AI Hallucination Detector

LLM hallucination refers to when an AI model generates text that sounds confident and plausible but contains false, fabricated, or unverifiable information. This can include made-up citations, incorrect facts, non-existent URLs, or fabricated statistics. Hallucinations occur because LLMs predict text based on patterns rather than retrieving verified facts from a knowledge base.

This tool scans for common hallucination risk signals: overly precise statistics without sources, citation patterns (e.g., '[Author et al., 2023]'), URL mentions that can't be verified, absolute confidence language ('definitely', 'certainly', 'always'), vague hedge phrases masking uncertainty, and numeric precision anomalies like suspiciously round or unverifiable percentages.

Key strategies: (1) Retrieval Augmented Generation (RAG) — ground the model in verified documents. (2) Ask the model to cite sources and then verify them. (3) Use temperature 0 for factual queries. (4) Prompt the model to say 'I don't know' when uncertain. (5) Add a verification step with a second LLM call asking 'Is this claim verifiable?'. (6) Use structured output formats to constrain responses.

All LLMs hallucinate to some degree. Smaller models (7B-13B parameters) tend to hallucinate more than frontier models. Models optimized for factual accuracy like GPT-4 with browsing, Claude with documents, or Gemini with Google Search integration hallucinate less on verifiable claims. Fine-tuned models trained on domain-specific data also reduce hallucination in their target domain.

These terms are often used interchangeably in AI contexts. Technically, confabulation (from neuroscience) refers to generating plausible-sounding but false memories without intent to deceive — which maps well to LLM behavior. 'Hallucination' has become the industry standard term for LLM-generated falsehoods. Both describe the same phenomenon: confident generation of unverifiable or false information.