What Is AI Hallucination? Definition, Causes, and Mitigation

Think of an AI like an overly eager intern who is terrified of saying “I don’t know.” When asked a question they don’t have the answer to, instead of admitting ignorance, they invent a story that sounds professional—citing names, dates, and details that never existed—just to be helpful. The result reads smoothly and sounds authoritative, but none of it actually checks out.

That is roughly what happens when a large language model hallucinates. The model is not lying—it has no concept of truth—but it fills gaps in its training data with statistically plausible patterns, then presents them with complete confidence.

How hallucinations happen

Large language models predict the next word (or token) in a sequence based on patterns learned from billions of text examples during training. They calculate probabilities, not verify facts. When a question falls near the edge of what the model has seen—or beyond it—fluency wins over accuracy. The model produces a response that sounds right because it pattern-matches to linguistic structure, not because it has a fact-checking mechanism underneath.

MIT research published in January 2025 confirmed this: models use more confident language when hallucinating than when stating actual facts. The fluency is not a signal of truth—it is a feature of the generation process itself.

Types of hallucinations

Researchers divide hallucinations along two useful axes:

Intrinsic hallucinations contradict the source material provided in the prompt. If you paste a document and ask for a summary, and the summary includes details not in the document—or contradicts the document—that is an intrinsic hallucination. The model was faithful to its own pattern-matching, but unfaithful to the source.

Extrinsic hallucinations introduce information that cannot be verified from the input or from any external knowledge base. These are the fabricated court cases, invented studies, and imaginary company histories. The key danger: extrinsic hallucinations are often harder to catch because there is no source to check them against without doing independent research.

Common specific manifestations include:

• Factual recall errors—a model states a historical event occurred in the wrong year, or attributes a quote to the wrong person.
• Fabricated citations—a model citing a journal article, court case, or research paper that does not exist, complete with plausible titles and page numbers.
• Logical inconsistencies—a response that contradicts itself halfway through, or gives advice that conflicts with its own stated premises.
• Temporal disorientation—getting dates, sequences, or timelines wrong, especially for events near or after the training data cutoff.

Confabulation vs. hallucination

Some AI researchers prefer the medical term confabulation—a phenomenon in psychiatry where a patient with memory gaps fills them in with plausible but false narratives, not intending to deceive. Smith et al. (2023) argued in a peer-reviewed paper that this is a better metaphor for what language models do: generating incorrect but coherent story fragments to fill gaps in knowledge, with no intent behind it and no awareness of the error.

The term is gaining traction in academic literature—for example, in Cossio’s 2025 taxonomic survey—but hallucination remains the dominant industry term. For most practical purposes, both words describe the same phenomenon: confident, fluent output that is not anchored to real facts.

Why it matters

In December 2023, lawyers in Mata v. Avianca submitted an AI-generated legal brief to a federal court. The brief cited six cases that simply did not exist—fabricated by ChatGPT with full case names, docket numbers, and quotes. The judge sanctioned the attorneys. The case became the signature example of what happens when human professionals take AI output at face value.

Beyond courtrooms, hallucinations affect:

• Medical systems—where models generating incorrect drug interactions or dosages can pose direct patient safety risks.
• Business decisions—where market projections or company data generated by AI can look authoritative but lead organizations in the wrong direction.
• Consumer trust—where a single confident hallucination in an AI assistant can undermine credibility across the entire product.

The OWASP Top 10 for LLM Applications (2025 edition) categorizes this as LLM09: Misinformation, recognizing hallucination not merely as an inconvenience but as a security-relevant risk category.

How teams reduce hallucination

No technique eliminates hallucination entirely—it is a structural feature of probabilistic text generation. The goal is layered risk reduction:

• Retrieval-Augmented Generation (RAG) is currently the most widely adopted approach. RAG connects the model to a curated knowledge base or external source before generating an answer, so the model must base its response on retrieved evidence rather than relying solely on parametric memory. Newer hypergraph-based RAG methods (Hyper-RAG, published in Nature Communications in 2026) further reduce error rates by capturing multi-entity relationships that flat retrieval misses.
• Verification workflows—using a second model or automated system to check the first model’s outputs against source documents, question internal contradictions, or score the confidence of claims. Multi-model debate and LLM-as-judge patterns are increasingly used in production systems.
• Prompt engineering—techniques like chain-of-verification and cite-then-answer, which structure prompts so the model must retrieve and reason before generating, reducing fabrication by up to 36% in published studies.
• Human oversight—keeping humans in the loop for high-stakes decisions, particularly in regulatory, medical, legal, and financial contexts. Human review remains the most reliable backstop, especially where the cost of an uncaught error is high.

Hallucination detection and evaluation

Establishing how often and how badly a model hallucinates requires structured evaluation. The most widely used benchmarks include:

• HaluEval (Li et al., 2023): A large-scale evaluation framework with approximately 35,000 samples testing hallucination across question-answering, dialog, and summarization tasks.
• HalluLens (2025): A newer benchmark that distinguishes between intrinsic and extrinsic hallucinations and uses dynamic test-set generation to reduce data leakage—a known problem with static benchmarks.

A key methodological note: the widely referenced TruthfulQA benchmark (Lin et al., 2022) was designed to measure truthfulness, not hallucination directly. By 2025, critics noted it had become partially saturated (models trained on it) and contained errors in its gold answer set—making it a less reliable proxy than often assumed.