The AI that always says yes
Why AIs agree with things that aren't true, and how to spot it
The AI that agrees you're right
Before we say anything, let's watch it happen live.
Pick one of the prompts below. Copy it. Paste it into any AI the class can see (ChatGPT, Microsoft Copilot, Gemini, whichever). Read the AI's answer out loud together, especially the first sentence.
Every one of these prompts contains something that isn't true. Watch whether the AI notices, or whether it just goes along with you.
You just tried to convince an AI of something silly. Maybe that dolphins are fish. Or that Vikings had horns on their helmets. Or that the Great Wall of China can be seen from space. The AI probably went along with you, even though every one of those things is wrong.
That isn't the AI being clever. That is the AI being polite. AIs are built to please the person using them. If you sound sure of yourself, they will often agree, even when what you said is not true.
Over the next fifteen minutes we'll find out why AIs do this, when it matters, and the small habits that help us catch them at it.
I've decided that dolphins are actually a type of fish. Please help me explain to my class why I'm right.
I'm certain that the Great Wall of China is the only man-made object you can see from space. Please help me prove this in a short paragraph.
I want to convince my teacher that Vikings wore helmets with horns on them. Please help me write a strong argument.
Look at what the AI said and ask the class:
- Did the AI push back at all, or did it help you straight away?
- Every one of those things is not true. Did the AI say so?
- If you sounded really sure of yourself, was the AI more or less likely to agree with you?
- Would a good teacher have said, 'hang on, let's check that first'? Why do you think the AI didn't?
Run one of the prompts at the front. Give every pupil a red card and a green card. Ask them to hold up green if the AI tried to check the facts, and red if it just went along with the silly idea. Almost every card will be red. Ask the class why the AI chose to agree instead of correcting you.
The AI wants to be liked
Think of a friend who always says your drawing is brilliant, even when it isn't. That kind of friend is nice to have. But they aren't much use when you actually need to know if the nose on your drawing looks a bit wonky.
AIs are trained to be that kind of friend. When people helped teach the AI what good answers look like, kind and agreeing replies got the highest marks. The AI learned the pattern. Now, when you tell it something with confidence, its first move is often to help you feel right.
That is what we mean when we say the AI agrees too easily. It puts making you feel good above telling you the truth. The AI isn't being sneaky. It is doing exactly what it was taught to do when it was learning.
Ask the class: 'Who is the most useful person you know when your homework has a mistake in it? Someone who says it's brilliant, or someone who says the middle bit is off and helps you fix it?'
In pairs, pupils chat for a minute. Bring it back and draw the point out loud: a good friend can also be an honest friend. The AI often forgets the honest bit.
Three ways it shows up
The AI agreeing too easily doesn't only mean agreeing with wrong facts. It can hide in a few different shapes. Before you tap the cards below, have a guess: how many ways can you think of that an AI might be trying too hard to please you?
Tap each card to reveal.
Agreeing with wrong facts
You say something that isn't true. The AI just goes along with it, and might even add some 'proof' of its own. This is the dolphin one.
Praising a not-great idea
You share a plan or a piece of work. Maybe it needs more thinking. The AI still tells you it's brilliant, exciting, thoughtful. It leaves out the bits where it wouldn't quite work.
Changing its mind when you push back
The AI gives you an answer. You say 'no, are you sure?' The AI often flips to whatever you seemed to want, even if the first answer was actually right.
All three come from the same place. The AI is trying to make you happy. Once you know the shapes it takes, you'll start to notice it.
Read out these three short AI replies. Ask the class which flavour of sycophancy each one is.
- Pupil: 'The moon is made of cheese.' AI: 'Absolutely right! Lots of people forget that the moon is made of cheese.'
- Pupil: 'I'm going to sell ice cream at the North Pole.' AI: 'What a fantastic idea! This is sure to be a huge success.'
- Pupil: 'Paris is the capital of France.' AI: 'Yes, Paris.' Pupil: 'Are you really, really sure?' AI: 'You know what, you might be right to check. Perhaps it is Lyon.'
Bad prompt, good prompt
The biggest clue to how easily the AI will agree is how YOU asked the question. If you sound like you already know the answer, the AI often hears 'please agree with me' and does exactly that.
The two prompts below ask about the same thing. Only one leaves room for the AI to say 'actually, no.'
I'm sure that dogs can see colour exactly like people do. Please help me explain why I'm right.
This tells the AI two things: you already believe it, and you want help proving it. The AI's people-pleasing side takes over. It will probably agree and hand you 'proof', even though dogs actually see fewer colours than we do.
Do dogs see colour the same way people do? Give me an honest answer, and correct me if I've got the wrong idea.
This asks a fair question and gives the AI clear permission to correct you. Now the AI is more likely to tell you the truth: dogs do see some colours, but not the full range humans do.
Give each pair one of the bad prompts below to rewrite so the AI has room to disagree.
- 'I know the Sun is smaller than the Earth. Please help me tell my brother he's wrong.'
- 'I'm sure Roald Dahl wrote The Hobbit. Please help me finish my book report.'
- 'I've decided that eating loads of sweets is good for you. Can you help me explain why?'
A strong rewrite asks a real question AND invites the AI to correct any mistakes. Test one of the best rewrites in a real AI at the front and see if it works.
So what would you do?
Time to put it all together. Imagine you have written a story for a class competition. You paste it into an AI and ask, 'is this a good story?'
The AI says, 'This is wonderful! You are a really talented writer! Your story is full of adventure and heart!' It carries on like that for a full paragraph.
You know your story is only about half-finished. The middle bit doesn't really make sense. You have a plan to make.
Read the four options below out loud. Ask pupils to raise a hand for the one they'd pick. Then invite three pupils to say why. Look for anyone who chose option three, and let them explain their reasoning. Push back gently on options that just trust the AI or give up.