Give AI your book page, your notes or a photo

It is nine at night. Question 7 in your science book will not come out, so you ask the AI. You type the question from memory, miss one line, and get a neat answer that does not match the book.

There is a better move. Show it the page.

Why the page beats your memory

The machine does two very different jobs. One job is remembering something it read long ago. The other job is reading something you just handed it.

It is much better at the second job.

Tests that give a model a document and ask it to stick to that document are updated through 2026. On those tests the strongest models contradict the given text in only a few answers out of a hundred. Now look at the other job. A company published its own test of short factual questions in 2025, with nothing supplied. Its bigger model was wrong about half the time it answered. Its small cheap model was wrong far more often than that.

Be careful with those low numbers. They only apply when you supply the text. They are not a general score for the app.

So the rule is short. Give it the paper. Do not ask it to remember the paper. This is the cheapest accuracy upgrade in the whole course, and it costs you nothing but one photo. It works best when you also ask well, so keep using the five things every good request needs.

Think about where this changes your day. A letter from the bank in English. A notice pinned up at the block office. Two pages of your own class notes the night before a test. In every case the paper is already in your hand. Hand it over instead of describing it.

One warning still holds. For a government scheme or a form, the paper you are holding may be an old one. The office itself, or the official website, is still the final word.

How to take a photo it can read

A bad photo makes a confident wrong answer. Six things fix almost every bad photo.

Press the book flat with your hand or a weight. A curved page bends the letters and the machine guesses at them.

Use daylight. Stand near a window during the day, or under a bright tube light at night. Then move so your own head and your phone are not throwing a shadow across the words.

Take one page at a time. Two pages in one picture means smaller letters and more mistakes.

Crop it. Cut away everything except the question you actually want. A small clear picture of one question beats a wide picture of the whole page.

Hold still and tap the screen on the text so the camera focuses. Blur is the worst enemy here.

Sometimes it still reads the page badly. Do not send the same photo again. Take a fresh one with more light, crop it tighter, and keep only the lines you need. Changing the picture works. Repeating it does not.

Hindi in Devanagari needs a better photo than English does. The letters join at the top, and a small blur turns one letter into another. If your page is in Hindi, take two photos and send the sharper one. There is more on working in your own language in the lesson on using AI in Hindi.

Photo first, then your question

Order matters, and almost nobody is told this.

Attach the picture first. Wait for it to finish loading. Then type your question in the space underneath it and send. Google’s own guidance for its models says to put the image first and the text after it.

The same idea works for typed notes. If you are pasting your class notes, a long message, or a notice from an office, paste that block at the top. Write your question at the bottom, under the block. Anthropic’s guidance says to put long material near the top, above your question, and that this improves answers across models.

Here is the shape to copy.

[paste your full notes or the whole notice here]

My question about the text above:
Make a table of the important dates. Columns: Date, What happens.
Use only the text above. If something is missing from the text,
write NOT IN THE TEXT.

That last line matters. It tells the machine to stay inside your paper instead of filling gaps from memory.

Make it type out the page first

This is the step that separates a useful answer from a costly one. Never let it answer a photo straight away.

Ask it to type out what it sees, then answer. Copy this.

I am sending a photo of one page from my science book.
Step 1: type out all the text you can see in the photo.
Step 2: tell me any word or number you could not read clearly.
Step 3: explain this page in simple words.
Step 4: answer question number 7.

A good reply starts like this.

Step 1, what I can see:
"Exercise 4.3, Question 7. A shopkeeper buys 3 kg of rice at
Rs 48 per kg. He sells it at Rs 55 per kg. Find his profit."

Step 2, not clear: the number after the first Rs could be 48 or 46.

Now do your part. Read that typed-out text against your book, number by number. If your book says 46 and it read 48, the whole answer will be wrong. It will still sound calm and sure, because sounding sure is free for it.

For a diagram, use words instead of a photo

Pictures are where this machine is weakest, and a lot of school questions are pictures.

In a 2025 test of physics questions, a model got about eight in ten right when the question was written out in words. On questions that could only be answered by reading the picture, it got about half right. Practical and laboratory questions were the worst group of all.

A 2026 test showed why. A model could name the parts inside a diagram very well. It was far worse at working out how those parts are joined together. It sees the pieces. It does not see the wiring.

So for a circuit, a ray diagram or a figure, do not send a photo. Type the picture out as a sentence. Like this.

A circuit has a 6 V battery. From the positive terminal, a 2 ohm
resistor is in series. Then the wire splits into two branches.
One branch has 3 ohm and one branch has 6 ohm. They join again
and go back to the battery. Find the current from the battery.

A ray diagram works the same way. Say which mirror or lens it is. Say where the object stands, and where the focus sits. Put the distances in as numbers. Then ask for the position of the image.

Writing that took one minute. It removes the machine’s worst weakness. It is also a test of you, because a student who cannot describe the circuit does not yet understand the circuit.

The same rule runs the other way too. Never copy a diagram the AI drew into your notebook, as the lesson on science diagrams explains.

What you must never photograph

Some pictures should never leave your phone. Learn this list before you get comfortable with uploading.

Do not send other people’s papers either. Your neighbour’s land record, your sister’s marksheet, a relative’s medical report. Those are not yours to hand over, even to help.

Do not photograph a friend’s answer sheet or a paper your teacher has not released. That is a different problem, and the lesson on when AI use is cheating draws the line clearly.

Think of it simply. A photo you send leaves your phone and lands on a company’s computers. Treat it like giving a paper to a stranger at a bus stand. The full list of what to hold back is in what you must never type into an AI.

What a photo costs you

A photo is many times heavier than a line of typed text. On a small daily pack you will feel it, and on a slow phone the upload also takes time.

Three habits keep the cost down. Crop before you send, because a smaller picture is a lighter picture. Send one photo instead of five. Type the question out when it is short enough to type.

There is a second cost people miss. Free plans put a limit on how much you can do in a day. Photos and files may have their own limit, apart from ordinary messages. Open the app’s own help page and read its current free limits before you plan a big upload. Plan the important photo first. Getting the most from a small data pack has more on stretching what you have.

Even a perfect photo does not promise a correct answer. It only removes one big reason for a wrong one. When the answer still comes back wrong, do not argue with it, and do not just say it is wrong. There is a way to say it that works, and that is how to correct a wrong AI answer.