[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"glossary-hallucination::en":3,"gloss-cluster-hallucination::en":23,"gloss-next-hallucination::en":9},{"slug":4,"category":5,"name":6,"definition":7,"meta_desc":8,"faq":9,"schema_markup":9,"related":10},"hallucination","core-ai","Hallucination","A hallucination is when an LLM generates text that is fluent, confident, and grammatically correct but factually false, fabricated, or unsupported by any real source — inventing a citation that doesn't exist, stating a wrong statistic, or confidently describing an API method that was never built. Hallucinations happen because LLMs are next-token predictors optimized to produce plausible-sounding text, not fact-checkers with a ground-truth database; when a model doesn't \"know\" an answer, it doesn't reliably say \"I don't know\" — it often generates the most statistically likely-sounding answer instead, which can be entirely invented. This is the single biggest trust and liability risk for SaaS builders shipping AI features, especially in regulated or high-stakes contexts (legal, medical, financial, support). A well-documented real-world example: a legal team using ChatGPT to draft a court filing submitted a brief citing several court cases that the model had completely fabricated, complete with plausible-sounding case names and docket numbers — the cases did not exist, resulting in sanctions. Mitigations builders actually use in production: grounding outputs in retrieved source documents via RAG (and instructing the model to answer only from provided context, refusing when the context doesn't contain the answer); lowering temperature for factual tasks to reduce creative drift; adding citation requirements so every claim links to a source the user can verify; using guardrails or a second \"verifier\" LLM call to check the first model's output against retrieved facts before showing it to a user; and clear UX signaling (confidence indicators, \"AI-generated, please verify\" disclaimers) so users don't over-trust output. No current mitigation eliminates hallucination entirely — it's a property of how these models generate text, not a bug that gets patched out — so production systems should assume some hallucination rate and design verification\u002Ffallback paths accordingly, especially wherever wrong output has real cost. Hallucination rates vary meaningfully by task type and are worth benchmarking before shipping: open-ended creative tasks tolerate some invention by design, while factual lookup, numeric extraction, and citation tasks demand near-zero hallucination tolerance because users will act on the output as ground truth. A useful production pattern is a self-consistency check — asking the model the same question multiple times (or via slightly different phrasings) and flagging disagreement between answers as a signal of low confidence, since a model that's genuinely certain about a fact tends to answer consistently, while a model that's hallucinating often varies its fabricated answer between attempts.","A hallucination is a confident, fluent LLM output that is factually wrong or fabricated — the central reliability risk in production AI features.",null,[11,14,17,20],{"slug":12,"name":13},"grounding","Grounding",{"slug":15,"name":16},"guardrails","Guardrails",{"slug":18,"name":19},"retrieval-augmented-generation","Retrieval-Augmented Generation (RAG)",{"slug":21,"name":22},"temperature","Temperature",[24,28,32,36,39,42,45,48,51,54,57,60],{"slug":25,"category":5,"name":26,"updated_at":27},"agentic","Agentic AI","2026-08-24T02:46:36+00:00",{"slug":29,"category":5,"name":30,"updated_at":31},"alignment-tax","Alignment Tax","2026-08-24T02:46:37+00:00",{"slug":33,"category":5,"name":34,"updated_at":35},"artificial-intelligence","Artificial Intelligence (AI)","2026-08-24T02:46:38+00:00",{"slug":37,"category":5,"name":38,"updated_at":27},"attention","Attention",{"slug":40,"category":5,"name":41,"updated_at":35},"beam-search","Beam Search",{"slug":43,"category":5,"name":44,"updated_at":31},"benchmark-contamination","Benchmark Contamination",{"slug":46,"category":5,"name":47,"updated_at":31},"catastrophic-forgetting","Catastrophic Forgetting",{"slug":49,"category":5,"name":50,"updated_at":35},"computer-vision","Computer Vision",{"slug":52,"category":5,"name":53,"updated_at":31},"constitutional-ai","Constitutional AI",{"slug":55,"category":5,"name":56,"updated_at":27},"context-window","Context Window",{"slug":58,"category":5,"name":59,"updated_at":35},"deep-learning","Deep Learning",{"slug":61,"category":5,"name":62,"updated_at":27},"diffusion-model","Diffusion Model"]