Which tool to choose
The choice depends on how much you trust the result and how long the material is.
- Long handout, serious exam, you need to be able to check every sentence: use a source-grounded tool like Google's NotebookLM. It answers only with what you've loaded and places next to each statement the reference to the exact point in the text. It's the best defense against fabrication.
- Scattered notes, photos of the whiteboard, photographed pages: a general-purpose AI assistant that reads images (the latest free version of ChatGPT, Claude or Gemini) transcribes and reorganizes in one go.
- Quick summary of a few pages you already know: any assistant will do, here the risk of losing the meaning is low because you already know what's in there.
How to do it
From a browser or an app, the path doesn't change in substance: you load the material, give a precise instruction, check.
- Gather the material in one place. If they're photos of notes, take them straight and in focus: a crooked or blurry page becomes a wrong transcription.
- Load everything into the tool. In NotebookLM you create a new workspace and drag in PDFs, documents or pasted text.
- Ask for the summary by concepts, not by quantity. The operational syntax:
Summarize this handout by key concepts, not page by page.
For each concept: a one-sentence definition, the reason it matters, an example if it's present in the text.
Use only the information from the loaded material. If a point isn't there, write "not present in the handout."
- Read the summary and open the citations of the three points you'd use in the exam. In source-grounded tools, each sentence has a clickable number that takes you back to the original passage. If the summary's sentence says one thing and the text says another, you've found a distortion: correct it by hand.
- Ask the AI to highlight what it cut. The operational syntax:
List the concepts from the text that you excluded from the summary and explain in one line why you judged them secondary.
This step is the one that saves the meaning: a wrong summary doesn't add lies, it removes them. Seeing what was removed tells you whether it removed the right thing.
A concrete example
Marta has forty pages of a private law handout and three days until the exam. She loads the PDF into NotebookLM and asks for a summary by concepts. She gets twenty entries, each with the reference to the paragraph. She opens the entry "contractual good faith": the summary says "it applies only in the performance phase." She clicks the citation and reads that the handout talks about good faith in negotiations too. The AI had narrowed the concept. Marta corrects her card and at the oral exam answers completely. Without checking the citation she would have studied a halved version.
When it does NOT work (and how to fix it)
If the AI invents a concept that isn't in the handout
This happens with general-purpose assistants, which fill the gaps with what they know from their training. Remedy: use a source-grounded tool, or add to the prompt the line "use only the loaded text, not your general knowledge" and then spot-check the statements you don't recognize.
If the transcription of the photos gets numbers or names wrong
A formula, a date or a proper name read wrong ruins the review. Remedy: after the transcription ask "reread and flag every number, date or proper name you're unsure of due to the image quality," then double-check those points against the original.
If you can't find the button to upload the files
Interfaces move around often. Remedy: look for a paperclip-shaped icon, a "+" or a "Sources" entry. If you really can't find it, paste the text directly into the chat box and write "treat this as the sole source, answer based only on it."
A tip from someone who actually uses it
Never ask "summarize everything." A good summary comes from an instruction that says what you need it for: reviewing before an oral exam is different from preparing an outline to memorize. Tell the AI what the summary is for and you'll get a different, more useful cut, and less material to discard.
Frequently asked questions
How long should a good summary be?
As long as it takes to reconstruct the reasoning, no shorter. A practical rule: if, rereading the summary, you wouldn't be able to re-explain the concept aloud, it's too short. Ask the AI to expand only the entries that are unclear to you.
Can I trust the summary to study directly from it?
For reviewing, yes; for studying for the first time, no. The first pass over the material serves to build your mental map; the summary is a shortcut that only works if you already have the map. Use it to fix things in place, not to replace the first reading.
Is it true that summarizing with AI makes you study less?
It's the opposite of how it's sold. The AI cuts the time of transcription and outlining, not that of comprehension: checking the citations and correcting the distortions require reading the text carefully. Those who skip that check study less, and badly.