A hallucination is an output that sounds plausible but is not supported by the available evidence. It may invent a citation, merge two people, use an outdated policy, misread a table, or confidently fill a gap the source never answered. Fluency makes these errors dangerous because presentation quality can be mistaken for reliability.
The behavior follows from how generative models work. They produce likely continuations under the supplied instructions and context; they do not consult an internal ledger that guarantees every sentence is true. Better models can reduce error rates, but no general-purpose system removes the need for controls.
Risk depends on the job. A surprising metaphor in a brainstorming session may be useful. A fabricated number in a financial report is not. Define which claims require evidence, what uncertainty is acceptable, and who bears the cost of a mistake before choosing the level of automation.
Grounding is the first defense. Provide current, relevant sources and instruct the model to use them. Retrieval-augmented generation can bring private or changing documents into the request. Grounding still needs evaluation: retrieving an irrelevant or outdated passage can produce a well-cited error.
Require traceability at the claim level. A source list at the end of an answer does not show which statement each source supports. Ask for citations next to factual claims, preserve document identifiers, and make it easy for a reviewer to open the referenced passage. Never treat a model-generated citation as verified until the destination and support are checked.
Give the model permission to abstain. If the evidence does not answer the question, the correct output may be “insufficient information,” a request for a missing document, or a list of unresolved points. An application that rewards completion at any cost will encourage confident guessing.
Validation should match the output. Structured data can be checked against a schema, range, or known vocabulary. Calculations can be recomputed. Names and identifiers can be matched to a database. Quotes can be located in the source text. Code can be compiled and tested. These deterministic checks catch failures that another free-form model review may repeat.
Separate extraction from interpretation for difficult documents. First capture the relevant values and passages with coordinates or citations. Then ask the model to explain them. This creates an intermediate artifact a human or test can inspect and reduces the chance that a polished narrative hides a reading error.
Use independent review for high-impact decisions. The reviewer needs the original evidence, not just the answer. Human oversight is most useful when the person has authority to stop the workflow and enough time to examine the risky claims; a rubber-stamp approval box adds little protection.
Monitor production failures rather than relying only on launch evaluations. Save representative prompts, retrieved evidence, outputs, corrections, and model versions within privacy limits. Group errors by cause: missing source, wrong retrieval, prompt ambiguity, unsupported synthesis, stale data, or validation failure. Each cause calls for a different fix.
Communicate uncertainty honestly in the interface. Show verification dates, source coverage, and whether the answer was generated from user-provided documents or general model knowledge. Avoid badges that imply a guarantee the system cannot make.
The practical goal is not to promise that an AI never hallucinates. It is to make unsupported claims harder to produce, easier to detect, cheaper to correct, and less likely to reach a consequential decision unnoticed.
AI Tools Radar separates product facts, editorial judgment, and commercial placement. Updated facts retain their verification date.