Why Claude makes things up — and how to prevent it

"Hallucination" is the wrong word

The term implies random confusion. It's not. When Claude invents a citation or gives a wrong date with total confidence, it's doing exactly what it was trained to do: predict the most probable next token. The model was trained to complete text, not to verify truth.

The 3 real types of hallucination

• Type 1 — Information not in training: Claude doesn't know your company's internal policies. It generates something plausible based on similar companies. RAG solves this directly.
• Type 2 — Information in training but degraded: Claude saw correct information, but the model didn't store it perfectly. It recovers it distorted. RAG helps here too.
• Type 3 — Correct facts, wrong reasoning: Claude has the right facts but the logic chain fails. The most dangerous type because everything sounds correct. RAG doesn't help. Chain-of-Thought does.

Why Claude doesn't say "I don't know"

Saying "I don't know" requires metacognition: knowing that you don't know something. That implies having an internal model of your own knowledge. Base LLMs don't have that.

What does happen: during fine-tuning and RLHF, humans penalized wrong responses and rewarded uncertainty acknowledgment. So Claude learned to say "I'm not sure" in some contexts. But it's learned surface behavior, not a real verification mechanism.

5 practical strategies to reduce hallucination

• 1. Give Claude the source: paste the relevant document or data. Tell it "answer only based on the following context." Forces grounding.
• 2. Ask it to cite: "include the paragraph you based this on." Uncitable claims become visible.
• 3. Chain-of-Thought for reasoning: "think step by step before answering." Forces the logic to be explicit — easier to catch errors.
• 4. Explicit scope: "if you don't know, say so explicitly instead of guessing." Works partially, not perfectly.
• 5. Cross-check critical answers: for high-stakes outputs (legal, medical, financial), use Claude to draft and verify with real sources.

When hallucination actually doesn't matter

Hallucination is a problem when you need factual accuracy. For many real tasks, it's irrelevant: drafting the structure of a document (you'll fill in the facts), brainstorming ideas, writing code to a spec you provide, transforming formats. Understanding when accuracy is critical vs. when fluency is enough saves a lot of over-engineering.