Hallucination

A hallucination is a confident, fluent output from an AI language model that is factually incorrect, fabricated, or not grounded in its input or knowledge, produced without any indication of uncertainty.

In the context of large language models (LLMs), hallucination refers to the generation of information that is plausible-sounding but factually wrong, unsupported by source material, or entirely fabricated. Unlike a simple error, a hallucination is typically presented with the same confident, fluent style as accurate output, making it difficult for users to identify without independent verification. The term is borrowed from psychology but describes a distinct computational phenomenon rooted in how language models are trained.

Hallucinations arise because LLMs are optimized to produce fluent, coherent text continuations rather than to verify factual accuracy. When a model encounters a query on a topic where its training data is sparse, contradictory, or absent, it generates a plausible-sounding completion that fills in gaps with invented details. Retrieval-augmented generation (RAG) systems reduce hallucination by anchoring outputs to retrieved documents but do not eliminate it. Additional mitigations include tool use (allowing models to verify claims via search or APIs), calibration training, and citation-constrained generation.

Hallucinations are a significant practical barrier to deploying LLMs in high-stakes domains. Documented cases include fabricated legal citations submitted to US federal courts in 2023 (the Mata v. Avianca case), invented scientific references in academic contexts, and incorrect medical information delivered with apparent confidence. These failures led to formal guidelines from legal and medical professional bodies advising against unreviewed AI-generated factual claims.

By 2026, hallucination remains an active research problem. Benchmarks including TruthfulQA, HaluEval, and FActScore quantify hallucination rates across models. Leading commercial models have substantially reduced hallucination frequency compared to 2022-2023 baselines through improved alignment training, RAG integration, and grounded generation techniques. However, no production model has achieved hallucination-free operation, and rates vary significantly by domain, query specificity, and whether external retrieval is available.