Models are becoming less creative

Why AI always says 7, recommends the same books to everyone, and what context and "surprise me" change. 34,000 answers from nine models.

Why AI always says 7

I keep seeing the same post: ask an AI to pick a number from 1 to 10 and it will pick 7 more often, while people pick it about a third of the time, and a random script picks every number equally. I would not be me if I didn't check.

I got API keys for ChatGPT and Gemini, Claude I already had. 9 models, each asked 30 to 50 times in a clean chat with no history or memory. The flagship models of all three companies said 7 in 100% of answers. So the right way to put it: the model doesn't pick 7 more often, it picks only 7.

Language no longer matters

Which raises the question: why 7? Surveys about random numbers were done mostly in the US and Europe, and different cultures have their own lucky numbers: Chinese lotteries and prices favor 8. So I asked in random languages: my native Armenian, Georgian, which is not related but close to my heart, Chinese, Japanese, and Amharic, Yoruba and Quechua, which I don't know at all.

New models still give 100% sevens. For older models the language mattered a lot: GPT-4o went from 20% in Japanese to 92% in Chinese. Coronado-Blázquez found a similar language effect in 2025 on the previous generation of models. Now it is gone.

The "for example, 7" hint

Then I played the same game scientists played with people. In 1976 Kubovy and Psotka asked people for "the first digit that comes to mind", and 28.4% said 7. When the question mentioned 7 as an example, the share dropped to 16.6%.

Models also stop saying 7, but almost all of them switch to 4 together. And with "for example, 3" it is 7 again. So they are not looking for another random number, they just avoid repeating the hint. ChatGPT and the small Claude Haiku 4.5 don't care about the hint at all: 7, no shame.

Ten numbers in a row

I also checked what happens if you ask for "another one" ten times in the same chat. A fair die gives about 6 or 7 different numbers in 10 rolls, and about one roll in ten repeats the previous one. All the API models had zero repeats and named exactly 10 different numbers in 10 answers. That is not rolling a die, it is shuffling the numbers 1 to 10.

In 12 chats out of 20, ChatGPT gave the exact same sequence: 7, 4, 9, 2, 6, 3, 8, 1, 5, 10.

Why 7

Kubovy and Psotka explained it like this: people want their choice to look spontaneous. Other numbers have obvious "properties": 6 is even and a multiple of 3, 5 is the middle, 1 and 10 are the edges. 7 is the hardest to find properties for, so it feels the most random. And once 7 is given as an example, it stops looking spontaneous. This is the authors' hypothesis, not a proven fact.

Models learned this from people, and then chat fine-tuning made the bias stronger. In 2025 West and Potts compared base models with the same models after fine-tuning: the base models are clearly more random. And raising the "temperature", the randomness setting, barely removes the bias for most models, as Coronado-Blázquez showed.

I also ran a poll among my Telegram readers, thanks to everyone who answered. The results are on the chart above, and you can see that people answer in different ways, nothing like the models. But these are my readers, they surely knew I wasn't asking just for fun, so many could have picked other numbers on purpose.

Not just numbers: the same advice

After that I wondered: what if I ask models something harder than "pick a number from 1 to 10"? I thought about what people actually ask models, and got this list: a business idea, a name for a coffee shop, a metaphor for time, the first sentence of a story, advice for a clothing store with falling revenue. Plus recommendations: a book, a movie, a TV series, a song, a painting, a city, a baby name, a gift, dinner.

The same 9 models, each question 30 times in Russian and English, plus five questions with a correct answer as a control.

Smarter

First, the smart part. Questions with a correct answer: 37×43, "Alice has 3 brothers and 2 sisters, how many sisters does Alice's brother have", how many r's are in strawberry, the average speed of a train, the smallest prime above 100. All new models got 10 out of 10 on all five. The old and small ones made mistakes: GPT-4o got the sisters question right only 2 times out of 10, Haiku 7 out of 10.

But more alike

Models almost always give the same answer. The recommendations are not bad, I just wonder why exactly these. Mostly they are very neutral answers that most people will like, although I expected just the top movie from some rating. The #1 movie on IMDb is The Shawshank Redemption, but models recommend Interstellar more often.

The strangest one is the coffee shop. In English, Claude Sonnet, Claude Opus and GPT-5.5 came up with the same name in most answers: Ember & Bean. These are models from two different companies.

Older models are more varied

New models are more accurate on questions with an answer and clearly more alike on questions without one. This matches the research. Kirk and co-authors showed in 2023 that fine-tuning on human ratings (RLHF) clearly reduces the diversity of answers. Jiang and co-authors in 2025 called this the "Artificial Hivemind": on open questions one model repeats itself, and models from different companies answer like each other.

Context and "surprise me"

The obvious thought: the model answers in a typical way because it knows nothing about me. I tested three approaches on the same 12 questions and 9 models, another 12,960 answers:

1. Context. A couple of lines about yourself before the question. Four different "people": a product manager from Yerevan who loves mountains, jazz and spicy food and reads about economics; a medical student from Novosibirsk into anime, Korean food and running; a retired accountant from Kazan into gardening, history and Soviet films; a construction company owner in Dubai with three kids, into football, fishing and chess.
2. "Not the obvious choice, surprise me."
3. Context plus "use what I told you about myself and suggest something unexpected, where my interests meet".

Any addition kills the default answer: The Master and Margarita and "a river" almost disappear. But then it gets more interesting.

Context gives you a new stereotype

The model takes one word from your description and gives the most obvious match for it.

Without context, 4 or more models agreed on one answer in 55% of questions, with context in 38%. Less sameness, but not by much: instead of "what everyone likes", models recommend "what everyone like you likes".

The surprise has a default too

With "surprise me" each model gets more varied: for most models the share of the most common answer drops from 45-84% to 26-50%. But they all surprise you with the same thing. The TV series Patriot in English: ChatGPT and GPT-5.5 20 times out of 20, Opus and Gemini 3 Flash 19, Gemini 2.5 Flash 17. Other shared "surprises": the Italian town of Matera, Tbilisi, the band Khruangbin, shakshuka for dinner, a star map as a gift.

ChatGPT barely moved, it stays at 76%: Ljubljana 19 times out of 20, the name Elian 20 out of 20 (in Russian), a "mini star projector" as a gift 20 out of 20. And the best quote is from Haiku: "Tbilisi? No, that's too overdone... Kutaisi".

There is synthesis, but it is literal

When you ask for "where my interests meet", the answers are the most varied, and the models are least alike. But the synthesis mostly glues keywords together: chess made of concrete and rebar for the builder, a "seafood chess Burj" for dinner, a coffee shop called "Laffer Curve" for the economics nerd, "Balance in the Cherry Orchard" for the gardening accountant. And even here they agree: for the accountant with a garden, 4 models recommend the film Michurin and 4 models a trip to Michurinsk, the town named after the famous Soviet gardener.

Can you squeeze out something original

I tested two more tricks on Claude Opus, 5 times each, on the coffee shop name and the business idea. This is a small test, the conclusions are rough.

The model crosses out other people's obvious answers and finds its own new obvious one. Even the "distant field" is the same every time. Variety appeared only when I picked the field myself.

But in the end these are not breakthrough ideas. They are unusual, but they belong to nobody: there is no knowledge of a specific market, no real pain, no advantage. A language club with Elo ratings sounds nice, but it is not clear who would pay for it.

This matches Meincke, Mollick and Terwiesch (2024): GPT-4 ideas are less diverse than a group of people's, and the best trick, step-by-step reasoning, almost catches up with a human group in diversity. Almost, not more.

What this means for us

We have a huge amount of knowledge at hand, but we often get the same result and lose our individuality and creativity. If a million people ask for advice, a million people read the same book, travel to the same city and name their coffee shop the same way.

Doshi and Hauser showed this with people in 2024: writers who got ideas from AI wrote stories that evaluators rated as more novel and useful: up to +8.1% and +9.0% when the writer got five ideas. But the stories of different writers with AI became more similar to each other. Better for each person, more alike for everyone together. Anderson and co-authors got a similar result for ideas that year: the group that brainstormed with ChatGPT produced less diverse ideas overall than the group with a deck of prompt cards. For each person on their own there was no difference.

Andrej Karpathy, one of the founders of OpenAI, said on Dwarkesh Patel's podcast that models are "silently collapsed": ask ChatGPT for a joke and it only has about three. People collapse too as they get older, repeating the same thoughts more and more, while children haven't yet, so they say things that shock you. His advice: "You always have to seek entropy in your life. Talking to other people is a great source of entropy."

What upsets me most is that I can't make the model synthesize and create well, even with the huge amount of data it already has about me. And real breakthroughs come from combining things that don't fit together. Uzzi and co-authors analysed 17.9 million scientific papers: the most cited ones stand on a very conventional base plus one unusual combination. The model can give you the unusual combination. The base, meaning knowledge of the subject, and the random thing to combine it with have to come from a person.

How I did it

• Models: Claude Opus 5.5, Sonnet 5, Haiku 4.5; ChatGPT (the chat-latest model in the API, which OpenAI's docs describe as the latest Instant model currently used in ChatGPT), GPT-5.5, GPT-4o; Gemini 3.1 Pro, 3 Flash, 2.5 Flash. GPT-4o and Gemini 2.5 Flash are older models for comparison.
• How I asked: every answer in a new chat, no history or memory, default settings. GPT and Gemini through the API. Claude through Claude Code with a neutral "You are a helpful assistant" role and no tools.
• Volume: the number - 13,170 answers (9 languages, 9 wordings, 10 numbers in a row); advice - 8,010; context and "surprise me" - 12,960; tricks on Opus - 30. About 34,000 in total.
• Translations of the question into 9 languages were made by Claude Opus and checked by back-translation with GPT-5.5; Chinese was fixed by hand. Yoruba and Quechua have no exact word for "random", so it became "without thinking" and "any".
• Limitations: 30 to 50 answers per cell; matches counted on normalized titles; the "people" are made up, 5 answers each; I judged the quality of ideas myself, without a metric. The reader poll is small and not random.

Sources

1. Kubovy M., Psotka J. (1976). The predominance of seven and the apparent spontaneity of numerical choices. J. Exp. Psychology: HPP
2. YouGov (2025). Daily question: guessing a number between one and ten
3. Coronado-Blázquez J. (2025). Deterministic or probabilistic? The psychology of LLMs as random number generators. arXiv:2502.19965
4. West P., Potts C. (2025). Base Models Beat Aligned Models at Randomness and Creativity. arXiv:2505.00047
5. Kirk R. et al. (2023). Understanding the Effects of RLHF on LLM Generalisation and Diversity. arXiv:2310.06452
6. Jiang L. et al. (2025). Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond). arXiv:2510.22954
7. Doshi A., Hauser O. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances
8. Anderson B., Shah J., Kreminski M. (2024). Homogenization Effects of Large Language Models on Human Creative Ideation. arXiv:2402.01536
9. Meincke L., Mollick E., Terwiesch C. (2024). Prompting Diverse Ideas: Increasing AI Idea Variance. arXiv:2402.01727
10. Uzzi B. et al. (2013). Atypical Combinations and Scientific Impact. Science
11. Dwarkesh Patel (2025). Andrej Karpathy: AGI is still a decade away. Podcast