AI4RA Tip of the Week #25
Why AI Is Bad at Picking Random Numbers (and What That Teaches Us)
Here’s a quick experiment. Open a new chat with an AI tool and ask it to “pick a random number between 1 and 10.” There’s a very good chance it says 7. Or ask it to “pick a number between 1 and 100?” Did it choose 37? New chats always start fresh, but ask again in the same chat, and you will often see the same handful of numbers come up again and again, like 37, 42, and 73. Ask it for a number between 1 and 100 dozens of times and you won’t get an even spread.
It’s a fun party trick, but why does this happen? It shows us something important about how AI works, and why you shouldn’t always take its answers at face value.
Why Does This Happen?
AI learned from people, and people are bad at being random. When someone is asked for a “random” number, they don’t really pick at random. They skip numbers that feel too obvious, like 1, 10, 50, or 100, and gravitate toward ones that feel random, like 7 or 37. AI models learned from millions of examples of human writing, so they picked up the same habit.
Training makes the habit stronger. When AI companies refine their models to give polished, confident answers, the models settle even more firmly on their “favorite” responses.
Each new chat starts from scratch. A fresh chat has no memory of what it said in other conversations, so it can’t notice that it has answered “7” a hundred times today. Within a single chat, it may avoid repeating itself, but it simply moves to the next most “human” answer, which is often 3.
Put simply, the AI isn’t rolling dice. It’s predicting what a person would most likely say. And people are pretty predictable.
It’s Not Just Numbers
The same pattern shows up in other places. For example, when AI writes multiple-choice quiz questions, it tends to put the correct answer in the same few positions, often B, instead of spreading answers evenly. Human test-writers have the same tendency, and the AI learned it from them. If you use AI to build training quizzes, check the distribution of correct answers and shuffle them yourself.
Why This Matters for Research Administrators
Randomness comes up more often in research administration than you might think. Some examples:
- Choosing transactions for an internal spot-check or audit sample
- Selecting files for a quality review
- Building training quizzes for new staff
If you ask an AI to “randomly select 10 awards to review,” the result may look random to you but follow hidden patterns. That could undermine the fairness and defensibility of your process.
The bigger lesson goes beyond numbers. AI produces the most likely answer, not necessarily the correct one. Its output can look confident and reasonable while quietly reflecting human habits and biases. Knowing that helps you decide when to trust AI and when to verify.
How to Use This Resource
- Use a real random tool when randomness matters. Use a random number generator, Excel’s RANDBETWEEN() or RAND() functions, or your audit software’s sampling feature.
- Ask the AI to run code instead of guessing. Many AI tools can run code. Try this prompt: “Use code to randomly select 10 rows from the attached spreadsheet. Show me the code you ran so I can verify the selection.”
- Check AI-generated quizzes. Count how often each answer letter is correct, and shuffle the options if they’re uneven.
- Document your method. For audit samples or reviews, record how your selection was made so anyone can reproduce and verify it later.
How can I learn more about this topic?
AI4RA Office Hours: every fourth Tuesday at 11 AM PST, with live Q&A and demos (https://ai4ra.uidaho.edu/#events)
Share your questions or success stories
Try the experiment yourself! Ask your favorite AI tool for a random number between 1 and 10 and share what you get on the AI4RA community forum (https://groups.ai4ra.uidaho.edu/g/main/topics). Your tips help the whole research‑admin community thrive.
Interested in authoring a blog post?
AI4RA is seeking authors for our blog! If you’re interested in writing a post about AI or data in research administration, fill out this brief (under 5 minutes) survey and we’ll reach out to you: https://uidaho.clickup.com/forms/9017952524/f/8cr628c-48817/PMFN4R8WWFVQBPVIOL
Have you deployed an AI4RA tool? Let us know!
Knowing who has deployed our tools tells us whether they’re being used by the community, shows us where to focus development, and supports our NSF reporting requirements. If you’ve deployed an AI4RA tool, please take this brief (under 1 minute) survey: https://uidaho.clickup.com/forms/9017952524/f/8cr628c-101017/LQRDE77XMJA1FR671E