# Hallucinations When AI sounds confident but gets it wrong. Track: Safe and Responsible Use ## What is it? An AI hallucination is an answer that sounds confident but is wrong, unsupported, or invented. That is the dangerous part. It may not look wrong. The writing can be smooth. The tone can be calm. The answer can include names, dates, sources, or numbers. But smooth writing is not proof. A hallucination can be a fake citation, a wrong policy, a made up feature, or a confident answer that the source does not support. ## Why it matters At work, a wrong AI answer does not stay inside the chat window. Someone copies it into an email. Someone adds it to a slide. Someone sends it to a client. Someone uses it to support a decision. That is where damage starts. Hallucinations matter most when the output affects money, legal exposure, customers, safety, security, reputation, or senior management decisions. For important work, verification is not optional. ## How it works AI models generate likely text. They do not always know whether something is true. They look at the prompt, the context, and patterns they have learned, then produce an answer that seems likely. When the model lacks enough grounding, it may still try to answer. Instead of stopping cleanly, it can fill gaps with language that sounds reasonable. The model may not be lying. It is completing a pattern. The result can still be wrong. ## Analogy Imagine a junior employee preparing a briefing note late at night. They have part of the file, but not the full file. The deadline is close. They want the note to sound complete. So they fill in the missing parts from memory and make the sentences flow. The document reads well. The problem is that some parts are guessed. That is how hallucinations often feel. Polished enough to pass a quick glance. Dangerous when nobody checks. ## Example usage Hallucinations show up in normal work. A chatbot may give a customer the wrong policy information. A legal draft may include case citations that do not exist. A market summary may invent a company statistic. A product note may mention a feature that was never shipped. The lesson is not to avoid AI completely. The lesson is to use AI for speed while keeping human judgment on the important claims. ## How to use this Use AI answers in layers. First, ask for the draft, summary, or explanation. Second, ask what claims need verification. Third, check the important claims against real sources. For serious work, ask the AI to separate what comes from the source, what it is assuming, and what needs human review. That habit catches many problems before they reach a customer, manager, or board pack. ## Common mistake The common mistake is trusting tone. People see a confident answer and relax. That is human. We are trained to associate confidence with knowledge. AI breaks that instinct. A hallucinated answer can sound more confident than a careful expert. The expert may say, I need to check. The AI may answer immediately. Do not reward the wrong behavior. If the answer matters, check the source. ## Question to ask - **Risk**: Which parts of this answer would create damage if they were wrong? - **Evidence**: What source supports each important claim? - **Citations**: Are the sources real, current, and relevant to the exact statement? - **Final use**: What must I verify before sending this to a client, manager, regulator, or customer? ## Quick quiz What is the safest first reaction to an important AI answer? ## Flashcard **Question:** What is an AI hallucination? **Answer:** A hallucination is an AI answer that sounds confident but is wrong, unsupported, or invented.