Study for an exam with AI
Four revision moves: twenty questions from your own notes, marking that names the misunderstanding, a folder per subject, two models on one past paper.
A clear summary can make you feel more prepared than you are. You reread the notes. You ask for a summary. It arrives, it is clear, and you feel ready.
Nothing was tested. Tomorrow the same page feels new again.
Below are four moves we would use instead, in the order we would use them. All four work on a free account. We name the limits as they come up.
Upload the notes and ask for twenty questions
Attach the lecture notes or the PDF. Ask for twenty questions on them, one at a time, with no explanations yet.
The wording does real work here. Ask for a mix: ten straight recall, ten that need two steps. Ask it to save the five hardest for last. Ask it to number them, so you can say "give me 7 again".
What the app allows:
24 of our 59 models accept a PDF attachment. Four of those are free: GPT-5.4 Nano and the three Gemini Flash Lite entries.
46 accept images, so photographing a page of handwriting is a normal way in.
A free account gets one attachment per message at up to 10MB, three a day, inside 50 messages per 24 hours.
The paperclip sits in the composer on the website and in both phone apps.
When the handout is a scan
Know this before you photocopy a term's worth of paper. A scanned handout has no text layer for us to read. So we run OCR on it first, up to 20 pages per PDF, and the model gets that transcription. Transcription can misread a figure, so check the ones you learn by heart. A longer scan skips that step, and a model with native PDF support gets the raw file instead. That is not behaviour we would tell you to plan a revision session around. So keep scans to 20 pages, or photograph the pages and send them as images.
Have it mark your answers, then explain only the misses
Answer first. Out loud, or in the box, before you scroll.
Then paste your answers back and ask for marking. Three instructions make the reply worth reading:
Mark each answer right, wrong or half right.
Explain only the ones I got wrong.
For each of those, name the thing I have misunderstood.
That last instruction is the whole exercise. A list of corrections is something you read once. A named misunderstanding is something you can go and revise.
Then repeat tomorrow on the misses alone. Ten fresh questions, drawn only from the five you dropped.
The marking itself needs checking
A model will sometimes mark a correct answer wrong, fluently, with reasons attached. Push back when you think you were right. Ask it to quote the line in your notes it is marking you against. That settles most of these in one turn. We wrote up why confident wrongness happens in AI hallucinations explained.
Reasoning effort matters on the marking pass, because the model has to follow your working rather than produce its own. 41 of the 59 models expose a reasoning step. On the website the effort button is hidden while Intent Detection is on. That setting is on by default, so the app decides effort itself, message by message. To choose it yourself, open Settings, then Tool Preferences, and turn Intent Detection off. On a free account that button appears greyed and opens an upgrade prompt, so effort stays wherever the model's own default puts it. The phone apps always show the button.
One folder per subject
Make a folder for each subject. Chemistry, French, stats. Free accounts get 25 of them, which is more than a timetable needs.
This is housekeeping, and it pays off around week three. Chats do not share context with each other, so nothing leaks from stats into French on its own. What you gain is finding the thread again in May. Name each one after the topic. "Titration calculations" is findable. "Chem 3" is not.
On a paid plan a folder can also hold a set of memories, and chats you start inside it pick those up. Pro and Pro+ also give each folder a Files section. Add the term's notes there once, and every chat in that folder can use them. Both are paid features, so a free revision folder is organisation and nothing more.
Run two models over the same past paper
This is the move that finds your weak topic without a teacher in the room.
Take one past paper question. Ask a model for a full worked answer. Then put the same question to a second model, cold, without showing it the first reply.
Two ways to do that:
Switch model in the same thread and ask again.
Branch the thread, which opens a new thread you can run on its own.
There is no side-by-side view, so the comparison is you reading two blocks of text.
Then look only at the differences. Where the two agree, move on. Where they diverge, on a number, a method or a definition, that is tonight's revision topic. It is also the point to open the textbook rather than trust either of them.
One catch with PDF past papers
Should the second model not be one of the 24 that take PDFs, it does not receive the file. It gets the text we pulled out of it, wrapped as a plain message. For a clean digital paper that usually reads fine. For a scan of up to 20 pages, it gets our OCR transcription. For a longer scan there is nothing to send.
A week built from the four moves
The order below is our advice rather than a finding from the literature:
Monday: upload the notes, get twenty questions, answer them, get marked.
Tuesday: ten questions on Monday's misses.
Wednesday: one past paper through two models, and write down where they disagree.
Thursday: twenty fresh questions on those disagreements.
Friday: Monday's twenty again, cold.
We would rather spread it over five short sessions than one long one, for a practical reason. You find out what you cannot do while there is still a week to fix it.
Two questions this page leaves alone
Which model suits which study job has a real answer, and we have written it up on its own. So has the question of where the line falls between getting help and handing in work that is not yours.
On what happens to your uploads: we do not train on your conversations. What the provider behind a given model does with them varies by model, and the models where we know of a difference carry a notice on their info card in the picker.
Monday is the only part that needs setting up. Open a chat, attach the notes, ask for the twenty questions.