What AI really is
is a fill-in-the-blank champion. Play a word game to see how AI that chats with us builds its answers, and why a smooth answer can still be wrong.
1. Can you fill in the blank?
A cat says “____”.
What sound fits? Say it out loud, then open the answer.
See a word that fits
“Meow” fits! You know the sound a cat makes. uses patterns learned from many examples to work out what words could fit.
2. Add a little more, then a little more
-
- A cat
-
- A cat
- says
-
- A cat
- says
- “meow”.
-
Look at the words so far
-
Pick the next piece
-
Add it to the end
Repeat until the answer is done
Read more
AI uses your question and the words so far to choose the next small piece of text. Then it does this again. The answer grows a little at a time.
These steps show the idea. Real AI adds tiny pieces of text that can be shorter than a word. It can build a whole story this way!
3. Friendly words come from a computer
| A person | ||
|---|---|---|
| Saying “Nice to meet you!” | Yes | Yes |
| Feeling happy or sad | Yes | No |
| How the answer is made | Thinking and experience | Calculating with numbers |
Read more
AI might reply, “Nice to meet you!” But friendly words do not show that it has feelings like a person. AI is a computer tool that makes replies by doing calculations.
4. A smooth sentence can still be wrong
A cat says “meow”.
- Sounds natural Yes
- Is true Yes
A cat says “woof”.
- Sounds natural Yes
- Is true No
Wrong answers have many causes
- It picked a word that fits, but it isn't true
- It learned from wrong examples
- It is missing information
- The has a problem
A wrong answer is not always your fault.
Read more
Imagine an answer like this: “A cat says woof.”
That sounds like a proper sentence, but woof is a dog’s sound! Making words fit and getting facts right are different things.
AI does not always check the facts before it answers. So even a clear question can get a wrong answer.
5. Can AI see inside your wardrobe?
“Which of my coats should I wear today?”
It can’t see your closet
“It’s cold today. I have a thin coat and a thick coat. Which would keep me warmer?”
It can compare the two coats
- Passwords
- Home address
- Photos of you or other people
Read more
If your clothes and the weather have not been shared with AI, it cannot know them on its own. Try giving just the clues it needs. You do not need to send a photo or your address.
Remember these three things
-
AI builds answers by adding small pieces of text.
-
Check important answers in a book, by calculating, or with a teacher.
-
Share useful clues. Keep your password and home address private.
Your turn!
AI wrote a long, confident answer. Does that mean it must be right? Think of a reason before opening the answer.
See the explanation
No! AI can join words smoothly and still get a fact wrong. For sums, calculate it yourself. For facts, check a book or ask a teacher.
Want to know a little more?
This page explains a common kind of . “Fill-in-the-blank champion” is a comparison, not a description of all AI. During , the adjusts numbers using examples. When making an answer, it uses those numbers and the text so far to calculate which small text pieces, called , could come next. It does not always pick the most likely one, so answers can vary. A likely answer is not necessarily true.
How do reasoning models answer?
-
Read the question
-
Write thinking notes 1st round of adding pieces — like scrap paper
-
Write the answer 2nd round — reading the notes too
Some AI shows “Thinking…” before it answers. This kind of AI is called a reasoning model. A reasoning model does not use a secret new trick. It answers in exactly the way you saw on this page: look at the words so far, pick the next piece, add it to the end, and repeat until it is finished.
There is just one difference: it uses this way one more time before the answer. First it writes thinking notes, piece by piece. It lists the facts, tries a calculation, and starts again if it finds a mistake. Then it treats the question and its own notes together as “the words so far” and writes the answer piece by piece. It is like solving a problem on scrap paper first, then looking at it while you write the answer on the answer sheet.
Because the working is already in the words so far, it often does better on problems with many steps, like maths or . Reasoning models have had extra practice at writing useful notes. But the notes come from the same calculation that picks the next piece, so a model can think for a long time and still be wrong. Thinking also takes more time and more computing. Always check important answers yourself.
One piece at a time outside, deep calculation inside
“Picking the next piece” describes how the answer comes out. It does not mean the model thinks only a little inside. Every time it picks one piece, it passes all the words so far through dozens of calculation steps called layers, one after another. Along the way it works out, in numbers, who is who, what is being asked, and what should come next.
It can also plan ahead. When researchers at looked inside a language model writing a rhyming poem, they found it chose the rhyming word for the end of the line first, then wrote the start of the line to reach that word. Pieces come out one at a time, but the calculation inside looks further ahead — like climbing stairs one step at a time while looking at the top.
Not every AI writes by adding pieces
- Picture-making AI (): it starts with a whole picture full of noise and cleans up the entire picture a little at a time, over many rounds.
- Diffusion language models: text can be made this way too. Models such as Gemini Diffusion and Mercury lay out a whole answer full of blanks at once, then fill in and fix many pieces together.
- Sorting and recommending AI: it decides whether a photo shows a cat, or picks videos you might like. Instead of writing a long reply, it gives a score or chooses from set answers.
Still, most AI that chats with you in writing — like Lyen, and — answers the way this page shows: by adding pieces one at a time.
Sources and further reading