# What is a hallucination in an AI agent?

A hallucination is an agent stating something its sources don't support. That single failure splits into three properties, and each one can be checked on its own. Faithfulness asks whether the answer represents the source it cites, which fails when the agent retrieves the right document and still misquotes it or inverts a number. Groundedness asks whether the answer came from that source at all, or from the model's own prior with the source sitting unused, which fails silently whenever the prior happens to be right. Context presence asks whether the needed information was retrieved in the first place, which fails before generation even starts.

Checking only the final answer conflates all three. A grader that reads output alone can catch a faithfulness break, but it can't tell you whether the context was there, used, or absent, and each of those points to a different place to fix.

---

Source: https://tessary.ai/answers/hallucinations/what-is-a-hallucination-in-an-ai-agent
More on Hallucinations: https://tessary.ai/answers/hallucinations
From Tessary, agent reliability for AI agents in production: https://tessary.ai