Word coins: teaching children what an AI answer costs
Every word an AI reads and writes is paid for by somebody. Turning that into coins in a jar is the fastest way we have found to teach a child that a longer answer is not automatically a better one.
We gave a group of children a jar of coins and one rule: every word an AI wrote back cost one coin, paid out of the jar in front of them. They asked for a five-hundred-word answer to a simple question first, purely to see what would happen, and watched the jar empty fast. Then they asked the same question for fifty words. The second answer answered the question exactly as well. Nobody needed us to explain what had just happened.
That is the entire idea behind tokens, minus the technical name. Every word an AI reads and every word it writes is paid for by somebody, and turning that into physical coins is the fastest way we've found to make the cost visible to a child.
What is a token in AI?
A token is roughly a small chunk of a word - sometimes a whole word, sometimes a piece of one. Every token an AI reads in your question, and every token it writes back in its answer, has a small real cost attached to it. More words in, more words out, more cost - which is why a rambling question and an over-long answer are not free just because nobody hands over actual money at the till.
A wandering three-paragraph question gets a nine-paragraph answer. Costs the most coins. Reader finishes about a third of it.
A tight, four-sentence question gets a four-sentence answer. Costs a fraction of the coins. Reader finishes all of it.
Why length is not the same as quality
The four-sentence answer above is not a worse answer because it's shorter - in the jar exercise, children consistently rated it just as good, and sometimes better, because they actually read the whole thing. Length has a price. Length is not automatically quality, and a model's default instinct, left unguided, tends toward longer rather than shorter.
One honest limit: cost is not the only thing that matters, and a child should not learn to always demand the shortest possible answer regardless of the question - some things genuinely need more words to explain well. The lesson is noticing the cost exists at all, not minimising it blindly. The next article in this level goes further into a related choice: not just how long an answer is, but which model answers it, and why the biggest one is not always the right pick.
Questions we get asked
What is a token in AI?
A token is roughly a small piece of a word - sometimes a full word, sometimes part of one. Every token in the text an AI reads and every token it writes back carries a small real cost, which is why longer questions and longer answers cost more, even though no money changes hands directly in a typical chat.
Does a longer AI answer cost more than a shorter one?
Yes - cost scales with the number of tokens involved, both in the question and the answer, so a nine-paragraph response costs meaningfully more than a four-sentence one answering the exact same question. Length adds cost regardless of whether the extra length adds any real value.
How do you teach a child that AI answers have a cost?
A simple physical exercise works well: assign one coin per word, ask for a long answer and count out the coins, then ask the identical question for a short answer and count again. Comparing the two piles, and then judging which answer was actually better, makes the abstract idea of cost immediate and concrete.