What is AI Hallucination? | AI Glossary
TLDR
AI hallucination is when a language model generates information that sounds confident and plausible but is factually incorrect, invented, or entirely unsupported by its training data.
The term "hallucination" in AI comes from psychology, where it describes perceiving something that isn't there. In language models, hallucination refers to the model producing content it presents as fact but that has no basis in reality: a CEO who doesn't exist, a scientific study that was never published, a legal case with a fabricated citation, or historical events that didn't happen.
Hallucination is not a bug in the traditional sense. It is a fundamental property of how large language models work. These models generate text by predicting what token is most likely to follow the previous one, based on patterns learned during training. They are very good at producing text that sounds correct. They do not have a separate fact-checking mechanism that verifies whether what they are saying is true.
Hallucination rates vary significantly between models and task types. Factual lookups, citations, names, dates, and statistics are the highest-risk categories. Creative writing, summarization of content provided in the conversation, and reasoning tasks generally have lower hallucination risk. Tools like Perplexity that ground responses in real-time web search reduce but do not eliminate hallucination.
In practice
Fabricated citations
Asking an AI to list academic papers on a topic and receiving a list of papers with plausible-sounding titles, authors, and journals that do not actually exist. This is one of the most common and dangerous forms of hallucination.
Incorrect statistics
An AI confidently stating that "67% of users prefer X over Y" with no source, generating a number that sounds authoritative but has no factual basis.
Invented product features
Asking an AI about a software product and receiving a description of features that sound real but were never actually built or released.
Frequently asked questions
Why do AI models hallucinate?+
Language models generate text by predicting the most statistically likely next word or phrase based on patterns in their training data. They do not have a separate fact-checking layer. When the model doesn't know something, it may still generate a confident-sounding response based on what seems plausible in context. This produces hallucinations.
Which AI model hallucinates the least?+
Perplexity hallucinates significantly less than ChatGPT or Claude on factual lookups because it grounds every response in real-time web search. Among pure language models, Claude and GPT-4 are generally considered more reliable than smaller models, but all language models hallucinate to some degree. No model is fully hallucination-free.
How do you prevent AI hallucination?+
Reduce hallucination risk by: asking for sources and verifying them independently; providing the information you want the AI to work with (rather than asking it to recall facts); using grounded tools like Perplexity for factual research; and treating AI-generated statistics, names, and citations as unverified drafts until confirmed.
Is hallucination getting better over time?+
Yes, meaningfully. Newer models hallucinate less frequently than their predecessors, and techniques like retrieval-augmented generation (RAG), tool use, and web search integration have significantly reduced factual errors in production applications. However, hallucination has not been eliminated and is unlikely to be fully solved with current architectures.
Bottom line
AI hallucination is when a language model generates information that sounds confident and plausible but is factually incorrect, invented, or entirely unsupported by its training data.
Put it into practice
Prompt packages that apply these concepts directly.