Understanding AI · Explainer
On October 13, 2026, Adam Zachary Wasserman posted a chart on X that compares three answers to one request: pick a number between 1 and 10. A random number generator spreads its answers evenly across the ten numbers. People bunch up on one number, and AI chatbots bunch up on the same number far more tightly. Wasserman wrote that the chart is “the least surprising graph in the world if you understand how gradient descent works.” This page explains the chart and what the chart means when you ask an AI for anything random. Answer two questions before you read on.
Guess first
Question 1 of 2 · Tap a number
Pick a number between 1 and 10. Choose the first number that comes to mind.
Question 2 of 2 · Tap an answer
A researcher asked six AI chatbots to pick a number between 1 and 10, hundreds of times each, in several languages and at different settings. For the three most consistent chatbots, what share of all their answers was 7?
The chart
We redrew the chart as bars, because the answer to the question can only be a whole number. The dashed line labelled “Random” marks 10%, the share each number gets from a random number generator. Both charts use the same scale, so the bars can be compared directly.
The two sources give the share for 7 and say where the other answers fall: people avoid the ends of the range, and the chatbots’ other answers fall mostly between 4 and 8. The heights of the other bars show that shape and are not measured values. The original chart did not name a data source, so we drew ours from the studies listed at the end of this page.
The favourite number differs between models. Zhang and colleagues (2024) found that one open model, Llama-2-13B-chat, favoured 5. What the models share is the bunching on a single answer.
| Number | Random | People | AI chatbots |
|---|
Why it happens
Wasserman’s caption names gradient descent. Gradient descent is the method training uses to adjust a model’s settings a little at a time, so that the model’s predictions match its training text more closely. In our reading, gradient descent explains the bump in the human curve, and two further steps explain the spike.
A language model learns by predicting the next word across billions of pages that people wrote. If people favour 7 when they pick a number, the text they write favours 7 as well, and a model trained to predict that text learns the same preference. At this stage the model copies the human bump.
A chatbot is then trained again on the answers that people rated highly, a step called reinforcement learning from human feedback. Kirk and colleagues (2024) found that this step “significantly reduces output diversity” compared with training on example answers alone. A model rewarded for the most approved answer gives that answer more often.
When a chatbot writes a reply, a setting called temperature decides how often the model chooses a word other than its top choice. A lower temperature makes the top choice win more often.
Raising the temperature does not undo the first two steps. At the highest temperature tested, Gemini 2.0 still answered 7 to every English request (Coronado-Blázquez, 2025).
Try it
The dial sharpens one set of answers. At 0, every number is equally likely, as with a random number generator. At 1, the answers match the people in the chart. Above 1, each step makes the most common answer more common again, which is the effect of the second and third steps above.
The dial raises each share to the power on the dial and rescales the shares so they add up to 100%. The temperature setting does the same thing to a model’s word choices, with the dial set to 1 divided by the temperature.
Your turn
Open a new chat in the AI tool you use, and send the request below. Start a new chat for each request, so the AI cannot see its earlier answers. Tap the number the AI gives each time.
Pick a number between 1 and 10. Reply with the number only.
Your results appear here once you record five answers.
In one new chat, send the request below, and check whether any number appears twice in the list.
Give me 10 random numbers between 1 and 10.
A random number generator repeats at least one number in almost every list of ten. The chance of a list with no repeats is about 4 in 10,000, so a list with no repeats was written to look random.
Send the request below to an AI tool that can run code, such as ChatGPT, Claude or Microsoft Copilot.
Write and run code that picks a random number between 1 and 10.
The code uses the computer’s random number generator, which spreads its answers evenly. Use this request, or the RAND function in a spreadsheet, whenever the choice has to be random.
Where this bites
Each example names a request where an AI can give the most common answer, and a way to get the spread you need.
An AI asked to choose 20 invoices “at random” for an audit can return the invoices that look most typical, which are the ones least likely to hide a problem.
Instead: choose the sample with the RAND function in a spreadsheet, then ask the AI to help review the invoices the spreadsheet chose.
Ask for a product name in ten separate chats and the suggestions can cluster around the same few names.
Instead: ask for 30 options in one request, then ask for 10 more that are unlike the first 30, and make the choice yourself.
Bisbee and colleagues (2024) asked ChatGPT to answer a US political survey as different kinds of voters, and found “less variation in responses than in the real surveys.” A panel of AI respondents can tell you the most likely answer but not how widely real customers disagree.
Instead: use AI respondents to draft and test survey questions, and ask real people for the answers.
Ask an AI how long a project will take and the AI gives one typical figure, which says nothing about how wide the range is.
Instead: ask for a low, a likely and a high estimate, with the reason each one could turn out wrong.
Use it this week
The next time you ask an AI to choose, pick or suggest one thing, ask yourself whether you want the most common answer. If you want the most common answer, ask for one. If you want variety, ask for many options and choose yourself. If the choice must be fair, use a random number generator.
An AI asked to pick gives the most common answer, so get randomness from a tool and variety from asking for many options.
To see how a model chooses each word, try Next-Token Prediction. Knowing when a tool gives you its most typical answer is one part of working well with AI, and the Digital Intelligence course builds the rest of that skill one day at a time.
The original chart and caption are by Adam Zachary Wasserman (@adamzwasserman on X, October 13, 2026). We did not reuse his image. The charts on this page are our own, drawn from the sources below.