The answer that calls for help.
Watch an AI answer a maths question two ways. First from memory — where it can sound completely sure and still get the number wrong. Then by handing the sum to a calculator and reading back the real answer. The interesting part is how it decides which to do.
Narrated by a synthetic voice — not a recording · 7 short stops
The catch: answering from memory, the AI never actually does the maths — it writes a number that only looks right. So how do you ever get a number you can trust? Watch what changes when it can hand the sum to a calculator instead.
Scroll to watch the hand-off ↓Scroll — then tap the buttons as you go ↓I wanted to watch the moment of hand-off, not just the polished answer. So this little demo lays the two paths side by side: answer from memory, or send a calculator a request and read back what it returns.
First, the answer from memory ↓Would you catch the wrong digits?
Answering from memory, the AI can produce a believable-looking number without ever doing the multiplication. A smooth sentence gives you the shape of an answer — but nothing in it tells you whether the sum was actually worked out.
Read the answer on the left. Try to tell, from the sentence alone, whether it is right.You can't tell. See what the other path does instead ↓What if the output isn't an answer at all?
This time the AI's output isn't a number. It's a short request: the name of a helper — a calculator — and the sum for it to work out. Because it's just a written request, a separate program can read it, check it, allow or refuse it, and only then actually run it.
Here's the part people get wrong. The AI never runs the calculator itself. It only writes the request and hands it over; the program around it — the app you're actually using, the thing that runs the AI and can pass its request to a real calculator — decides whether to run it, and hands the answer back for the AI to use.
The request is ready. Run it ↓Which number survives?
Run the request and watch. It moves from Ready, to Running, to Complete — and the result stays on the page so you can compare the two answers.
Try it — press “run the request” below and watch which digits are real.The from-memory guess didn't survive. So how did it know to ask? ↓How did it know to ask for help?
The AI never stops to check how sure it feels. From the huge pile of text it was built on, it simply learned a habit: a question that looks like maths is best handed to a calculator.
Where did that habit come from? In the huge pile of writing it was built from, it saw the same move over and over — a hard sum, then someone reaching for a calculator, then the checked number. Copy that pattern enough times and “reach for the calculator on a sum like this” becomes a reflex — no understanding required, just a well-worn groove.
So a chatty or general question, it just answers from memory. A question that looks like a sum, it reaches for the calculator. It isn't measuring some inner feeling of certainty — only reacting to the shape of what you asked.
Try it — clicktap chit-chat or general knowledge, and watch the path switch while the calculator stays quiet.A learned habit, not a hunch. Now steer it yourself ↓Can you steer it yourself?
Type a sum, and it should reach for the calculator. Type something chatty, and it should answer from memory.
Try it — type any sum and run it; watch which path lights up before the answer arrives.That request carried your numbers to the calculator. One more thing to take with you ↓
The calculator's answer comes back as new information. The AI drops it into its sentence — without pretending it worked out the digits itself.
The AI doesn't run the calculator itself. It writes the request and hands it over; the program around it runs the calculator and passes the answer back.
The evidence stays readable even without scripts.
{ tool: "calculator", expression: "847 * 293" }