Sometimes a lesson plan starts with a curriculum objective.
Sometimes it starts because you ask an AI to pick a number.
I was chatting with ChatGPT and, for reasons that aren't particularly important, asked it to choose a number between 1 and 30.
It chose 17.
This wasn't especially interesting until Teddy pointed out that AI models have a bit of a habit of choosing 17.
So, naturally, we tried it again.
And then we tried other AIs.
Seventeen kept appearing.
At which point a throwaway conversation became much more interesting.
Is 17 actually random?
Large language models are very good at producing something that looks like an appropriate answer.
That isn't necessarily the same as doing the thing we've asked them to do.
Ask an ordinary random-number generator for an integer between 1 and 30 and it uses a mechanism designed to generate numbers across that range.
Ask a language model to "pick a number between 1 and 30" and, unless it deliberately calls another tool to generate one, it is still doing what language models do: predicting a plausible response.
And apparently 17 looks wonderfully random.
It's odd. It's prime. It's not too close to either end. It isn't a neat multiple of five or ten. It's just beyond the midpoint.
It has, essentially, excellent random-number vibes.
Except that if everyone picks it, it isn't doing a particularly convincing job of being random.
This isn't unique to my experiment. Other people have found the same tendency, with 17 and other numbers ending in seven appearing disproportionately often when language models are asked to make apparently random choices.
And that raises a much more interesting question:
If an AI learns from human-generated information, how much of our own bias does it learn along the way?
This sounds like a maths lesson
At this point my teacher brain kicked in.
My pupils are growing up in a world in which AI will be completely ordinary. They'll use it at work, encounter it in services, search with it, create with it and probably interact with AI systems in ways we haven't thought of yet.
Teaching them simply how to use AI isn't enough.
They need to learn to question it.
So instead of telling them that AI has biases, we're going to investigate one.
The experiment is simple.
First, the pupils privately choose their own number between 1 and 30.
Then we ask a selection of AI models to choose a number from the same range, repeatedly.
Twenty trials gives us something to work with.
For each AI, pupils will record the results in a tally and frequency table. Then we'll turn the results into graphs and compare them.
Does 17 dominate?
Do other numbers appear unusually frequently?
Are some numbers hardly chosen at all?
Do different AI models behave differently?
And perhaps most interestingly:
Are the humans actually any better at being random?
Suddenly our silly number 17 has given us tallying, frequency, data collection, graphing, comparison, probability and interpretation.
But the maths isn't really the most important part.
"The computer said so" isn't evidence
What I want the pupils to discover is that an AI can give a perfectly reasonable-looking answer without necessarily using the process they assumed it was using.
17 is a harmless example.
Nobody particularly cares if ChatGPT's favourite pretend-random number is 17.
But the principle scales.
If we ask an AI to recommend something, classify something, summarise evidence or make a judgement, its answer may contain patterns inherited from the enormous quantity of human-generated material from which models learn.
Humans aren't neutral.
Our writing isn't neutral.
Our choices aren't random.
So we shouldn't expect AI magically to become neutral simply because there's a computer involved.
That's a much more useful understanding of AI than either "AI knows everything" or "AI is bad".
It's a tool.
A remarkably useful one.
But understanding what sort of tool you're using matters.
And then the lesson escaped my classroom
Once we'd planned the experiment, I realised there was no particular reason to keep it to my two pupils.
The task is accessible. The question is intriguing. Different classes could repeat exactly the same experiment and contribute their results.
So the plan has grown.
We'll create a shared space for results and invite other classes across the school to run the experiment too.
That gives us a much larger dataset.
It also gives pupils another question:
Do we still see the same pattern when we collect more evidence?
Which is rather important.
We started with an observation.
We formed a hypothesis.
Now we're going to collect data to see whether the evidence actually supports it.
If it doesn't, that's interesting too.
Accidental learning
This is one of my favourite ways for technology to enter education.
Not:
"Today we are learning about Artificial Intelligence."
But:
"That's odd. Why did it do that?"
Curiosity gives us the question.
Maths gives us a way to investigate it.
Digital literacy helps us understand what might be happening.
Critical thinking stops us accepting the first plausible explanation.
And AI itself becomes both the subject of the experiment and one of the tools we can use to explore it.
All because of one suspiciously popular little prime number.
17.