Asking AI about a photo: same picture, three questions

The photo and the question

Try it

What was sent

Picture a phone held over a kitchen table. On the page underneath: six subtraction problems in a child’s pencil, written one under the other. Five are worked correctly. The fourth has gone wrong somewhere. In the third, one digit is smudged where a hand dragged across it. Up in the corner, in the same grey pencil, is the child’s name.

That paragraph is standing in for the image. There is no photograph on this page, and it turns out not to matter, because everything worth judging here is in the reply rather than the picture. You send a page; what comes back is a few sentences, and the sentences are the part that needs reading carefully.

The other half of what gets sent is the question typed underneath. This demo holds the photo still and changes only that, three times.

What goes in, what comes back

Image in
1. Photo of the page
Text in
1. Your question
One AI system
Written explanation

Text out

Two inputs, one system, one written answer. Nothing in between stops to ask what you actually wanted to know.

What came back

The same photo, three questions

These three replies were generated once and saved, so every reader sees the same ones. Step between them and watch what changes. The photo does not.

The open question

what’s wrong with this?

This looks like a page of handwritten two-digit subtraction practice, with around six problems worked out.

Most of the work looks right. There may be an error in one of the middle problems, though a couple of the digits are hard to make out.

If you tell me which problem you mean, I can go through it step by step.

An open question got an open answer. It described the page back to you, hedged twice, and asked you to narrow it down. Nothing in it is wrong. It is simply the reply a vague question earns.

Crop first. The model does not need what you would not hand to a stranger, and the third attempt shows what leaving it out costs: nothing. Then ask about one spot rather than the whole page — that is what turned a hedge into a step.

Look closely

A fourth answer, worth reading twice

This one is invented for the demo, written to look like a good answer to the uncropped photo. It is fluent and well organized, and most of it holds up. Three lines claim more than a photograph can support. Tap the ones you would want to check.

The page has six handwritten subtraction problems, worked in pencil.
The fourth, 62 minus 27, is answered 45; the tens column was not reduced after borrowing.
Overall, one slip on a page of otherwise solid borrowing practice.

What to take away

Three things to carry

Crop before you send. The third attempt got the identical answer from a picture with the name trimmed off, so this one costs you nothing. Send the problem, not the page it was sitting on.
Ask about a specific spot, not the whole page. “What’s wrong with this?” gets you a description; naming the problem and saying what you want checked gets you the step. That is one extra clause, and it is the difference between the first two answers.
A confident reading of a smudged digit is still a guess. How steady an answer sounds tells you nothing about whether the page was read correctly, so glance at the numbers it repeats back before you act on anything built on top of them.

Asking about a photo