Hallucination
A hallucination is an AI-generated output presented as fact that is not actually true — a fabricated citation, invented person, made-up API endpoint, or wrong number stated confidently.
Hallucinations are the canonical AI failure mode. They happen because language models are trained to produce plausible-sounding text, not verified text. When the model doesn't know, it doesn't always say so — it sometimes generates something that fits the pattern of an answer. The modern mitigations are: retrieval (give the model the source documents), tool use (let it look things up instead of guessing), structured output (force it to fill known fields), and verification passes (a second model checks the first). None eliminates hallucination; together they make it rare enough for production.
Example
Asked for the citation for a legal argument, an unguarded model invents 'Smith v. Jones, 2014, 9th Cir.' that doesn't exist. The lawyer files it. The judge notices. This actually happened.
How OpenLabor uses it
OpenLabor employees ground answers in your data using retrieval and tool calls, and surface confidence so you know when to verify.
Can hallucinations be eliminated?
Not entirely with current model architectures. But in well-designed systems they can be made rare and easy to catch.
Why don't models just say 'I don't know'?
They sometimes do, and modern models do it more. But the training objective rewards confident, fluent output, which biases toward an answer.
How do I detect hallucinations in production?
Verification models, source-attribution requirements, output schemas with required citation fields, and human spot-checks on a sampled basis.
Related: guardrails, rag, tool-use.
AI Labor Glossary