What Are AI Hallucinations? Why AI Makes Things Up

What Are AI Hallucinations?

AI hallucinations (also called confabulations) occur when a large language model generates information that sounds plausible and is stated confidently, but is factually incorrect, fabricated, or misleading. The term borrows from psychology, where hallucination refers to perception without a corresponding external stimulus — the AI is "perceiving" information that doesn't exist.

Hallucinations are not bugs that can be simply patched. They arise from the fundamental nature of how LLMs work — and understanding why they occur helps you use AI tools more safely and effectively.

Why Do LLMs Hallucinate?

Large language models are trained to predict the next token in a sequence — to generate text that is statistically plausible given what came before. They are not retrieving facts from a database; they are generating text that resembles what appears in their training data.

This has several implications:

• No internal "truth checker": The model has no mechanism to verify whether a statement is factually accurate before generating it. It optimizes for plausibility, not truth.
• Pattern interpolation and extrapolation: The model fills in gaps by generating plausible-sounding patterns, even when it has insufficient training data for a specific question.
• Confident generation regardless of certainty: The same text generation mechanism produces highly confident-sounding sentences whether the model "knows" the answer well or is essentially guessing.
• Training on imperfect data: If training data contained errors, the model may reproduce them. If training data had little coverage of a topic, the model may extrapolate incorrectly from related patterns.

Common Types of Hallucinations

Fabricated Citations

Among the most notorious hallucination types: generating fake academic papers with plausible-sounding titles, authors, and journal names that don't exist. In 2023, a lawyer submitted a legal brief citing cases generated by ChatGPT — none of them were real, and the lawyer faced sanctions. Always verify citations from AI output.

False Factual Claims

Incorrect statements presented confidently — wrong dates, incorrect statistics, mistaken biographical details, inaccurate scientific claims. These are particularly dangerous because they often sound authoritative.

Confident Extrapolation Beyond Knowledge Cutoff

LLMs have training data cutoffs and don't know events that occurred after them. They may answer questions about recent events confidently by extrapolating from older patterns, producing plausible-sounding but incorrect information.

Inconsistent Self-Contradiction

Models may contradict themselves within the same conversation or between sessions, because they don't maintain a consistent internal knowledge state — each generation is probabilistic.

Sycophantic Hallucination

When users push back on correct answers, models may agree with the user's incorrect assertion to avoid conflict — a form of hallucination driven by training to produce agreeable responses.

Hallucination Rates: How Common Are They?

Hallucination rates vary significantly by task, model, and topic:

• For well-documented topics with lots of training data: relatively rare
• For obscure topics, recent events, or very specific details (exact quotes, precise statistics): much more common
• For tasks requiring multi-step reasoning: errors can compound

Studies have found that in some domains (medical, legal, scientific fact verification), leading models hallucinate on 5–20% of specific factual queries. This is far too high for high-stakes use without verification.

Mitigation Strategies

Retrieval-Augmented Generation (RAG)

RAG systems provide LLMs with relevant retrieved documents at query time, grounding responses in verified sources rather than relying purely on parametric memory. The model is instructed to base its answer on the provided documents. This dramatically reduces hallucinations for factual queries in well-defined domains.

Prompting for Uncertainty

Ask the model to express uncertainty: "If you're not certain about any facts, say so" or "Tell me what you're confident about vs. what I should verify." Well-instructed models can partially calibrate their uncertainty expression, though this doesn't eliminate hallucinations.

Verification Habits

The safest approach: treat LLM outputs as a knowledgeable but sometimes unreliable colleague who needs to be checked on specifics. Use LLMs for reasoning, drafting, and brainstorming. Verify specific facts, statistics, citations, and claims from authoritative sources before relying on them.