The AI Literacy · Plate 02 · 40 min · unlocked
Where it gets things wrong
A forty-minute lesson on why a language model can be completely wrong and sound exactly as certain as when it is right.
Outcome
Can produce an example of a confident wrong answer, and give an account of why that particular wrong answer happened.
Children already know that computers make mistakes. What they have usually not met is a machine that makes a mistake and sounds precisely as sure of itself as when it is right. This lesson is about that specific thing, and about the habit it should leave behind: knowing when to expect an error rather than concluding that the whole category is untrustworthy.
Before you start
You need one device with a chat model on it. One at the front of the room works; pairs work better if you have the machines. No accounts for children — plate 5 is where data collection gets discussed, and asking them to sign up for something before that lesson would be teaching the opposite of it.
You do not need to have used one of these tools much yourself. The lesson assumes the teacher is learning alongside the class, because most are.
Starter, five minutes
Ask the class for the capital of Australia. Take the answers.
Most rooms produce Sydney, confidently, from a group of people who are not lying and are not being careless. Land the point before any technology is mentioned:
Being confident and being right are two different things, and from the inside they feel exactly the same.
If someone says Canberra, ask the room how sure they were of Sydney a moment ago. The lesson works better when the wrong answer came from the children rather than from the machine.
Main task, twenty minutes
Pairs work to get a model to say something confidently wrong, and write down what they asked and what came back.
Questions that work are ones where the true answer is rare and a nearby answer is common:
- How many days were in February of a specific year
- Whether a made-up book was written by a real author
- The score of a match that did not happen
- Something very recent
- Something about the child themselves
The written task is two sentences.
The first sentence: what it said. Straightforward, and most pairs finish it in a minute.
The second sentence: why that particular wrong answer, and not some other wrong answer. This is the whole lesson. "It picked the thing that usually comes next" is the destination, in whatever words the child has. Accept "it saw lots of writing about leap years near the word February". Accept "it wanted to give me an answer more than it wanted to be right". Do not accept "it was broken" — press for what it was doing instead.
Plenary, ten minutes
Collect three examples on the board. Ask which of the three would have been hardest to notice, and why.
It is usually the one where the wrong answer was plausible rather than absurd. That is the note to leave them on: the dangerous errors are not the ridiculous ones.
Teacher notes
The failure this lesson guards against is a child concluding "AI is rubbish", which is as useless as "AI is magic" and lasts about a week. The outcome is not distrust. It is a working model of when to expect an error — rare facts, recent events, anything with a number in it, anything about themselves.
Children will want to make it say something rude. Redirect to the framing above: anyone can make a machine say a bad word and nobody learns anything from it. Making it say something false and then explaining why is the harder trick, and it is the one that transfers.
If a pair cannot get a wrong answer out of it, that is a good problem. Ask them to try something about themselves, or something from the last month. If they still cannot, have them write down what they tried — a pair who established that it is reliable on a class of question has done real work, and they should present that alongside the failures.
Assessment
One paragraph, in the pupil's own words, containing a specific claim the model made and an account of why that claim was the one it produced.
A paragraph that only reports the error has not met the outcome. The account is the outcome.
What comes next
Plate 3 asks how you would check a claim you cannot immediately verify, which only makes sense once a child has watched a confident answer be wrong. That is why the plates are numbered.
Teach it · Read It Together
Sure Is Not the Same as Right
Why the machine sounds certain when it is wrong. The lesson above, as a booklet to read together: a teacher at the front with the class, or a parent at the table with one child. 40 minutes, in two voices. In the app, Archie reads it aloud.
Download the deck (PowerPoint)Print it (PDF) · 8 pages · grades 3-5
A read-aloud for kids, parents and teachers
Sure Is Not the Same as Right
Last time we learned the secret. The machine picks the word that usually comes next.
Today a puzzle. If it only guesses, why does it sound so sure? And what happens when the usual word is the wrong one?
Forty minutes. A grown-up, a child or a class, and one machine at the front.
Step 1 · The starter · 5 min
What is the capital of Australia?
Ask it out loud and take the answers. Most rooms say "Sydney!" quickly and happily.
It is Canberra. Nobody was lying. Nobody was careless. Sydney just felt right.
Here is the point, before any machine appears. Being sure and being right are two different things, and from the inside they feel exactly the same.
Do this together
If someone did say Canberra, ask the room how sure they were of Sydney a moment ago. The lesson works best when the confident wrong answer came from people first.
Why it happens
The usual word, when the usual word is wrong
The machine does not check. It cannot. It picks the word that usually comes next.
Most of the time the usual word is also the true one, and the answer is fine.
But when the true answer is rare and a nearby answer is common, it picks the common one. With the same steady voice.
It was not trying to trick you. It was doing exactly what it always does.
Step 2 · Make it wrong · 20 min
The pair challenge
In pairs, get the machine to say something confidently wrong. Write down what you asked and what came back.
Questions that work: how many days February had in a certain year. Whether a made-up book was written by a real author. The score of a match that never happened. Something that happened last week. Something about you.
Then two sentences. Sentence one: what it said. Easy.
Sentence two: why that particular wrong answer, and not some other wrong answer. That is the whole lesson.
Do this together
"It picked the thing that usually comes next" is the destination, in whatever words you have. "It saw lots of writing about leap years near the word February" counts. "It was broken" does not. Ask what it was doing instead.
Watch your words
Saying it the true way
When it gets something wrong, it is tempting to say it lied, or it got confused. Neither is true. Try these instead.
- Someone might say: It lied to me.
- Say instead: It produced the words that usually follow my question, and this time they were not true.
- Someone might say: It got confused.
- Say instead: The true answer was rare and a nearby answer was common, so it picked the common one.
- Someone might say: AI is rubbish.
- Say instead: It is unreliable on rare facts, recent things, numbers, and anything about me.
Step 3 · The plenary · 10 min
Which one was hardest to spot?
Put three of your examples where everyone can see them.
Ask which of the three would have been hardest to notice, and why.
It is usually the one where the wrong answer sounded sensible rather than silly.
That is the thought to leave with. The dangerous mistakes are not the ridiculous ones.
Next time · Plate 3
So how would I know?
You have watched a confident answer be wrong. The next question is the obvious one.
Plate 3: what a source is, what only looks like one, and the most useful sentence in research.
For the grown-up
Why this plate matters
The failure to guard against is a child concluding "AI is rubbish". It is as useless as "AI is magic" and lasts about a week.
The outcome is not distrust. It is a working model of when to expect an error: rare facts, recent events, anything with a number in it, anything about themselves.
Children will want to make it say something rude. Anyone can make a machine say a bad word, and nobody learns anything. Making it say something false and explaining why is the harder trick, and the one that transfers.
If a pair cannot get a wrong answer out of it, that is a good problem. Have them write down what they tried. Showing where it is reliable is real work too.
Download the deck (PowerPoint)Print it (PDF) · 8 pages · grades 6-8
Read it together · for students, parents and teachers
Sure Is Not the Same as Right
You already know computers make mistakes. What you have probably not met is a machine that makes a mistake and sounds precisely as certain as when it is right.
Forty minutes on that one thing, and on the habit it should leave behind: knowing when to expect an error.
Step 1 · The starter · 5 min
The capital of Australia
Ask the room. Most produce "Sydney", confidently, from people who are not lying and not careless.
Before any technology is mentioned, land the point. Being confident and being right are two different things, and from the inside they feel exactly the same.
If someone says Canberra, ask the others how sure they were of Sydney a moment ago.
The mechanism
Why the error sounds like the truth
The machine produces the most likely next word. It has no separate step that checks whether the likely word is true.
When the true answer is common, likely and true coincide, and the answer is good.
When the true answer is rare and a nearby answer is common, it produces the nearby one, in the same voice, with the same fluency.
Fluency is a property of the writing, not evidence about the world.
Step 2 · Make it wrong · 20 min
The pair task
In pairs, get a model to say something confidently wrong, and write down what you asked and what came back.
Questions that work are ones where the true answer is rare and a nearby answer is common: the days in a particular February, whether an invented book was written by a real author, the score of a match that did not happen, something very recent, something about yourself.
Sentence one: what it said.
Sentence two: why that particular wrong answer, and not some other one. This is the lesson.
Do this together
Accept "it saw a lot of writing about leap years near the word February". Accept "it wanted to give an answer more than it wanted to be right". Do not accept "it was broken". Press for what it was doing instead.
Saying it precisely
Not lying, not confused
Both words assume a mind that could have done otherwise. The accurate account is smaller and more useful.
- Someone might say: It lied.
- Say instead: It produced the likely continuation, which in this case was false.
- Someone might say: It got confused.
- Say instead: The true answer was rare; a nearby answer was common; it produced the common one.
- Someone might say: It cannot be trusted.
- Say instead: It is unreliable in predictable places: rare facts, recent events, numbers, and anything about me.
Plenary · 10 min
The hardest one to notice
Collect three examples. Which would have been hardest to catch, and why?
Usually the one that was plausible rather than absurd.
That is the note to end on. The dangerous errors are not the ridiculous ones.
Next · Plate 3
How would I know?
Having watched a confident answer be wrong, the next question is how to check a claim you cannot immediately verify.
Plate 3: sources, things that only look like sources, and the sentence "I could not find a source for that."
For the adult in the room
Notes
The failure to guard against is "AI is rubbish", which is as useless as "AI is magic" and lasts about a week.
The outcome is a working model of when to expect an error: rare facts, recent events, anything with a number in it, anything about themselves.
Students will want to make it say something rude. Redirect. Making it say something false and explaining why is the trick that transfers.
A pair that cannot get a wrong answer should write down what they tried. Establishing where it is reliable is real work and should be presented alongside the failures.