Let’s talk about artificial intelligence and learning without the hype. There is no revolution under way that will solve education. There are concrete tools, with uneven evidence, that do specific things. Some work very well. Others cost more than they give back. And many are free.
Let’s start with two experiments with GPT-4 that produced opposite results. At a secondary school in Turkey, students who practiced with an unrestricted chatbot scored lower on the final exam than those who didn’t use it [1]. In a physics course at Harvard, students who learned a topic with an AI tutor learned more, and in less time, than in an in-person active-learning class [2]. Same language model, different methodology.
What the evidence says
The Turkish experiment followed almost a thousand secondary school students in three groups: no AI, a ChatGPT-style interface, and a version programmed to give hints instead of solutions. During practice, the regular chatbot raised performance by 48% and the hint version by 127%. In the later exam, without AI, the first group scored 17% lower than those who never used it [1].
The OECD’s Digital Education Outlook 2026, which synthesizes evidence from several countries, reaches the same point. General-purpose chatbots raise performance on the task without necessarily making people learn more, and with pedagogical guidance they do support learning [3]. An experiment with 117 university students confirms it. The ChatGPT group improved its essay more, but knowledge gain and transfer to another task were no different from those of other groups [4].
A 2026 meta-analysis of 49 STEM studies finds a positive average effect (g = 0.84), with enormous variation between studies. After correcting for publication bias, the average drops to 0.08. What holds after the correction is the pattern: effects are larger when the goal is acquiring knowledge than when it is developing skills, and AI helps when it adds to the student’s activity and can hurt when it replaces it [5]. A 2025 theoretical framework sorts this into four cases. AI can invert the student’s mental activity (they think less), replace it, augment it or redefine it [6].
Socratic tutors don’t give you the answer: they ask you questions and guide you. Source-grounded assistants answer only from the documents you give them. AI search engines synthesize information from the web with links. Symbolic calculators solve and verify math. Spaced-repetition programs decide when to ask you something again.
Free tools worth using
Gemini Notebook (formerly NotebookLM). Google changed the name in July 2026 and the product is the same [7]. You upload your sources (PDFs, web pages, YouTube videos, audio) and it answers with citations pointing to your material, so you can check every claim against the original. It generates audio and video summaries, mind maps, flashcards and quizzes. If you get a question wrong, an explain button justifies the answer with citations from your sources [8]. The free plan includes these features with daily quotas. It is the best option when you have notes and a closed syllabus.
The study modes of ChatGPT, Gemini and Claude. They do essentially the same thing: they ask before they explain. ChatGPT launched its mode in July 2025, also for free accounts [9]. Gemini added “Guided Learning” a week later, with steps, images and quizzes [10]. Claude has a “Learning” style in its style menu, available to everyone since August 2025 [11]. You turn them on manually. If you don’t, you have the regular chatbot from the Turkish experiment.
Anki. Flashcards with spaced repetition. Free on desktop and Android, one-time purchase on iPhone, and since 2023 it has included an optional review-scheduling algorithm called FSRS [12]. If Anki feels awkward, Quizlet is simpler to start with [13]. The combination that works is asking a chatbot to turn your notes into flashcards and letting Anki decide when to ask you.
Perplexity. It searches the web and answers with links to its sources [14]. Use it to find current information and articles, and open the links it cites before trusting the synthesis.
Elicit and Consensus. These are search engines for scientific literature. Elicit searches over a hundred million papers, summarizes them and extracts data from the studies [15]. Consensus answers questions from the published literature [16]. Both have free plans with limits. The difference from a general chatbot is in the question you ask. Instead of “is intermittent fasting good?”, ask what trials and meta-analyses published since 2020 say about weight loss with intermittent fasting.
Wolfram|Alpha. For math and quantitative sciences it is still the reference. It is not a language model: it calculates and shows the procedure [17]. Use it to check a calculation a chatbot gave you.
Paid tools: when paying makes sense
ChatGPT Plus, Claude Pro and Google AI Pro cost roughly 18 to 23 euros a month in Europe, VAT included [18]. What you buy is more daily usage and more powerful models. You are not buying a different study method. In Gemini Notebook, paying mainly raises the daily quotas for audio, video, flashcards and quizzes. Perplexity, Elicit, Consensus and Wolfram|Alpha also have paid plans with more searches, more data extraction or step-by-step solutions.
My criterion is usage. For studying a normal syllabus, I think the free tools are enough. Paying makes sense if you hit the limits every day, if you work with many long sources, or if you do literature reviews with Elicit or Consensus. That is an opinion, not a fact.
Simple tutorial for using Gemini Notebook
It works with a chapter, a topic or a whole syllabus. Allow 40 to 60 minutes the first time. After that, 20 or 30 minutes per session is enough.
Before you start. Create one notebook per subject or topic, not one for everything. The fewer sources you mix, the more focused the questions will be. Before uploading anything, check that your notes are legible: a crooked scanned PDF or a blurry photo will give worse results than clean text.
- Write down everything you know about the topic from memory, without looking at anything, for five minutes. Do it on paper or in a separate document. It doesn’t matter if it comes out messy or incomplete. The gaps are the starting point of the method, because they tell you what you don’t know yet. If you go blank after two minutes, stop. That is also information.
- Upload your notes to Gemini Notebook, or paste them into a chatbot with study mode turned on. It accepts PDFs, web pages, YouTube videos and audio. Give each source a clear name (“Topic 4 – class notes”, “Book – chapter 2”), because the answers cite where each fact comes from and this way you will recognize it. Start with the sources for the topic you are studying and add more only if needed. If you work in a general chatbot instead of Gemini Notebook, paste the excerpts and add: “Use only this material. If something isn’t there, tell me.”
- Ask it to quiz you, not to explain. A prompt that works: “You are my tutor. Ask me one question at a time about this material. Don’t give me the answer. If I get it wrong, give me a hint and let me try again. When I get it right, explain the why in two sentences and move on to the next one.” Answer before looking at anything. If you hesitate, write an approximate answer instead of asking for the solution. Trying to recall, even if you fail, is what fixes learning [19]. There are two useful variants for later. The first raises the difficulty: “Mix easy and intermediate questions with some that make me apply the concepts to a new situation.” The second checks that the question comes from the material: “Indicate which source each question comes from.”
- Compare what you wrote in step 1 with the material and ask it to point out what you were missing. Paste your text and write: “Compare this with the sources. Tell me which important concepts I haven’t mentioned and what I’ve said that is incorrect or imprecise.” Note down the gaps. They are the list of what you need to reinforce, and they keep you from reviewing what you already know.
- Turn your mistakes into flashcards, using Gemini Notebook’s or Anki’s, and review with spacing. Ask: “Generate flashcards only with the concepts I got wrong. One idea per card, with a question that makes me recall, not recognize.” Check each one before saving it. If a card is confusing or too long, edit it or discard it. Move the cards to Anki if you want long-term review. Anki schedules the intervals for you. The longer you want to remember something, the more you should lengthen the time between reviews [20]. If you’d rather not install anything, Gemini Notebook’s flashcards work for reviewing during the few days before an exam.
- Once a week, give yourself an exam with no help at all and write down your score. Close the notebook and the chatbot. Answer in writing, with the time you would have in a real exam. Grade it afterward with the material in front of you. It is the same measurement the studies used with the final exam: what counts is what you can do without the AI beside you. Record the score week by week. If it goes up, the method works and you can keep it. If it doesn’t, go back to step 3: it almost always means you are asking for explanations instead of questions.
Perhaps the common ingredient of the tutors that worked is forcing you to attempt the answer before receiving it. In the Turkish experiment both tutors used the same model and only the instructions changed [1], and the STEM meta-analysis finds that AI helps when it adds to the student’s activity [5]. If so, any decent model with the right instructions should reproduce a good part of the benefit.
Audio summaries are good for a first pass, for example on a commute, but they do not replace retrieval, because listening is passive [19]. As far as I know, there are no controlled trials of them with delayed exams.
References
[1] Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122 — Reliable
[2] Kestin, G., Miller, K., Klales, A., Milbourne, T., Ponti, G. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/ — With reservations
[3] OECD (2026). OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education. OECD Publishing. https://www.oecd.org/en/publications/2026/01/oecd-digital-education-outlook-2026_940e0dd8.html — Reliable
[4] Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. Preprint: https://arxiv.org/abs/2412.09315 — With reservations
[5] Boolzen, C., Kuhn, J., Flegr, S., Rott, E.-M., Stausberg, N., Küchemann, S. (2026). Evidence of impact and interpretational limits of generative AI in STEM education: a systematic review and meta-analysis on cognitive learning outcomes. Artificial Intelligence Review, 59, 223. https://doi.org/10.1007/s10462-026-11665-9 — Recent
[6] Bauer, E., Greiff, S., Graesser, A. C., et al. (2025). Looking beyond the hype: Understanding the effects of AI on learning. Educational Psychology Review, 37, 45. https://opus.bibliothek.uni-augsburg.de/opus4/frontdoor/deliver/index/docId/121589/file/121589.pdf — Theoretical framework
[7] Google (16 Jul 2026). NotebookLM is now Gemini Notebook. https://blog.google/innovation-and-ai/products/gemini-notebook/notebooklm-gemini-notebook/ — Product documentation
[8] Google (2025). 6 NotebookLM features to help students learn. https://blog.google/innovation-and-ai/models-and-research/google-labs/notebooklm-student-features/ — Product documentation
[9] OpenAI (29 Jul 2025). Introducing study mode. https://openai.com/index/chatgpt-study-mode/ — Product documentation
[10] Google (6 Aug 2025). Guided Learning in Gemini: From answers to understanding. https://blog.google/outreach-initiatives/education/guided-learning/ — Product documentation
[11] Engadget (Aug 2025). Anthropic brings Claude’s learning mode to regular users and devs. https://engadget.com/ai/anthropic-brings-claudes-learning-mode-to-regular-users-and-devs-170018471.html — Product documentation (press coverage)
[12] Anki. https://apps.ankiweb.net. FSRS integration: Anki 23.10 release notes, https://changes.ankiweb.net/changes/23.10.html — Product documentation
[13] Quizlet. https://quizlet.com — Product documentation
[14] Perplexity. https://www.perplexity.ai — Product documentation
[15] Elicit. https://elicit.com — Product documentation
[16] Consensus. https://consensus.app — Product documentation
[17] Wolfram|Alpha. https://www.wolframalpha.com — Product documentation
[18] Notebookcheck (12 Jul 2026). Cuánto costarán las suscripciones a servicios de IA en 2026 y cuál merece la pena. https://www.notebookcheck.org/Cuanto-costaran-las-suscripciones-a-servicios-de-IA-en-2026-y-cual-merece-la-pena.1341641.0.html — Product documentation (press comparison of official prices)
[19] Rowland, C. A. (2014). The effect of testing versus restudy on retention: A meta-analytic review of the testing effect. Psychological Bulletin, 140(6), 1432–1463. https://doi.org/10.1037/a0037559 — Reliable
[20] Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. Psychological Bulletin, 132(3), 354–380. https://pubmed.ncbi.nlm.nih.gov/16719566/ — Reliable
Notes on the labels.
- [2]: a single course, a short topic and students from a highly selective university. I found no replications.
- [4]: 117 participants and a single writing task.
- [5]: August 2026 meta-analysis. The authors rate the included studies as moderate-to-high risk of bias, with none at low risk.
- [6]: cited for its framework, not as proof of effects.
- [7]–[18]: describe features, launch dates and prices as of October 4, 2026, and these change often.