← All insights

A · Manage · 2026-10-18

Use AI, but Don’t Let It Pick 7 for You

Illustration · not a service record: A packed lecture hall — an AI talk in progress, speaker small against the venueIllustration · not a service record

Ask AI to pick a number from 1 to 10, and it picks 7 almost every time. A chart showing this has been going viral online lately, and it can actually be understood with one math term from the seventh grade of junior high school.

This week there was a chart on X that drew several million views from a single post alone. For the same prompt, “pick a number from 1 to 10,” true randomness is a flat line, human answers go up and down, and AI shoots up in a sharp spike at 7. The original poster said just one sentence: “If you understand gradient descent, this is the least surprising chart in the world.”

I want to explain it with junior high math, paired with an Opus 5.5 animation.

The value that appears most often in a set of data is called the “mode.” People already favor 7: in psychology experiments, about 30% of people choose it, because 1 and 10 feel too close to the edges and 7 feels most like “a random pick.” AI has read vast amounts of human text and learned that 7 is the most common; research that tested several AI models found that they chose 7 about 80% of the time.

But there is a detail here that deserves more attention than gradient descent. If AI were merely imitating humans, the share choosing 7 should be around 30%. The push from 30% to 80% happens mainly in post-training: when humans score AI’s responses, they unconsciously prefer answers that are more typical and more familiar, which researchers call “typicality bias.” So AI picks 7 not because it especially likes 7, but because it is too good at guessing what everyone wants. In other words, part of AI’s “typicalization” is something we taught it with our own hands.

(Incidentally, that viral chart is an illustrative diagram; the curves cite no source, but the direction is right.)

This does not happen only with numbers. Ask AI to draw a watch and the hands often stop at 10:10, because nearly all the watch ads online set them that way. If a whole class writes essays with AI, the work grows more and more alike: a study in Science Advances found that stories written by people who were given AI ideas were rated as more creative, yet were more similar to one another. The individual gets better; the collective gets narrower.

Yet what changes the world is often that short bar in the bar chart. In 1968, everyone crossed the bar facing forward; only Fosbury jumped with his back to it. He was described as looking like a fish flopping about on a boat, yet he took the gold medal at the Mexico City Olympics. Four years later, 28 of 40 competitors had switched to his technique. Today, almost every high jumper jumps this way.

So this is what I would say to children: AI is very good at giving you the mode, but your answer does not have to be the mode.

There are two small methods. One is to think it through yourself first, and then ask AI. The other is to ask it to “give me 5 answers, and mark how common each answer is,” and then start revising from the least common one. This is also why Junyi Academy’s AI Notebook is designed not to give the answer “first”: let children think for themselves first, so that not everyone hands in the same 7.

Using AI is fine. Just don’t let it pick 7 for you.

What was the first number you thought of? Tell me in the comments. If you want to get the “mode” straight together with your child, Junyi has a worked-example video that teaches exactly “how to read the mode from a bar chart.” The link is in the first comment.