GPT-5.6 Sol has a 92.2% hallucination rate. Claude Fable 5.1: 72.6%. Why do AI models, even the most advanced ones, hallucinate so much?

Artificial Analysis tests AI models on 6,000 factual questions, closed book: no web search, no documents, everything from memory.

They track 2 rates per model:
- ✅ Accuracy: the share of right answers.
- 👻 Hallucination rate: when the model doesn't have the right answer and makes one up.

The latest results (max settings):
- Claude Fable 5.1: ✅ 67%, 👻 73%
- GPT-6 Astra: ✅ 63%, 👻 51%
- Gemini 3.1 Pro: ✅ 55%, 👻 51%
- Grok 4.7: ✅ 47%, 👻 29%
- Gemini 4 Argon (not public yet): ✅ 49.9%, 👻 15.1%

Fable 5.1 attempts 93% of the questions. It knows more than any other model, but when it doesn't know, it rarely says so. It makes up an answer instead.

So why not train models to say "I don't know"?

They do. But the problem runs deeper, in the very way these models work.

Models generate their answers word after word, they don't look them up in a database. What they learn isn't stored as facts, it's blended into billions of internal settings, locked in at the end of training. For each word, those settings calculate, based on the context, the probability of every possible next word, and the model picks one of the most likely. And the most likely may not reflect reality.

By default, a model says "I don't know". But two things weaken that reflex:
- Its generative nature: conversing and arguing also means creating combinations of words that are not necessarily backed by facts.
- Recognition: recognizing a word doesn't mean knowing everything about it, but it's enough to switch the reflex off and make up an answer from scratch.

You get the idea: hallucinations are a byproduct of how generative models work. And the more a model tries to answer everything, the more it makes things up when it doesn't know.

How to reduce your model's hallucinations:

1/ Work from the most complete context possible, with clear sources and targeted questions. The more precise the context, the more precise the answer.

2/ Force document or web search, and explicitly tell the model not to work from memory.

3/ Cite your sources and ask for a quote behind every claim.

4/ Ask the question several times, or ask several models.

5/ Pick the model for the task. Some are more cautious than others.

Anthropic's documentation says that allowing the model to say "I don't know" is enough to drastically reduce false information. I haven't tested it yet.