The analogy
If you measure a room with a tape measure just once and misread the mark, you start with a wrong number. If you measure it five times and take the value that comes up most often, the error of the single reading cancels out in the bunch. Or think of a hard question put to five competent people: you go with the answer the majority gives, not the one from a single person.
Self-consistency is this majority vote applied to the AI. You don't trust a single attempt, which might have taken a wrong path by chance: you gather several and follow the one they converge on.
How it really works
The same question, asked several times, can lead the AI down different reasoning paths, because of the dose of randomness in the choice of words. On a problem with a verifiable or clearly more likely answer, generating several independent solutions and taking the majority one beats the single attempt: the wrong paths tend to diverge each on its own, while the right one keeps coming back. It's a way to reduce the weight of an occasional error, not to make it omniscient.
What you can do in practice
- For a calculation or a delicate decision, ask the same question in three new, separate chats, then compare the answers and follow the one that recurs.
- Alternatively, in a single request: "give me three independent solutions to this problem, then tell me which is the most recurring and why."
- Go with the answer the attempts converge on; if instead they all diverge, it's the signal that the question is hard or poorly posed, and needs clarifying.
- Keep this method for problems with a correct answer, not for opinions.
A common misconception
People think the majority answer is always the true one. It's not so. If the AI has a systematic error — a wrong idea rooted in the data it was trained on — it repeats it in every version, and the majority will confidently give you the same wrong thing. Self-consistency reduces random errors, those of the single unlucky attempt, not the deep, recurring ones. The facts still need verifying.
Frequently asked questions
How do I use it without technical tools?
Just repeat the question in new, independent chats, or ask in one go for several separate solutions and their point of convergence. There's no need to program anything: it's a technique you apply with the normal conversation.
Does it also work for opinions or advice?
No, it makes sense only where there's a correct answer to bring out: a calculation, a deduction, a fact. On opinions and tastes, the "majority" of answers doesn't mean "more correct," it just means "more frequent."
How many times is it worth repeating?
Generally three to five attempts are enough to bring out a convergence. Beyond that, the gain is small compared with the effort. If already at three the answers are all different, the problem isn't the number of attempts.
If the AI always repeats the same answer, then is it certainly correct?
No, and it's the illusion to avoid. An AI can be consistently confident and consistently wrong: internal consistency reduces random noise, but it doesn't guarantee truth, especially if the error is rooted. The converging answer is a good candidate, not a proof. For the facts that matter, the confirmation comes from an external source, not from repetition.