How do I stop AI from hallucinating?
You cannot make a single AI model stop hallucinating, because a hallucination looks identical to a correct answer from the inside: both are fluent, both are confident. What you can do is stop hallucinations from reaching your final answer. Ask the same question of several different AI models independently, then accept only the claims that name a source you can open and that survive the other models trying to disprove them. A made-up fact almost never survives that, because it has no source to point to. This is the Council Method, and it catches hallucinations instead of trying to prevent them.
Why "stop hallucinating" is the wrong goal
A hallucination is not a glitch the model knows it is having. The model has no internal signal that separates "I know this" from "I am completing a plausible sentence." So asking a single model to stop hallucinating is like asking a person to stop being wrong about things they are certain of. The confidence is the problem, and the confidence is baked in.
The move that works is to change what you trust. Stop trusting one model's answer. Start trusting what survives a room.
The three controls that catch hallucinations
- Independence. Ask two or three different models (from different companies, ideally) the same question, without letting any of them see the others' answers first. Shared blind spots are real, but a hallucination invented by one model usually is not invented the same way by another.
- Provenance before consensus. For every load-bearing claim, require a source you can actually open. A hallucinated fact has no real source, so this is where most of them die. Do not accept "agreement" as a substitute. Several models repeating something is not evidence.
- A human verdict. You read what survived, you decide, you sign it. The method does not remove you. It gives you a checked answer to sign instead of a confident guess.
A real example
In a live run on this site, five models were asked to interpret a scrap of code. Four read it correctly. One invented an entire company, complete with fake pricing and citation markers, that had nothing to do with the question. The other four models, shown all five answers, named exactly why that answer was baseless and threw it out. No human expertise was needed to catch it. The room caught it. That is the whole idea: read the run, or run your own question.
The full method, including how to run it with one model when you have to, is in Let the AI Be Smart, and the free Council Protocol gives you the exact prompts.
Go deeper: this site's hub page on the Council Method is the full definition. Related questions: What is an AI hallucination?, How do I know if an AI answer is correct?, What is the Council Method?.