Before you read on, pick a number between one and ten. Don’t think about it. Just pick one and hold on to it.
This week a graph popped up in my social media feed. Someone had asked AI chatbots to pick a random number between one and ten, over and over again. If the answers were truly random, the ten bars would have been roughly level. They weren’t. One bar towered over the rest, and it was seven.
It’s no fluke either. Kieran Browne of Springboards, an Australian AI company, put the same question to three of the best known chatbots in 100 fresh conversations each. ChatGPT said seven 92 times. Claude said it 90 times. Google’s Gemini said it every single time. A larger study of six models found the same pull. In the chart below, each bar shows how many times out of 100 a chatbot said seven. The dotted line shows what a truly random pick would give: seven about 10 times in 100. All three bars blow straight past it.
So why seven?
Partly because of us. Seven has always been a lucky number. I can remember buying lollipops as a child, and if its wrapper revealed the number seven, you won a free one. It isn’t just nostalgia. In 1976, two psychologists stopped 558 people on the Yale campus and asked for the first number between 0 and 9 that came to mind. If people picked at random, each number would get about one vote in ten. Seven got 28 votes in every 100, close to three in ten.
Why do we do it? When we’re asked for something random, we steer clear of anything that looks too neat. One and ten sit at the edges. Five sits in the middle. Even numbers feel tidy. Seven is odd, off centre and nobody’s idea of a round number, so it feels like the random choice, which is exactly why it isn’t. People wrote the text these chatbots learned from, so our habit came along for the ride.
The rest is down to how the machine works. A chatbot doesn’t roll a dice. It writes by predicting what is most likely to come next, so ask it for a random number and you get the most likely one. Put the two studies side by side. People in 1976 chose seven about three times in ten. The chatbots chose it more than nine times in ten. The chatbot didn’t just copy our habit. It magnified it.
That makes it the best ten second AI lesson I know. Mention “large language models” at a staff meeting and you can watch the room drift. So don’t. Ask the staff to pick a number, then type the same question into a chatbot on the projector. When seven appears, the penny drops. It works just as well with a class.
And that penny matters, because the same tool that picks seven also answers “summarise the causes of this war”, “draft a comment for this report” and “explain both sides of this debate”. It answers them in exactly the same way. It reaches for the most popular answer, the one that turns up most often in what it learned from, and delivers it with the same confidence.
That is the real problem when anyone uses AI for research. With a number, we can ask a hundred times and count. With a paragraph, we see one answer. We never see the answers it might have given, or the views that were never likely enough to reach the top. Good research means weighing the full range of views. AI hands you the most popular one. A polished paragraph tells you nothing about whose voice is missing. Knowing that should be part of everyone’s AI literacy, not just our students’.
So, what number did you pick? If it was seven, don’t worry. The chatbot only picked it because it was the likeliest answer. You probably had a better reason. I know I did. I was born on the 7th. I never had a chance.
Patrick Hickey (@aiteachingguru on all major social media) is an AI CPD Provider, AI in Education Consultant and Media Contributor on AI in Education. He is an ASTI nominee to the Assessment working group of the Department of Education and Youth’s AI in Schools External Advisory Taskforce.
If you have any queries you can email www.aiteachingguru@gmail.com