In language models, hallucination means generating content that is fluent, confident, and plausible but factually incorrect or unsupported. The model presents fiction as fact. Examples include inventing case law citations, fabricating statistics, misattributing quotes, or describing features that do not exist in a product. Hallucination is not random noise. It emerges from how LLMs are trained and optimized: to produce probable text, not verified truth.

Next-token prediction objective: Models are trained to predict likely continuations, not to verify claims against ground truth. Plausible-sounding wrong answers can score as highly probable. Training data gaps: When the model lacks knowledge about a topic, it fills gaps with statistically likely fabrications rather than saying "I don't know." Alignment pressure: RLHF and preference tuning reward helpful, complete answers. Models learn that confident responses score better than cautious uncertainty. Context overload: In long prompts, models may misread or ignore retrieved context and fall back on parametric memory. Reasoning limits: Complex multi-step tasks expose failures where the model loses thread and invents connecting facts.

Factual hallucination: Wrong dates, numbers, names, or events stated as fact. Citation hallucination: Fabricated references, URLs, or paper titles that look real. Logical hallucination: Invalid reasoning chains that reach incorrect conclusions through plausible-sounding steps. Context hallucination: Ignoring provided documents and answering from outdated training memory instead. Sycophantic hallucination: Agreeing with incorrect premises in the user prompt and building false narratives on top.

RAG with citation: Ground answers in retrieved documents and require source attribution. Confidence calibration: Train models to express uncertainty and refuse when evidence is insufficient. Verification loops: Secondary models or rules check outputs against knowledge bases before delivery. Human-in-the-loop: Expert review for high-stakes outputs in legal, medical, and financial workflows. Domain-specific fine-tuning: Models trained on verified domain data hallucinate less on in-domain tasks. Structured output: Force JSON or schema-constrained responses to reduce free-form confabulation. Evaluation: Continuous benchmarking on faithfulness metrics, not just fluency.

Hallucination is the primary barrier to using general-purpose LLMs for professional work without additional architecture. Regulated industries cannot deploy systems that invent legal citations, clinical facts, or financial figures. The enterprise response is not to wait for zero-hallucination models. It is to build systems where: Knowledge is retrieved and cited, not memorized. Domain models are fine-tuned on verified corpora. Workflows include human review gates for material decisions. Task-specific agents maintain context across long interactions without degrading. Post-LLM architectures that embed reasoning inside governed workflows address hallucination structurally, not just through better prompts.