For one prompt, I needed ChatGPT to pick a random number from 0 to 9. The content of the prompt itself isn’t important—the point is that the next step depended on the number.
I don’t really trust generative AI with math tasks, but I decided to give it a try anyway:
Me: Generate a random number from 0 to 9
ChatGPT: Random number: 7. Want another one?
Me: yes
ChatGPT: Random number: 3.
Me: again
ChatGPT: Random number: 9.
Well, there’s some variety! Looks like it works. In my case, I didn’t need any special accuracy—so it seemed fine.
But…
I tried asking the model without any prior context (in a new chat), and got a different result—GPT-4o gave the same number, 7, several times in a row. Only once out of ten did it give a 6.
I googled it and found a tweet saying that ChatGPT really does have a bias toward “seven.” This has been noticed before, both with single-digit and more complex numbers.
So you shouldn’t rely on this “randomness”—at least not if you need real entropy. For critical tasks, it’s better to use other tools.
P.S.
I also thought of a simpler way—just take the last digit of the current time’s seconds. That would have been enough for my task. But here’s another disappointment—it turns out ChatGPT doesn’t have access to real time 😳. Even though the model gives an answer to “what time is it?”, it’s not tied to actual time at all. As the chat explained to me, “I just have a feeling…” It’s strange that the developers still haven’t “hooked up” some kind of internal function like this to ChatGPT.
