Ai is predictably unreliable

Marc

Registered User
Messages
1,998
A simple test of the limitations of AI

Open up an AI tool and ask for a random number between 1 and 25 and the chances are you will get 17.

Now think about the implications of how much trust you are putting into the answers it is giving you when it is this predictable.
 
They are LLMs and don’t have a built in random number generator.

I presumed if you ask a bunch of people for a random number 17 will pop up often too.

Us humans aren’t very good at selecting random numbers. We avoid the extremes, even numbers and seem to be drawn to prime numbers.

Neither are computers very good at real randomness. A dice roll would be better.

Try random.org if you need a random number.
 
Last edited:
Chatgpt also gave me 17. I asked why. See response below. I suppose you would need to prompt it to actually use a random number generator.
 

Attachments

  • Screenshot_20260724_093259_ChatGPT.webp
    Screenshot_20260724_093259_ChatGPT.webp
    102.4 KB · Views: 26
Now think about the implications of how much trust you are putting into the answers it is giving you when it is this predictable.
I was watching a talk on YouTube the other day with some academic talking about Google and it's adoption of Ai in it's search engine

And what I took from it was years ago when you search on Google for an answer to a question, you got links to pages/sites that might give you information to help answer that question

Now with the integration of Ai, Google gives you the answer, which is all well and good

But the flip side of it, is because people are accepting that that is the answer people are not doing their own research and traffic to these pages/sites is falling at an alarming rate

Now that to me is more worrying then the number 17
 
Simply asking the question will limit them to predicting words in sequence from within the vast knowledge they have ingested. Instructing them to use a random number generator will get you a better result, for example Gemini will get you one by running the Python random command in the background. That will be a pseudo random generated number, but good enough for most purposes.

To get a truly random number, you'll need to use the paid versions which can be configured to call external APIs, but you're getting into code there.
 
I presumed if you ask a bunch of people for a random number 17 will pop up often too.
If you ask people for a number between 1 and 10, 7 will be the one you get the most.

OPs post is funny because if anything it's showing that LLMs may contain the same biases as humans, who are also bad at choosing random numbers. It's also partly a skill issue. The better your prompt is the better your results will be. Try this prompt instead for example:

Randomly choose a number between 1 and 25. Make it really random, don't just choose the one people choose most often.
or
Randomly choose a number between 1 and 25 with equal likelihood of each number
 
Last edited:
Back
Top