AI Notes vs Writing Notes Yourself

For studying, choose an AI page only, a page you write, or a hybrid.

A note has three jobs. Capture gets the lecture or the chapter into a record you can reopen. Encoding is the selecting and phrasing you do while producing the note: what belongs, in words you will later have to produce. Retrieval comes later, with the page closed. Producing the idea from a cue, then checking, is a separate act from reading a finished page.

AI is often the better first capture: a long lecture, a moment when writing at that pace is a poor fit, or a first pass over unfamiliar material. Writing the note yourself is the study act when you must select what belongs and phrase what you will later produce — exact wording, a formula, or a condition a draft is likely to smooth over. The useful hybrid is an AI draft you rewrite, or a page you write with AI used only for draft questions. Do not leave the untouched AI page as the version you review from.

A transcript or a generated summary is storage. Reading it until it looks familiar is a pass over someone else's sentences. In the studies below, students who compared an LLM with note-taking mostly preferred the LLM, while retention and comprehension three days later were higher for note-taking, alone or with the LLM, than for the LLM alone. In a separate lab, people preferred the more automated setup because it felt easier, while immediate scores were higher for notes built from shorter summaries than for notes built from structured AI blocks.

The three jobs of a note

The three jobs can be split across an AI page, a page you write, and a hybrid. The running source is one textbook section: why a binding price ceiling creates a shortage.

The section's chain is short. A price ceiling is a legal maximum price. It binds when that maximum sits below the price at which quantity demanded equals quantity supplied. At that lower legal price, quantity demanded is higher than at the market-clearing price, and quantity supplied is lower. The gap is a shortage. A ceiling set above the market-clearing price does not bind, and the section does not treat it as creating that gap.

On this section, capture means the definition, the binding condition, and the two quantity movements are somewhere you can reopen. Selecting and phrasing means putting the chain in an order you can produce: legal maximum, below the market-clearing price, quantity demanded up, quantity supplied down, the gap, and the case that does not bind. Later retrieval means covering the page and producing that chain from the heading. That later pass is how to review notes effectively: repair the page while the source is still clear, then retrieve with it covered.

An AI page can be useful storage and still skip the encoding that comes from writing it. Producing notes can help a little even before you review them. Kobayashi (2005) found a positive but modest encoding effect of taking notes versus not taking notes: a mean weighted effect size of .22 across 57 studies. Keeping notes and reviewing them is a different benefit (Kiewra 1989). The .22 compares taking notes with taking none. It says nothing about an AI page.

When an AI note is the better first draft

Use an AI draft when the bottleneck is getting the source down. Selecting and phrasing still come before you rely on the page.

A long lecture

Speaking speed and careful phrasing are different jobs. If you spend the hour composing the shortage paragraph, you can miss the graph, the case that does not bind, or the example the lecturer used. A transcript or a generated page can hold that pass. The study act starts when you return and decide which sentences belong on the page you will review.

When writing at that pace is a poor fit

A draft is also the right first tool when live capture is a poor fit for how you take in the source. That includes a long recording, a dense PDF, a section you are meeting in a second language, or a day when the goal is to get the material into a form you can edit. The draft lowers the capture cost. What the section claims is still your decision, after the draft exists.

A first pass over unfamiliar material

Before you know the subject, a central condition and a side comment can look alike. On a first reading, "price controls cause shortages" can look like the whole lesson. A draft gives you a map to check against the section: where it defines the ceiling, where the maximum has to sit below the market-clearing price, and where the two quantity movements appear. You still make that check. The draft is the first tool. The check, and the phrasing you keep, are the study.

Stop at the clean summary and you have stored the section in fluent prose. You have not yet selected the binding condition or phrased the chain you will have to produce.

When writing the note yourself is the study act

Write the page yourself when the valuable work is the selecting and the phrasing. A draft would hand you a finished version of that work.

You already know what belongs

After the lecture, or after one careful reading, the ceiling chain is no longer a search. The page needs the legal maximum, the binding condition, the two quantity changes, and the case that does not bind. Writing those lines is practice at producing the explanation. A generated page of a section you already understand mostly saves you that practice.

Wording and details you must produce as written

Some lines have to come back in a fixed form. The gap as quantity demanded minus quantity supplied at the ceiling. A definition the course states in set words. A label on the graph. A worked example that uses the section's own prices and quantities. Those lines are worth writing yourself. Recognizing a model's sentence with the page open is a weak forecast of producing the expression later.

When a draft is likely to smooth the condition over

Watch this section for a smooth miss. The draft says price ceilings cause shortages, and it drops the binding condition, so a ceiling above the market-clearing price is described as if it created the gap. If the section's point is that condition, place it yourself, against the textbook, in words you chose. Check any draft against the section before a sentence becomes the version you trust.

A page you already have

The note is already the record and the phrasing. The missing piece is a cue for later retrieval. Draft questions from the page you wrote, then edit them until a prompt can be missed. "When does this ceiling bind?" can be missed. "Explain price ceilings" often cannot. Retrieve from the page you already trust. Active recall with notes is that closed-page attempt. The page stays the page.

A hybrid that does not leave the AI page as the version you review

Two sequences. Both keep a human pass over the lines you will rely on.

Draft, then rewrite what you will review. Paste the ceiling section, or start from the lecture transcript, and let the model produce a first page. Edit it against the source. Keep a sentence the section supports. Rewrite the binding condition in your own words. Put the non-binding case on the page if the draft omitted it. Cut a mechanism the section does not claim. The page you save for review is that edited page. The untouched draft can stay in the chat as a record of the first pass. Review from the edited page, where you had to decide what the shortage depends on.

Write the page, and use AI only for draft questions. You compose the ceiling note. Then you ask for recall questions on that page. Keep the prompts that force the chain. Repair the ones that are too broad to miss. Leave the questions optional. The learning unit is still the page. A question is a cue.

Once a draft exists, how to use AI for studying is the work after that draft. This page is the prior choice: who produces the note.

A hybrid that stops at generate-then-reread still has only storage. The sentences are fluent. Fluency with the page open is a weak forecast of producing the shortage chain with the page closed.

What the studies actually compared

These studies compared specific setups. They do not show that you should always write the note yourself. In the school study, the activities were using an LLM, taking notes, or doing both, and the test came three days later. In the lab, everyone still composed notes, and the result is an immediate test.

In a pre-registered school experiment, Kreijkes and colleagues studied 405 students aged 14–15. The students studied two passages and were tested three days later. Both note-taking alone and note-taking combined with an LLM had significant positive effects on retention and comprehension compared with using the LLM alone. Students were split into two comparisons. In the group that compared the LLM with note-taking, most preferred the LLM, and they rated it more helpful for understanding than note-taking. In the group that compared the LLM alone with the LLM plus notes, most preferred the combination, and helpfulness did not differ between those two. This was chatbot help while reading, not a test of a finished AI study page against a page the student wrote.

Chen, Ruan, Ju, Yap, and Wang ran a within-subject lab with 30 people. Everyone still composed notes during lecture videos of about 10 minutes. Immediate closed-book scores were higher when people built notes from shorter AI summaries (mean 13.88, SD 2.34) than when they composed notes from structured AI blocks (mean 10.28, SD 3.75). The Tukey mean difference was 3.61. When they then revised answers with their notes, means were 15.29 versus 12.12, difference 3.17. Participants preferred the automated setup because it felt easier and took less effort. This is an immediate test. It is not a test of reading a finished AI page with no composing, and it is not a delayed exam.

Preference and the test did not always match. Where students compared the LLM with note-taking, they preferred the LLM and called it more helpful, while retention and comprehension were higher for note-taking than for the LLM alone. Where they compared the LLM alone with the LLM plus notes, they preferred the combination, which also scored higher than the LLM alone on retention and comprehension. In the lab, participants preferred the automated setup because it felt easier and took less effort, while the higher immediate scores were for notes built from shorter summaries.

People often remember a word better when they had to produce it than when they only read it (Slamecka and Graf 1978; Bertsch et al. 2007, overall effect about .40), which is item generation, usually of words, and not a measurement of an AI note against a note you write.

What this comparison is not

Pen versus keyboard answers a different question. Some lecture studies found that typing word for word was weaker for conceptual questions than writing more selectively by hand (Mueller and Oppenheimer 2014). A direct replication did not find a consistent longhand advantage (Morehead, Dunlosky, and Rawson 2019). The contrast is how selectively a person writes during a lecture. An AI page versus a note you compose is a separate choice.

Notes versus flashcards is a format choice about connected explanations and independently testable items. Who should produce the page is a different decision. Which study app to buy is a product choice, and it leaves the same question open: whether the page you review from is one you selected and phrased.

How Yuhimo fits this choice

Yuhimo can take either hybrid path. The selecting, the phrasing, and the retrieval judgment stay with you.

Creating study material with AI drafts pages from a topic, pasted text, or a PDF or DOCX file. You review an outline, then the pages. Nothing is added to your notes until you save. That generation is online-only. On the ceiling section, the draft is the capture. Before you save the page you intend to review, rewrite the binding condition and the quantity chain so the saved page is the one you phrased. Saving does not turn learning on. Generation does not turn it on either. Learning stays optional.

If the page is already yours, AI question generation drafts recall questions from that page. Questions are optional prompts. The page stays the learning unit. You can review the page without them.

During review you judge your own recall: Forgot, Barely, Remembered, or Easily. AI does not grade answers, decide what you know, or schedule reviews. Those limits are in what AI does not do. A clean generated page can sit in your notes as storage. The review is still your attempt to produce the shortage chain, and your judgment of how that attempt went.