Why Is the Answer Different Every Time?
Yesterday you asked it to name a milk-tea shop and got three artsy options. Today you ask the exact same thing, and the whole vibe has changed. Bad memory — or is it blowing you off? Neither. Click a few buttons and you can watch how each word actually arrives.
Because every time AI writes a word, it is rolling the dice: it picks one from a few high-probability candidates. That's intentional design, so answers stay natural, not stiff. When you need stable output, you can turn the randomness down.
Below is a fixed question, not one word changed. Tap "Ask again" and see how far three answers can drift.
AI writes a sentence by finishing the sentence: each time it only decides the next word. It scores every candidate first — high scores are more likely to be drawn, low scores less so — then it draws one. After that word, it draws again for the next. A sentence of a few dozen words is a few dozen draws. Two answers coming out identical is the rare event.
Don't take my word for it. Below is a real moment at one position: AI has already written "Name idea: Jasmine Milk," and is deciding the next word. The probability bars for four candidates are right here — you roll this die for it. Roll a few times, and watch the draw tally on the right.
You might ask: why not just pick the highest-scoring word every time, so the answer stays stable? People tried. It was terrible. Always picking the top score turns the answer into a stiff parrot: ask for a milk-tea shop name ten times, and it answers "Jasmine Milk Tea" ten times. The copy all sounds the same, and tasks like naming, poetry, and brainstorming fall apart. Let it draw the second- or third-place word, and the path suddenly widens — the language sounds more like a living person talking.
Creative work · Welcome the randomness
Naming, copy, brainstorming — randomness is where the ideas come from. Ask a few more times and you've hired a few extra colleagues for free. Just pick the best version.
Serious work · Turn randomness down
Contract clauses, code, data processing — if the answer drifts, things go wrong. Many products let you lower the randomness (often called the "temperature" parameter). At the lowest setting it almost always picks the top score. For a deeper look, see the advanced lesson on temperature.
One practical reminder: randomness also means a single answer only has so much reliability. When important facts are involved, ask the same question a few different ways. The parts that stay consistent are more trustworthy; the parts that keep changing should be checked against an authoritative source. Inconsistent answers and lying are two different things — but for you, the move is the same: verify either way.
✅ What this page wants to share with you
- Randomness is a feature: different answers mean the dice are working; a real fault is a different story
- Every word is one draw: high-probability words get drawn often; low-probability words still get a chance
- Welcome randomness for creative work, lower temperature for serious work: use it by the occasion
- For important facts, ask a few times and cross-check: what's consistent is more trustworthy; what's inconsistent, go verify