n8n's guide to catching AI hallucinations says no single check is enough. Layer them and give every flagged answer a place to go

A hallucination looks exactly like a normal answer, which is why it slips into the next workflow step.

What happened

n8n published a practical guide to hallucination detection in AI agent workflows. It splits hallucinations into four types: factual (claims that conflict with verifiable facts), contextual (answers not supported by the supplied context, a big issue in RAG), confabulations (gaps filled with invented detail) and self-contradictions.

Its core argument is that no single detector covers all four, so checks should be layered:

• Faithfulness checks break an answer into claims and compare each with the retrieved documents. Tools like DeepEval can score this; its faithfulness metric uses an LLM as a judge against the retrieval context.
• Self-consistency asks the same question several times and compares answers. n8n warns that runs can agree on the same wrong answer, so treat it as a signal, not proof.
• Source verification checks key claims against an authoritative database, API or documentation, because a faithful answer can still be wrong if the retrieved document is outdated.
• Confidence thresholds route high-confidence results forward, send ambiguous ones to more checks and hold high-risk cases for a human before an agent acts.

In n8n, these checks sit on the canvas next to the agent, an Error Trigger node can start a separate error workflow, and execution history helps debug past runs.

My take

The most useful line in the guide is that a flagged response needs somewhere to go. Most client builds I audit have no such path. The model answers, the next node fires, and a wrong value lands in the CRM.

Match the check to the cost of the mistake. An internal summary can pass with a light check. An email to a customer, or an update to a deal stage, should verify facts against the CRM record itself and stop for approval when confidence is low. Source verification is the cheapest win for CRM work, because the source of truth already exists: compare what the model wrote with the contact or deal fields before you save it.