I’m an artist and creative professional who uses GPT extensively.
I genuinely appreciate what GPT has given me. I use it for research, creative work, and sometimes even as a sounding board when I need to think through something.
Recently, though, I encountered something very small that made me think about a much larger problem.
There was a question circulating on social media asking people to generate a few random things:
-
A random number between 1 and 30
-
Another random number between 1 and 30
-
A planet from the Solar System
-
A geometric shape
A surprisingly large number of GPT users reported receiving the same answers:
17, 23, Saturn, triangle.
So I tried exactly the same questions myself.
I got:
17, 23, Saturn, triangle.
At first, this is just funny.
I looked into it afterward and found that there are already discussions about numerical biases in language models, including the tendency of certain models to repeatedly select particular numbers when asked for something “random.”
And obviously, there is nothing inherently special about 17.
But the more I thought about it, the more interesting the distinction became.
A language model does not necessarily produce randomness in the same way that a mathematical random-number generator does. Instead, it can generate an answer from a learned probability distribution of what a “random” answer is likely to look like.
In other words, the model may not be reproducing randomness itself.
It may be reproducing what humans have historically made randomness look like.
And the person receiving the answer may then interpret that output as genuinely random.
That is a tiny example.
But what happens if the same mechanism applies to things much more important than numbers?
If “randomness” can become something like “what has historically appeared random in human-generated data,” what happens when AI begins to mediate concepts such as:
creativity, normality, relevance, plausibility, importance, taste, originality, or even truth?
My concern isn’t simply that AI might give us wrong answers.
It is that AI might gradually make the range of possibilities we encounter narrower.
Why this feels personal to me
I think I belong to a somewhat unusual generation.
I grew up without smartphones. I remember communicating with friends through text messages and landlines.
I started using smartphones during adolescence.
And I only began using AI as an adult.
Sometimes I wonder what it would have been like to be born into a world where all of these technologies were already ubiquitous.
But if I could choose again, I think I would still choose to be born at the end of the 20th century.
Not because I think the past was better.
But because I had time to make mistakes.
To become interested in things that seemed useless.
To pursue irrational interests.
To misunderstand things.
To change my mind.
To meet people I never expected to meet.
To fail.
To become obsessed with something for no particular reason.
To take inefficient paths and sometimes discover something precisely because the path was inefficient.
None of those experiences were deliberately designed to make me more creative or intelligent.
They were simply part of growing up.
And I don’t think we can fully recreate that through education.
If we tell children:
“You need to make mistakes.”
“You need to experience boredom.”
“You need to think independently.”
“You need time without AI.”
then even those experiences become intentional and structured.
And that creates a strange paradox.
The moment we try to teach the conditions that once naturally produced unpredictable human development, we may already be narrowing the range of experiences ourselves.
Perhaps these aren’t skills that need to be taught.
Perhaps they are conditions that need to remain possible.
What worries me about AI and childhood
Children may need opportunities to:
-
be wrong without immediately receiving the correct answer;
-
remain curious without immediately resolving the question;
-
become interested in something without knowing why;
-
experience boredom without an algorithm filling the space;
-
make something badly and experience failure firsthand;
-
encounter people and situations that cannot be optimized;
-
form interpretations that are initially strange or incorrect;
-
experience consequences that cannot simply be delegated to an AI.
These things are valuable partly because we don’t know where they will lead.
That unpredictability may be the point.
And then there is synthetic data
This is where my concern about the “17 problem” connects to something much larger.
As AI-generated content becomes an increasing part of our information environment, some of that content will inevitably become part of future datasets, references, and training material.
Could this create a feedback loop?
AI produces statistically plausible outputs.
Those outputs become part of the information environment.
Future AI systems learn from increasingly AI-mediated information.
The statistically common and plausible becomes more represented.
Rare, unconventional, ambiguous, poorly documented, or simply strange human perspectives become less visible.
Not because they are false.
But because they are less represented.
This seems different from traditional “AI bias.”
It may be closer to a gradual narrowing of the space of possibility.
The question I’m trying to get at
What happens if AI becomes the primary interface through which children learn, search, create, interpret, and communicate?
Will it only change what they know?
Or could it eventually change what they spontaneously imagine as possible?
I don’t think the answer is necessarily that AI should be banned.
I use AI myself. I benefit enormously from it.
I think AI can expand human capabilities in extraordinary ways.
But perhaps we need to distinguish between:
AI helping humans think
and
AI thinking before humans have had the chance to think.
And perhaps we should ask whether there are forms of human experience that should remain deliberately unoptimized precisely because they are inefficient, unpredictable, and impossible to reproduce on demand.
Maybe AI literacy shouldn’t only mean:
“How do I use AI effectively?”
or even:
“How do I recognize when AI is wrong?”
Perhaps it should also include:
“When should I not ask AI?”
I keep coming back to a simple question:
What kinds of thinking should humans continue to do for themselves?
And an even more uncomfortable one:
What experiences must remain genuinely open-ended so that future humans can still become something we cannot predict?
I started thinking about all of this because an AI gave me the number 17 when I asked for something random.
Maybe it’s nothing.
But maybe the interesting part isn’t why it chose 17.
Maybe it’s what happens when billions of people begin receiving similarly plausible answers, over and over again.
I’d genuinely like to hear what other people think.
Have you noticed similar patterns in ChatGPT or other AI systems—particularly cases where the answer wasn’t obviously wrong, but somehow felt too predictable, too representative, or strangely convergent?