How it was built11 of 11
Not all AI is the same
There is no single thing called "the AI". Different companies built different systems, gave them different rules, and they behave differently on purpose.
How to explain it, by age
Same idea every time. What changes is how much of the machinery a kid can hold, so pick the version that fits and go up a band when it stops being interesting.
Ages 4-6
Use the idea of different makers. "There isn't just one robot. Different companies made different ones, and they act differently, like how different shops have different rules."
Try this
Ask two different assistants the same silly question and let her hear the answers differ. Different is the whole point at this age.
Listen for
Her saying they are not the same. That is enough.
If she already has more access than most. Name the two she encounters most, so they stop being one undifferentiated thing.
Ages 7-9
Add the choosing. "Each one was made by a different company, and each company decided what it should be like. How polite it is, what it refuses to talk about, whether it makes jokes. Those are decisions somebody made, not things that happened by themselves."
Try this
Ask two of them the same question and compare. Then ask her which answer she preferred and why, which is a better question than which was right.
Listen for
Her attributing a difference to a decision somebody made.
If she already has more access than most. Ask her to guess what kind of person a company was imagining when they made it sound the way it does.
Ages 10-12
Introduce the layer on top. "Underneath, they are all prediction machines trained on huge amounts of text. On top of that, each company adds rules about what it will and will not do. Those rules reflect what that company worries about: being sued, being embarrassed, being seen as biased. So when one refuses and another does not, you are seeing a company's caution, not a fact about the world."
Try this
Find a question one assistant declines and another answers. Ask her what she thinks each company was worried about.
Listen for
Her treating a refusal as information about the company rather than about the subject.
If she already has more access than most. Introduce the idea that the same underlying model can behave very differently depending on the instructions wrapped around it.
Ages 13+
Move to values and who set them. Every one of these systems encodes choices about what counts as harmful, balanced or appropriate, made by small teams inside companies, largely in one part of the world. Those choices then reach hundreds of millions of people. Open models complicate this further: anyone can remove the rules, which is both the strongest argument for open systems and the strongest against.
Try this
Have him probe the same politically contested question across several systems and document the differences. The pattern of what each hedges on is more revealing than any single answer.
Listen for
Him asking who decided, and noticing that "neutral" is itself a position somebody chose.
If she already has more access than most. Discuss whether a genuinely neutral assistant is possible, and what it would even mean.
Why the plural matters
A child who says “the AI said” has already conceded the argument. It sounds like a single authority, the way people once said “the encyclopaedia said”.
A child who says “ChatGPT said, and Claude said something different” is doing something else entirely. They are treating each one as a product made by a company with its own reasons, which is exactly what it is.
That habit costs nothing to build and it changes how every future answer gets received.
The personality was designed
Worth saying plainly, because it is not obvious: the warmth is a product decision. Someone chose how friendly it should be, tested versions, and picked one.
That is not sinister. Shops train staff to be pleasant for the same reasons. But a child who knows the friendliness was designed is much less likely to mistake it for a relationship, and that matters more every year as these systems get better at sounding like they care.
Misconceptions to address
- That all these tools are the same underneath. They differ in training data, in the rules layered on top, and in what the company wanted them to be like.
- That the personality is a personality. The friendliness was designed, tested and chosen, in the same way a shop's greeting is.
- That refusing to answer means the topic is forbidden. It usually means one company drew a line there, and another company drew it elsewhere.
- Adults describe "what AI thinks" as though there were one opinion. There is not.
Running this with a class
Put the same question to two or three different assistants on the board and compare answers side by side. Ask the class to account for the differences. They will propose training, rules and company choices unprompted, which is the whole concept arrived at inductively.
Last reviewed August 2026