How to Reduce LLM Hallucinations: Model Selection, Web Search, and Verification Agents

Choose Models with Lower Hallucination Rate

Hallucination is generated information that looks plausible but is factually wrong. The hallucination rate is the proportion of such erroneous outputs among all generated content; a lower rate usually indicates higher factual consistency and reliability.

Hallucination: generated information that looks reasonable but is actually wrong. Hallucination rate: the share of erroneous outputs among all generated content.

Hallucination Leaderboard: https://github.com/vectara/hallucination-leaderboard

Moonshot AI Kimi‑K2‑Instruct – 1.1%

Google Gemini‑2.5‑Pro‑Exp‑0325 – 1.1%

OpenAI GPT‑5‑high – 1.4%

Qwen3‑Max‑Preview – 3.8%

DeepSeek‑V3 – 3.9%

Claude‑4‑Sonnet – 4.5%

MoonshotAI Kimi‑K2‑Instruct‑0905 – 6.2%

DeepSeek‑R1 – 14.3%

DeepSeek‑R1 tends to fabricate nonexistent books, reports, or paper titles, invent data, and use overly flowery language, while DeepSeek‑V3 shows a much lower hallucination rate.

Enable Web Search

Missing factual grounding is a major cause of hallucination. Activate the model’s internet‑search capability or employ dedicated AI‑search tools that provide supporting citations.

Tools such as Mita AI Search or Perplexity answer queries and return source references.

Ask from a Different Angle

Re‑phrasing a question can surface inconsistencies. Example with the claim about Geoffrey Hinton:

“Is Geoffrey Hinton the father of artificial intelligence?”

“Who is considered the father of artificial intelligence?”

Comparing answers from different phrasings helps cross‑validate the information.

Build an Information Verification Assistant

Define a prompt that instructs the assistant to assess supplied text, perform a web search, judge truthfulness, and explain the reasoning.

You are a professional misinformation detection expert. Your task is to analyze the text I provide and judge its factual accuracy.

Please follow these steps:
1. **Assess**: Carefully read the supplied text.
2. **Web Search**: Perform an online search for related information.
3. **Judge**: Determine whether the information is **true**, **false**, or **cannot be verified**.
4. **Explain**: Briefly state the reasons for your judgment. If false, point out the inaccuracies; if unverifiable, explain why.

**Output format:**
**Judgment:** [True/False/Cannot Verify]
**Explanation:** [Bullet‑pointed reasons]

Implementation in Claude:

Create a new Project in Claude and paste the prompt into the Instructions field.

When a statement needs verification, send it to the “information verification assistant”.

For the Hinton example the assistant returns that Geoffrey Hinton’s formal title is “father of artificial intelligence”, not “the father of AI”.