A convincing answer is not necessarily a factual one
An AI hallucination is content that a model presents as an answer but that is unsupported or false. It might be a made-up book title, an incorrect date, a citation that does not exist, or a confident explanation that misses a crucial exception. The term does not mean the model is deliberately deceiving you. It describes a mismatch between a plausible output and the facts.
This matters because polished writing can hide mistakes. A model may use the right terminology and produce a tidy explanation while getting a name or number wrong. Treat fluency as a feature of the writing, not evidence that the underlying claim has been checked.
Why hallucinations happen
Many AI assistants generate text by predicting what should come next based on patterns learned during training and the context in the conversation. That process helps them explain familiar ideas and draft useful text, but it does not automatically verify each sentence against a trustworthy record. When the prompt asks for a specific fact the model cannot reliably determine, it may still produce an answer that fits the expected shape.
Several conditions make errors more likely: an ambiguous question, missing context, a niche subject, or a request for precise details such as page numbers and quotations. Information can also change after a model’s training or differ by place and circumstance. Access to a search or document tool can help, but it does not guarantee accuracy: the tool may return weak material, or the model may misread an otherwise sound source. A model’s expression of confidence is therefore not a substitute for evidence.
Separate the useful explanation from the claims you must verify
Start by identifying what the answer is doing. A general explanation may be useful as a starting point even when you would not trust it for exact details. Mark claims that could be checked independently: names, dates, figures, quotations, product features, technical requirements and statements about what a source says. The more specific or consequential a claim is, the stronger the reason to verify it.
You can ask the assistant to separate established points from assumptions and to say what information would resolve uncertainty. That can make its answer easier to inspect, but do not treat its self-assessment as proof. Asking “Are you sure?” may simply produce a more confident restatement. A better follow-up is: “Which parts of this answer require independent verification, and what would I need to check?”
Use a simple verification workflow
First, break the answer into individual claims rather than checking the whole response at once. Next, look for an appropriate source for each important claim. For a product feature, consult its documentation; for a quotation, find the original text; for a calculation, inspect the inputs and redo the arithmetic. If a claim depends on a location, version or time period, confirm that the source matches those conditions.
Then compare the claim with what the source actually says. Check whether a cited page exists, whether it supports the specific statement, and whether the answer leaves out a qualification. A link or a convincing-looking citation is only a pointer to inspect, not confirmation. If you cannot find adequate support, label the point unverified, remove it, or rewrite it more cautiously. For important decisions, do not let an unchecked AI summary stand in for the underlying material.
Prompt for answers that are easier to check
Good prompting cannot eliminate hallucinations, but it can make errors easier to catch. Define the scope of the question and supply relevant context. If you provide a document, ask the model to answer only from that document, identify the passage supporting each key point, and explicitly say when the document does not contain the answer. You still need to compare its response with the text: models can misinterpret passages or attribute claims to the wrong section.
For open-ended questions, try: “Give a brief answer, list the factual claims that need checking, and flag anything you are uncertain about. Do not invent citations.” For a draft you plan to publish, keep a separate verification pass for names, numbers and source-dependent statements. The practical goal is not to make an AI answer sound certain. It is to use the model for what it does well while keeping responsibility for factual checks visible.
