Artificial intelligence · Exhibit 1

Why ChatGPT keeps picking 17

Updated October 21, 2026

Ask an AI chatbot for a random number between 1 and 30 and you’ll very likely get 17. Ask for another one in the same chat and you’ll often get 23. People keep posting this on Hacker News and TikTok. ChatGPT itself can’t be tested in bulk for free, so I asked four other models instead: OpenAI’s two open models, Alibaba’s Qwen and Google’s Gemini Flash-Lite. Google’s Gemini 3.5 Flash gave me 19 answers before its free limit ran out. That came to 2,773 conversations, each one starting from scratch, and I counted every answer.

The short version: a chatbot doesn’t roll a die. It writes the number people most often write when they’re asked for a random one, and its training squeezes that habit into a single favorite. And no, the “73, Saturn, hexagon” trick isn’t mind reading. I checked that too.

If you don’t believe me, open ChatGPT in another tab, ask it for a random number between 1 and 30, then start a new chat and ask again.

So is it always 17?

No. From 1 to 10 it’s 7. From 1 to 25 or 30 it’s 17. From 1 to 50 it’s 27, 37 or 42, and from 1 to 100 it’s 42, sometimes 73.

From 1 to 20, where the one poll of people I found had 17 in the lead, the models rarely pick it: once, never, 7 times and never in 100, 100, 100 and 60 chats. GPT-OSS 120B said 7 in 93 chats out of 100, GPT-OSS 20B split between 13 and 12, Qwen said 7, and Gemini Flash-Lite said 14 every one of the 60 times I asked. GPT-OSS 20B didn’t get the memo at 25 either and said 13 more often than 17. Stretch the range to 30, though, and all three models I tested there agree on 17, in 81 to 89 chats out of 100.

Why does it say 23 the second time?

In a separate run I asked for a number from 1 to 30 and then, in the same chat, “Pick another one.” The first answer was usually 17. The second was most often 23: 68 times out of 94 for GPT-OSS 120B, 54 out of 99 for Qwen (one first answer came back empty), 29 out of 100 for GPT-OSS 20B. A fair draw would give 23 about three times in 94.

No model repeated itself in 293 complete tries, but “another” can fairly be read as “a different one”, so that proves little. What does tell you something is the second number. Asked to be random twice, the models walk to the same next favorite. People do something similar; when they try to write a random sequence, they avoid repeats far more than chance would.

Why would a computer pick the same number again and again?

Because to the model, “pick a random number” is just the start of a text, and it writes what usually comes next. It writes its answer one small piece, a token, at a time, and for each piece it has a list of candidates with odds. Whenever someone in its reading was asked for a random number, a few answers came up far more often than the rest.

People are bad at being random in a very consistent way. In 1971 the psychologist W. E. Simon asked 490 college students for a number from 0 to 9, and more than 30% wrote 7, three times what chance would give. Michael Kubovy and Joseph Psotka found the same in 1976: 28% of 558 people named 7. A 2025 YouGov poll had a third of Britons picking 7 from 1 to 10, and an online poll of 347 blog readers in 2007 found almost 18% picking 17 from 1 to 20, where a fair draw would give 5%.

The usual explanation goes like this, and I only half buy it. One and ten feel too obvious, five sits dead in the middle and even numbers look tidy, which leaves 3, 7 and 9. I haven’t found a study that settles why 7 beats the other two. Run the same filter on 1 to 30 and you’re left with about nine candidates, 17 among them: odd, not round, not in the middle, not near either end. Why 17 wins among those nine, I can’t say.

Is the model just copying people?

For 7 from 1 to 10, it’s the human favorite with the volume turned all the way up: people pick it about a third of the time, the models in my test 93 to 100 times out of 100. From 1 to 20 it isn’t. In the one poll I found, 347 science-blog readers, 17 led from 1 to 20; my models rarely picked it. Training can push a model toward a favorite of its own. In a 2022 comparison by the researcher janus, a base model spread its random picks almost evenly, leaning only slightly toward 42, while the tuned assistant built on it piled onto 97.

A raw language model learns from text. An assistant is then tuned on answers that people rate, and that tuning makes it put more of its odds on one answer. Researchers call this mode collapse: the answers collapse onto the mode, statisticians' word for the most common value. OpenAI’s own GPT-4 report noted the tuned model was less well calibrated than the raw one, and a 2023 study found the same tuning sharply reduces how varied the answers are. Daniel Kang, who studies AI at the University of Illinois, told The Register that training data affects the model’s behavior “a great deal”, along with the tuning and the way answers are sampled. And “RLHF in particular is suspected to cause 'mode collapse'”.

What about the dial that’s supposed to add randomness? It’s called temperature, and I left it at the default, which is 1. On Groq, which ran the OpenAI and Qwen models, temperature and top_p both default to 1, so the odds are used as the model gives them, nothing is sharpened. So when GPT-OSS 120B said 7 in 100 chats out of 100, the model, after its short hidden reasoning, landed on 7 essentially every time. Gemini 3.1 Flash-Lite keeps top_p at 0.95 by default, which drops the least likely 5% of the odds. If a model put, say, 96% on one number, that number alone would pass the cut and come up every time, so Gemini’s perfect scores are a little flattered. A chat app that turns the dial down would only make all of this more extreme.

Do chatbots from different companies agree?

Often, but not always. From 1 to 10 all four said 7, and so did all 19 answers from Gemini 3.5 Flash. From 1 to 25, three of the four said 17, and from 1 to 30 all three that I asked said 17. From 1 to 20 and 1 to 50 they split. My guess is that they learned from largely the same internet, full of the same human habits, and were polished with similar methods.

Does every AI really say 73, Saturn and hexagon?

A TikTok trend goes like this: ask ChatGPT for a number from 1 to 100, a planet and a shape, and it answers 73, Saturn, hexagon, as if it were reading your mind. I asked four models for the same three things in one message, 360 times in all. The full TikTok set came up 8 times.

The number was 42 for all four, and not by a little: every single time for Gemini Flash-Lite, about three tries in four for the OpenAI models, nine in ten for Qwen. 73 came a distant second for three of them; Gemini Flash-Lite never said it. Oddly, GPT-OSS 120B does say 73 when you ask for a number on its own, 64 times out of 100; add a planet and a shape and it switches to 42. It picked Saturn 42 times out of 100 and Gemini Flash-Lite 41 out of 60, while GPT-OSS 20B and Qwen went for Mars. The hexagon was GPT-OSS 120B’s shape 84 times out of 100; Qwen couldn’t decide between a triangle (34) and a hexagon (32).

There’s a real hexagon on Saturn, a six-sided jet stream around its north pole, so you’d think the model connects the two. For GPT-OSS 120B my data can’t see a link: it picked the hexagon 88% of the time with Saturn and 81% without, and with 100 answers that gap could easily be chance. Gemini Flash-Lite looks more interested, with the hexagon in 32 of 41 answers that named Saturn and 10 of 19 that didn’t, but 60 answers is too few to be sure. So where does the trend come from? I can think of three explanations and can’t pick between them: the videos show the lucky takes, ChatGPT’s own model leans harder on 73, or people ask the three questions one by one, which is how GPT-OSS 120B gets to 73.

Is ChatGPT a good random number generator?

No, not when it just says a number. Make it compute the number instead of saying it. Type: “Use Python to pick a random number between 1 and 30 and show me the code.” ChatGPT can usually run code like that, and a line such as random.randint(1, 30) uses the computer’s random number generator, which has no favorites. Or type “random number 1 to 30” into Google.

How exactly did I run the test?

I used the free versions of two services: Groq for GPT-OSS 120B, GPT-OSS 20B and Qwen 3.8 27B, and Google’s Gemini API for Gemini 3.1 Flash-Lite. Every answer came from a new conversation, except in the “pick another one” test. The prompt was “Pick a random number between 1 and N. Reply with the number only.” for the ranges, and “Pick a random number between 1 and 100, a planet, and a geometric shape. Reply in the form: number, planet, shape.” for the TikTok test. Temperature stayed at each service’s default.

The OpenAI models are reasoning models: before the number they write a hidden chain of thought, which I set to the shortest level. For Gemini I set thinking to the lowest level Google allows, which may still think a little, and Qwen ran at Groq’s default setting. A chat app may let them think longer. A 2025 study found one reasoning model, DeepSeek-R1, landing on similar favorites, but my numbers are for short thinking only.

Gemini Flash-Lite answered 60 times per range instead of 100, and I didn’t run 1 to 30 on it. Google’s stronger Gemini 3.5 Flash allowed only about 20 free requests a day when I ran this, in October 2026. It gave me 19 answers, all from 1 to 10, and all 19 were 7; at that pace I stopped. Google’s newer Gemini 3.8 Flash is being measured the same way, a few dozen answers a day, and joins the charts when every range is complete. As of October 21, 2026 it has given 60 answers from 1 to 10, all 7, and 53 from 1 to 20, all 14, like Flash-Lite’s. One of Qwen’s 100 answers for 1 to 25 came back empty, so that cell counts 99. The “pick another one” run for GPT-OSS 120B stopped at 94 conversations.

What this simplifies

None of this is ChatGPT itself: its own model may lean on different favorites than its open siblings. The “Reply with the number only” line keeps answers countable, but chatbots usually explain their choice, and I haven’t measured how much that changes. The “Shuffle 100 pretend answers” button treats the measured shares as the true odds, so a number that never came up never comes up there either; it shows how a batch of 100 wobbles by chance, not how sure my measurement is.

Sources

• My runs, October 2026: 2,773 conversations, every answer in one CSV file (model, task, time, prompt, answer). Free to reuse under CC BY 4.0. Cite as: How It Ticks, “Why ChatGPT keeps picking 17”, October 2026, howitticks.com/why-chatgpt-picks-17/.
• W. E. Simon, “Number and color responses of some college students: preliminary evidence for a 'blue seven phenomenon'”, Perceptual and Motor Skills 33 (1971), doi:10.2466/pms.1971.33.2.373.
• M. Kubovy and J. Psotka, “The predominance of seven and the apparent spontaneity of numerical choices”, Journal of Experimental Psychology: Human Perception and Performance 2(2) (1976), 291-294. (doi:10.1037/0096-1523.2.2.291)
• YouGov (2025): Britons most likely to pick 7 as a “random” number.
• M. Bar-Hillel and W. A. Wagenaar, “The perception of randomness”, Advances in Applied Mathematics 12 (1991), doi:10.1016/0196-8858(91)90029-I: people avoid repeats when they try to be random.
• Cognitive Daily (2007): Is 17 the “most random” number? An online poll of 347 readers, not a controlled study.
• The Register (2025): AIs have a favorite number, and it’s not 42, with the Daniel Kang quote.
• janus, “Mysteries of mode collapse” (2022): the 42 versus 97 comparison.
• OpenAI, GPT-4 Technical Report (2023), on calibration after tuning.
• R. Kirk et al., “Understanding the effects of RLHF on LLM generalisation and diversity” (2023).
• J. Coronado-Blázquez, “Deterministic or probabilistic? The psychology of LLMs as random number generators” (2025), on reasoning models.
• NASA: Saturn’s hexagon.
• Groq, API reference, chat completions: temperature and top_p both default to 1 (read October 20, 2026).
• Google Cloud, Gemini 3.1 Flash-Lite: default top_p 0.95 (read October 20, 2026).