AI study tools: the ones that help you learn, and the ones that just feel productive

Upfront, because it changes how you should read this: this site is about Gemini Notebook, the Google research tool formerly called NotebookLM. That tool appears below, in the one category it's genuinely best at, and it's explicitly not the answer in three of the others. Every competing roundup we checked while researching this either ranked its own product first without saying so, or was two years stale.

The distinction that predicts whether a tool helps

Every AI study tool sits somewhere on one axis: does it do the cognitive work, or does it make you do it?

This isn't a philosophical preference. It's the oldest well-replicated result in the learning-science literature, and it long predates AI. Robert and Elizabeth Bjork's work on desirable difficulties established that conditions making study harder in the moment — spacing sessions out, mixing topics, testing yourself instead of rereading — reliably improve what you retain months later, while making you feel like you're doing worse. The related generation effect is just as durable: producing an answer yourself encodes it more strongly than reading the same answer. Soderstrom and Bjork's framing is the useful one — performance during a session and learning that survives it are different things, and they often move in opposite directions.

Which is precisely the trap. An AI tool that summarises a 40-page chapter in nine seconds produces an enormous feeling of progress and removes the exact effort that would have created the memory. The tool isn't lying to you. Your sense of having learned something is.

What the research actually found

Four studies are worth knowing, and they disagree in an informative way. None of the competing articles on this topic cite any of them.

The pessimistic one. An MIT Media Lab study wired 54 participants to EEG while they wrote essays either with ChatGPT, with a search engine, or unaided. The ChatGPT group showed the weakest cortical engagement, reported the least ownership of their own work, and — the finding that travelled — recalled about 17% of what they had just written a day later, against roughly 46% for the unaided group. The authors called the effect "cognitive debt." Read it with its limits attached: 54 people, only 18 of whom completed the final session, still a preprint rather than peer-reviewed, and one narrow task. Its lead author has publicly pushed back on the "brain rot" headlines it generated.

The large one. A 2026 analysis of **3.2 million** maths interactions on the ALEKS platform did something cleverer than a lab study: it compared problem types AI can solve (worded questions) against types it cannot (interactive graph manipulation), before and after ChatGPT's release. Time spent learning the AI-solvable problems fell steadily — cumulatively about 27% for college students and 31% for high-schoolers over eleven quarters. On proctored retention questions sat without AI access, the odds of answering correctly dropped 25%. It is quasi-experimental rather than a true randomised trial, so the causal claim deserves some caution, but the scale and the design are hard to wave away.

The optimistic one. A randomised controlled trial published in Scientific Reports found Harvard undergraduates learned roughly twice as much, in less time, from an AI tutor than from in-class active learning. This is the result AI-in-education advocates cite, and they usually omit the design detail that explains it: the tutor was purpose-built to be given correct solutions in advance, to reveal one step at a time, to stay brief, and to refuse to hand over full answers. It was engineered to force retrieval. It was not a chatbot.

The deflationary one. A small randomised trial at Georgetown's medical school gave 33 first-year students either ChatGPT, general web resources, or their own textbooks for two weeks. Retention quiz scores showed no significant difference between the three groups. Too small to conclude much from, but a useful corrective to both louder camps.

Put together, the picture is consistent and not especially ambiguous: AI helps when its design preserves the effortful step, and hurts when it removes it. The variable that matters isn't which model is underneath. It's whether you end the session having retrieved anything.

Five jobs, and what is actually good at each

"Best AI study tool" is an unanswerable question, in the same way "best kitchen implement" is. Studying is at least five separate jobs, and the tool that's excellent at one is usually poor at another.

1. Understanding dense material you must not misquote

The job: a 60-page paper, a statute, a set of lecture slides. You need to interrogate it, and you cannot afford an invented answer.

This is the one case where a source-grounded tool is categorically better than a general chatbot, and it's worth understanding why. Ask ChatGPT about your reading and it answers from everything it has ever seen, with your document as a suggestion. Ask Gemini Notebook — Google renamed NotebookLM in July 2026 — and it answers only from the sources you uploaded, with an inline citation back to the sentence it used. You click the citation and land on the paragraph. For anything you will be examined on or will cite, that difference is the entire ballgame.

Free tier limits, from Google's own documentation: 100 notebooks, 50 sources per notebook, three Audio Overviews a day. Generous enough that most students never reach them.

Where it is not the answer: it will not teach you anything absent from your sources, it has no spaced repetition, and it won't help with a maths derivation. Feed it a thin notebook and you get thin answers — the grounding cuts both ways.

2. Making facts stick

The job: 400 anatomy terms, irregular verbs, case names. Pure durable memorisation.

No AI tool has improved on spaced repetition, and the best implementation remains **Anki**, which is free on Windows, macOS, Linux, Android and the web, with free syncing. The iPhone and iPad app is a one-time purchase of about $24.99 — that single fee funds everything else, and it is not, despite what several roundups claim, an annual subscription. Anki is ugly, its interface is genuinely hostile to beginners, and it beats every polished competitor on the only metric that matters, because its scheduling is built around exactly the desirable-difficulty principle above.

The AI contribution here is real but narrow: generating the cards. Writing 400 good flashcards is the reason most people never start. Having a model draft them from your material and then reviewing them yourself is a legitimate speedup, because the effortful part — answering them, repeatedly, over weeks — stays yours. Quizlet ($7.99 a month, or $35.99 a year) and Knowt are gentler entry points with weaker scheduling. RemNote merges notes and cards for people who want a single system.

3. Capturing lectures

The job: you cannot write fast enough, and transcription is a solved problem. Otter and its competitors handle recording, speaker separation and search well.

One warning that belongs here rather than in the marketing copy: a transcript is not a study aid. It is a longer version of the thing you already failed to absorb, and possessing one produces a strong, false sense of having captured the material. The transcript is raw input — it becomes useful only when it feeds job 1 or job 4. Check your institution's recording policy first; many require consent.

4. Testing yourself

The job: finding out what you don't know, before an exam does it for you.

This is the highest-value use of AI in studying and the most underused, because it's the only one that feels bad. Retrieval practice is the most robustly supported technique in the literature, and generating practice questions used to be a teacher's job. Now anything can do it: Gemini Notebook builds quizzes and flashcards from your sources, ChatGPT and Claude will examine you on a topic if you tell them to, and dedicated tools automate the loop.

The prompt that matters, in any of them: ask to be tested one question at a time, with no answer until you have committed to one. Left alone, a chatbot will helpfully print the question and its answer together, which converts a retrieval exercise into reading. The Harvard trial's tutor beat live teaching largely because it was forbidden from doing that.

5. STEM working

The job: a derivation, an integral, a plotted function. Wolfram Alpha remains the sane choice, because it computes rather than predicts — it shows steps that are correct by construction, where a language model produces steps that look correct. Modern models are far better at maths than they were, and they still fail silently in the middle of long derivations. The ALEKS finding above is specifically about maths, and it is the strongest evidence in this article that outsourcing this particular job is expensive.

What "free" means on each of these

"Free AI study tool" covers at least four different arrangements, and the differences bite at the worst possible moment — the night before an exam, at the point you hit a cap.

• Actually free, permanently. Anki on desktop and Android, including sync. Funded by iOS sales, with no upsell.
• Free tier of a paid product. Gemini Notebook, ChatGPT, Quizlet. Genuinely usable, with ceilings — daily generation caps, source limits per notebook, message limits. Fine until a deadline.
• Free trial dressed as a free tier. Common in the newer study apps: "free" until a credit balance runs out, with the balance undefined and the expiry unstated. Check what a credit actually buys before building a workflow on it.
• Free because you are the product. Anything that wants your uploaded coursework and offers nothing but a login is worth a look at the privacy policy first.

Tool              Free tier                        Paid
----------------------------------------------------------------------
Gemini Notebook   100 notebooks, 50 sources each   $4.99-$199.99 / mo
                  3 audio overviews / day          (bundled in Google AI)
ChatGPT           daily message limits             $20 / mo (Plus)
Anki              everything, all platforms        $24.99 once (iOS only)
Quizlet           flashcards, limited AI           $35.99 / yr
Wolfram Alpha     answers without full steps       $9.99 / mo (Pro)

Prices checked 5 August 2026, USD.

One correction worth making, since two widely-read roundups get it wrong: Gemini Notebook's top tier is $99.99 a month for the 20 TB Google AI Ultra plan, rising to $199.99 for 30 TB — not the $249.99 one competitor lists. It also cannot be bought on its own at any price; every paid tier arrives bundled inside a Google AI subscription. Our pricing breakdown has the full tier table.

When not to use an AI study tool at all

Four situations where the correct move is to close the tab:

• You're inside two weeks of an exam on material you've already covered. At that point nothing beats sitting past papers under timed conditions. Generating fresh summaries this late is displacement activity that feels like revision.
• The assessment is the thinking. If you're marked on constructing an argument, having a model construct it removes the thing being assessed — and, per the MIT result, you won't remember what you submitted.
• The material is short and hard. A dense three-page proof doesn't need summarising. It needs reading four times. Summarisation is a tool for volume, not for difficulty.
• Accuracy is critical and you can't verify it. If you can't check the output against a source, you're not studying, you're gambling. This is exactly the gap source-grounded tools with citations exist to close.

A workflow that puts them in the right order

The sequence matters more than the tool selection, because it decides where the effort lands. What works, in practice:

1. Capture the lecture with a transcription tool, and don't mistake having it for knowing it.
2. Read the actual material first, before any AI touches it. This is the step everyone skips, and the one that makes the rest work — you cannot spot a wrong summary of something you have never read.
3. Interrogate what you didn't understand with a source-grounded tool, following the citations back to check every answer against the original.
4. Generate flashcards from the parts you got wrong, not from everything. Review and correct the cards yourself; that pass is where a surprising amount of the learning happens.
5. Drill them in a spaced-repetition system over weeks. Non-negotiable, and the part with the strongest evidence behind it.
6. Test yourself with generated questions, one at a time, answers hidden. Whatever you fail goes back to step 4.

Steps 1 to 3 are where every AI study product concentrates its marketing, because they're the pleasant ones. Steps 4 to 6 are where the learning is. A tool that makes 1 to 3 effortless while you skip 4 to 6 has made you faster at not learning.

People also ask

What is the best AI tool for studying?

There isn't one, because studying is several different jobs. Gemini Notebook is best for understanding source material you must cite accurately; Anki for long-term memorisation; Otter for capturing lectures; Wolfram Alpha for STEM working. The most valuable thing any of them does is test you — that's the part with real evidence behind it.

Do AI study tools actually help you learn?

It depends entirely on whether the tool preserves the effort. A randomised trial found students learned about twice as much from an AI tutor built to withhold answers and force step-by-step work. Separately, an MIT study found essay writing with ChatGPT cut next-day recall to roughly 17%, and an analysis of 3.2 million maths interactions found retention dropped 25% on problems AI could solve for students. Same technology, opposite outcomes, decided by whether you did the retrieving.

What's the best free AI study tool?

Anki, if you mean permanently and genuinely free — it costs nothing on desktop, Android and the web, including sync, and only the iOS app is a one-time purchase of about $24.99. Gemini Notebook has the most generous free tier among the source-grounded tools, at 100 notebooks and 50 sources each.

Is using AI to study cheating?

Using it to produce work you submit as your own generally is, and rules vary by institution — check yours. Using it to test yourself, explain a concept or generate flashcards from your own materials is ordinary studying, and it's also the use with the better evidence behind it.

Can ChatGPT replace a tutor?

Not by default, though it can come close if you constrain it. The trial where an AI tutor beat classroom teaching used a system built to reveal one step at a time and never hand over a full solution. An unconstrained chatbot does the opposite. Instruct it to ask one question at a time and withhold answers until you have attempted them.

What's the difference between Quizlet and a source-grounded tool like Gemini Notebook?

Quizlet is a flashcard system with a large library of decks other people have made. Gemini Notebook generates material only from documents you upload, and cites which passage each answer came from. Use the second when accuracy against your own reading matters, and the first when you want quick cards on a common topic.