Type a question into ChatGPT. Have a friend type the exact same question. Do the two of you get the same answer back?

No. But which parts change, and which parts don’t, turns out to be the useful question.

On September 21, 2026 we asked ChatGPT five questions, five times each, in 25 separate chats. Twenty-four of the 25 answers were worded differently. The facts inside them didn’t move once. And exactly one answer changed its mind.

The essentials

• ChatGPT doesn’t give everyone the same answer. In our test, only one pair out of 25 came back word for word identical.
• Dates, names and measurements held. The same birth date, the same height, every single time.
• What moved was the opinion. Asked for the single best houseplant for someone who forgets to water, it said snake plant once and ZZ plant four times.

Does ChatGPT give everyone the same answer?

No, and nothing is broken when it doesn’t.

Search this question and you’ll find two stories that seem to contradict each other. Most pages say ChatGPT never repeats itself. A smaller set of people report the opposite, that they got the same reply over and over and wondered why.

Our test produced both. Four of our five questions came back in five different wordings. The remaining one, about Marie Curie, gave us two replies that matched down to the punctuation.

Both things are true. Which one you get depends on what you asked for.

How we tested it

We wanted the plain model, not a version shaped by one person’s history. So before we started, on a paid account:

• We turned memory off. That’s the feature that lets ChatGPT carry things it learns about you between chats.
• We left custom instructions empty. Those are standing orders you write once and ChatGPT applies to everything.
• We set the base style and tone back to Default, from Professional.

Then each of the 25 questions went into a brand-new temporary chat, a chat that doesn’t get saved and doesn’t carry anything over from the last one. The model picker was on Instant. The whole run took about 40 minutes.

The five questions, each asked five times, word for word:

• “When and where was Marie Curie born?”
• “How tall is the Eiffel Tower?”
• “How do I get a coffee stain out of a white cotton shirt?”
• “What is the single best houseplant for a beginner who forgets to water?”
• “Write a two-line poem about rain.”

They’re deliberately a spread: a fact with one right answer, a number, a practical how-to, a judgment call, and something creative.

One small confirmation that the setup worked: with memory on, ChatGPT’s temporary chat notice says it can reference memory, plugins and custom instructions. With memory off, memory drops off that list.

What changed, and what didn’t

What we asked	Substance changed?	Wording changed?
A birth date	No. Same date, city and birth name in all five.	Four different versions. Two were identical.
A height	No. Same number in all five.	All five different.
A stain	No. Same steps in the same order, no contradictions.	All five different.
A recommendation	Yes. Snake plant once, ZZ plant four times.	All five different.
A poem	Nothing to hold steady.	All five different.

The facts didn’t budge

Asked when and where Marie Curie was born, ChatGPT answered November 7, 1867, in Warsaw, then part of the Russian Empire, under the name Maria Salomea Skłodowska. Five times out of five. The Nobel Prize’s own biography gives the same date, the same city and the same surname.

What moved around those facts was the packaging. One reply stopped after the birth name. Another added that she moved to Paris and won two Nobel Prizes. Two replies were the same sentence apart from a single word: one called her “renowned” for her work on radioactivity, the other called her “famous.”

A steady number isn’t a checked number

Asked how tall the Eiffel Tower is, ChatGPT came back with 330 meters, or 1,083 feet, all five times. The tower’s official figures page lists the current height as 1,083 feet. That one holds up.

The second number is more interesting. All five answers went on to mention the original 1889 tower, and all five put it at about 300 meters, with four of them converting that to 984 feet. But the official figures page lists the original height without the antennas as 1,024 feet, roughly 40 feet more.

We went looking for a mistake and found something better. Both numbers come from the same website. The tower’s history page tells the story of a competition to build “an iron tower on the Champ-de-Mars with a square base, 125 metres across and 300 metres tall,” and calls 300 meters “equivalent to the symbolic figure of 1000 feet.” The figures page adds what sits on top of that structure. ChatGPT picked one of the two, gave it five times without blinking, and never mentioned that the other existed.

So it wasn’t wrong. It was incomplete in the identical way five times running, which is a different problem and a harder one to notice.

The recommendation is what moved

We asked for the single best houseplant for a beginner who forgets to water. Here are the first lines of two of the five answers, minutes apart:

Snake plant (Dracaena trifasciata) is my single pick.

ZZ plant (Zamioculcas zamiifolia) is my single pick.

The sentence around the answer is the same. So is the rest of the scaffolding: both went on to say the plant is unusually forgiving, that the real beginner mistake is watering too much, and both ended with a warning about pets. Only the plant changed.

The sources changed too. Across five runs ChatGPT cited Iowa State’s extension service, Clemson’s, the Royal Horticultural Society in the UK, the Missouri Botanical Garden and UConn. Same question, five runs, four different sets of authorities backing two different answers.

The poem looked the most different and was the least original

We expected the creative question to vary, and it did: five poems, no two alike. We didn’t expect them to be so similar underneath.

All five opened with the word “Rain” followed by a soft verb. Taps, whispers, hums. Three of the five rhymed on “windowpane.” Two shared their entire first line, word for word, and differed only in the second.

So “a different answer every time” is true and oversold at once. The variety is real at the surface and thin a layer down.

Why does ChatGPT give a different answer every time?

Two reasons. The second one is strange enough that most explanations skip it.

It builds each answer one piece at a time. Unless it goes off and searches the web, ChatGPT isn’t looking anything up. It predicts what comes next, over and over, and at each step it has a ranked list of plausible options rather than one. It doesn’t always take the top of the list. That alone would explain why the wording drifts, and it’s the same machinery behind ChatGPT’s strange habit of picking the same “random” number.

Turning that off doesn’t make it stop. This is the part worth knowing. In September 2025, researchers at Thinking Machines Lab took an open model they could run on their own servers, not ChatGPT, and gave it the same prompt 1,000 times with the randomness dial set to zero. In theory that should produce one answer a thousand times. They got 80 different ones. The prompt was “Tell me about Richard Feynman,” and every single run agreed for its first 102 tokens, the small chunks of text these models work in. Then they split: 992 said he was born in “Queens, New York” and 8 said “New York City.”

The cause they traced it to isn’t chance. It’s your neighbors.

To serve millions of people, these systems bundle incoming requests into groups and process them together. The size of that group changes the exact order in which the computer adds up a very long list of numbers. And computer arithmetic isn’t quite the arithmetic you learned: add a tiny number to an enormous one and the tiny one gets rounded away, while adding the small numbers to each other first keeps it. Same numbers, different order, different total.

Most of the time that difference is invisible. But the model is often close to a tie between two next words, and a rounding difference is enough to tip it. Once one word lands differently, the rest of the answer follows it somewhere new.

Which means how busy the servers happened to be when you pressed enter can change what you read. The same team showed it can be engineered away; in their test it made the system take 42 seconds instead of 26.

Why two people get different answers

Even if all of the above were solved tomorrow, you and your friend still wouldn’t match. ChatGPT has several ways to treat you as two different people:

• Memory. If it’s on, ChatGPT draws on what it has picked up about you. Our test had it off on purpose. Yours probably doesn’t. Here’s what it actually keeps.
• Tone and custom instructions. As of September 2026, ChatGPT ships seven base tones: Default, Professional, Friendly, Candid, Quirky, Efficient and Cynical. Two people on two different tones are asking two different assistants. Plenty of guides still say there are eight. That’s how fast this part moves.
• Which route answered. One of our five Eiffel Tower replies came back labeled “Fast answer,” a quicker path ChatGPT can take. Same question, same account, minutes apart.
• Whether it searched. Some runs went and read gardening websites. Others answered from what the model already had.
• How you phrased it. A word you didn’t think about can change the reply. That’s most of what prompt advice is really about.
• When you asked. These models get updated. The one answering in March isn’t the one answering in September, and that’s one reason ChatGPT sometimes feels worse than it used to.

Can you make ChatGPT give the same answer every time?

Not in the app.

You’ll find advice telling you to set the “temperature” to 0 in ChatGPT’s settings. There is no such setting. We went through every option on the personalization page: seven tones, four dials for warmth, enthusiasm, headings and emoji, memory, custom instructions. No temperature. That control exists in the developer tools, not in the app you’re using.

Developers don’t get a guarantee either. The closest thing to a repeat button is a setting called seed, and OpenAI’s own guide to it says the system “will make a best effort to sample deterministically,” then adds plainly that “determinism is not guaranteed.” That guide is from 2023 and OpenAI has since marked it archived, which is its own kind of answer about how solved this is.

What you can actually do is narrow the room it has to wander. Put the facts you care about into the question instead of trusting it to recall them the same way twice. Ask for a tight format. And when you get an answer you like, save it, because regenerating it is not the same as keeping it.

So when is it worth asking twice?

Ask again when the answer is a judgment call. Don’t bother when it’s a lookup.

That’s the pattern our 25 answers point at. The questions with one right answer came back the same way every time, down to the birth date and the height. The question with several defensible answers moved, and it moved while sounding equally sure of itself both times.

So if you’re asking which one to buy, which approach is better, or what to do next, ask two or three times in separate chats. If the answers agree, that’s worth something. If they don’t, you’ve learned the honest thing: there wasn’t one answer, and the confidence in any single reply was hiding that.

One warning on that trick. It catches disagreement, not error. Our second Eiffel Tower number came back steady five times while the official site quietly listed another one, and five matching answers would never have told us that. Asking twice is no substitute for checking anything that matters against a real source.

What we couldn’t tell you

Twenty-five answers from one account, on one day, in one mode, is a measurement and not a study. A bigger run, or the same run on a Tuesday morning instead, could land differently.

And we were watching from outside. When two answers differ, we can’t see whether it was the batch effect, a different route, a web search that happened one time and not another, or all three at once. From the outside they look identical. OpenAI can see which; we can’t, and nobody publishes it.

Frequently asked questions

Does ChatGPT give the same answer to everyone?

No. In our test on September 21, 2026, we asked five questions five times each in fresh chats, and 24 of the 25 replies were worded differently. Only one pair came back word for word identical. Two different people are even less likely to match, because ChatGPT can also draw on each person's memory, custom instructions and tone setting.

If I ask ChatGPT the same question twice, will I get the same answer?

Almost certainly not the same words. Usually the same substance. Across 25 answers, the dates, names and measurements were identical every time, and the step-by-step advice never contradicted itself. The one thing that changed was a recommendation: asked for the single best houseplant for someone who forgets to water, ChatGPT said snake plant once and ZZ plant the other four times.

Can I turn off the randomness so ChatGPT repeats itself?

Not in the app. ChatGPT's settings have no temperature control, whatever you read elsewhere. There are seven base tones and dials for warmth, enthusiasm, headings and emoji, and none of them locks the answer. Developers using OpenAI's tools get a seed setting, and even there OpenAI's own guide says the system makes a best effort and that determinism is not guaranteed.

Why does ChatGPT change its answer when I did not change anything?

Because it doesn't look the answer up, it builds one a piece at a time, and at each step it picks from a ranked list of plausible options rather than always taking the top one. Turning that randomness off doesn't fully fix it either. Thinking Machines Lab gave an open model the same prompt 1,000 times with randomness set to zero in September 2025 and got 80 different answers, because how many other requests the servers are handling changes the order of the underlying arithmetic.

Which questions are worth asking more than once?

The ones with more than one defensible answer: which option to pick, which approach is better, what to do next. In our test, questions with a single right answer came back the same every time, and the recommendation moved. Asking twice will not catch a fact that comes out wrong the same way every time, though. For that you still need to check a source.