Hallucination

What it is

IBM describes AI hallucinations as outputs that sound plausible but are factually wrong, irrelevant or entirely fabricated. Examples include invented studies, URLs that don’t exist and wrong details about real people or companies. It’s a known property of language models, not a rare bug.

How it works

1. Models generate text that looks statistically likely. When they don’t have the answer, a plausible-sounding guess can come out.
2. Risk rises with obscure topics, precise numbers, citations and questions beyond the knowledge cutoff.
3. Anthropic’s guidance to reduce it:
   • Explicitly let the model say “I don’t know”.
   • For long documents, have it extract word-for-word quotes first and answer only from those quotes.
   • Have it cite a quote or source for each claim, and retract any claim it can’t support.

Benefits

• Knowing the failure mode lets you design checks into workflows.
• Grounding answers in your documents (RAG) reduces the risk but doesn’t remove it.
• Clear review rules let teams use AI drafts safely.

Limits and risks

• No current technique eliminates hallucinations. Anthropic says even the most advanced models can produce text that’s factually incorrect or inconsistent with the given context.
• Fabricated citations look real, so open every link.
• A support or sales bot that invents policies creates real liability.
• IonBin doesn’t quote hallucination-rate benchmarks. Rates depend on the task, and vendor figures aren’t independent, so test on your own material (see evals).