Diagram

Feel free to share. Please credit the source.

Common Misconceptions +

✕ Misconception 1

× Myth 1: The more certain Claude sounds, the more likely it's true. The confidence level when hallucinating is not fundamentally different from when generating accurate information. Research suggests models are more prone to expressing high confidence when they know less about a topic. Always verify factual claims independently regardless of how certain Claude sounds.

✕ Misconception 2

× Myth 2: The latest version of Claude doesn't hallucinate. Every version of Claude hallucinating — just at different rates. Hallucination is a structural characteristic of language models, not a bug that can be fixed. Don't relax your verification habits just because you're using the latest version.

The Missing Link +

Direct Impact

Speed of AI-generated factual content vs. verification cost: Claude can generate an analysis report in seconds, but if it contains hallucinations, you need time to verify each claim. For daily work, the speed advantage usually offsets verification costs; for high-risk scenarios (legal documents, medical information, financial reports), verification costs may exceed the speed advantage. RAG vs. pure language model: Using RAG architecture dramatically reduces hallucination rates, but introduces additional system complexity. For applications with high accuracy requirements, this investment is worthwhile.

Ask a Question

Related Terms

Useful Resources

Claude API Status → Model Pricing → Prompt Playground → Token Counter → MCP Servers → LLM Benchmarks → Model Comparison →