The best way to learn in 2026
The two techniques that lead the evidence have not changed since 2013, and in 2026 software can run them for one learner at a time.
The best way to learn in 2026 is practice testing and distributed practice, the two techniques Dunlosky and colleagues (2013) rated high utility out of the ten they reviewed. Read once, get tested the same day, explain the concept back the next day after sleep, then review on a spaced schedule.
AI changed who can run that routine for you. In a randomized trial at Harvard, median learning gains more than doubled with an AI tutor designed around pedagogical best practices (Kestin 2025). In a field experiment with high-school students, an AI that supplied answers left exam scores 17% lower (Bastani 2025).
The best way to learn in 2026 is a routine built on two techniques
The best way to learn in 2026 is to test yourself on the material and to spread the sessions across days. Dunlosky and colleagues (2013) graded ten study techniques against the experimental evidence and rated exactly two as high utility: practice testing and distributed practice. Both predate every AI study tool on the market.
What 2026 added is a machine that will run those two techniques for you, for one learner, at whatever hour you sit down. That matters because both techniques are hard to run alone. Writing your own quiz means knowing the answers first. Spacing your reviews means tracking what is due on which day, for months.
Two randomized studies mark the range of what AI does to learning. Kestin and colleagues (2025) compared an AI tutor designed around pedagogical best practices with an active-learning physics class at Harvard, and median learning gains in the AI group were more than double. Bastani and colleagues (2025) gave nearly a thousand high-school students a ChatGPT-like interface, and their scores on a later unassisted exam fell 17% below students who had no AI at all. The same technology produced both results. The best way to learn in 2026 is therefore a routine you run, and you judge any tool by whether it runs that routine or replaces it.
What the evidence ranks highest, and what it ranks lowest
Practice testing and distributed practice are the only two techniques the 2013 review rated high utility. Dunlosky and colleagues graded each technique on how broadly its benefit generalizes. Three techniques landed in the middle: elaborative interrogation, self-explanation and interleaved practice. Five landed low: summarization, highlighting and underlining, the keyword mnemonic, imagery for text, and rereading.
The low group contains the methods most students actually use. Reading a chapter twice and running a highlighter down it are the default study habits in most schools, and they sit at the bottom of the ranking.
An AI can run any of the ten. Which one it runs is a decision its designer makes, and a general chat window defaults to the low ones: it summarizes, it restates, it produces a clean page you read once and close.
| Technique | Dunlosky rating | What it looks like when an AI runs it for you |
|---|---|---|
| Practice testing | High | It quizzes you the same day you read, with the lesson closed, and grades your answer against it |
| Distributed practice | High | It puts the next review days out, tracks what is due, and widens the gap each time you recall an item |
| Self-explanation | Moderate | It asks you to state the mechanism in your own words, then marks what you left out |
| Interleaved practice | Moderate | It mixes questions from several concepts into one session instead of drilling one concept |
| Rereading | Low | It regenerates the same explanation whenever you ask for it again |
| Summarization | Low | It hands you a summary that you did not have to write |
Why rereading and highlighting feel like learning
Rereading feels like learning because familiar text is easy to read, and you read that ease as memory. Roediger and Karpicke (2006) had students read prose passages and then either take a recall test or study the passage again. On a final test five minutes later, the restudy group scored higher. On final tests two days and one week later, the tested group retained substantially more. Restudying also raised students' confidence about what they would remember, so the method that felt strongest at the moment of study left the least behind a week later.
Karpicke and Blunt (2011) measured that misjudgment directly. On a test one week later, students who practiced retrieval scored 0.67 against 0.45 for students who built concept maps, and 101 of 120 students did better after retrieval practice. Before the test, 75% of them had predicted that concept mapping would be equal or better. The feeling of fluency is a poor measure of what you will recall a week later, which is why learning fast and learning comfortably rarely happen together.
What changed in 2026: a tutor for one learner
What changed is the cost of teaching one learner at a time. Bloom (1984) compared students tutored one-to-one under mastery learning with students in conventional classrooms and found a gap of about two standard deviations. The average tutored student scored above 98% of the conventional class. Bloom named the search for a scalable method with that effect the 2 sigma problem, and it stayed open for four decades because human tutors do not scale.
Kestin and colleagues (2025) ran the first strong test of whether software closes part of that gap. In a randomized trial with 194 students in an introductory physics course at Harvard, the researchers set an AI tutor designed around pedagogical best practices against an in-class active-learning session on identical material. Median learning gains in the AI group were more than double, students spent less time on the material, and they reported more engagement and motivation. The comparison group was already using active learning, which itself outperforms lecturing.
One trial with 194 students does not settle the 2 sigma problem. It does show that one learner working with software can beat a well-run class on the same material.
The one way AI makes learning worse
AI makes learning worse when it supplies the answer during practice. Bastani and colleagues (2025) ran a field experiment with nearly a thousand high-school math students in Turkey across the 2023–24 school year. Students given a ChatGPT-like interface, called GPT Base, scored 48% higher on the practice problems they solved with it in front of them. On a later exam taken without AI, the same students scored 17% below classmates who had never used it. They had copied answers, and they did not perceive the harm while it was happening.
The same paper contains the correction. A second version, called GPT Tutor, gave teacher-designed hints instead of solutions. It raised practice scores 127%, and exam scores in that group were statistically indistinguishable from the control group. One change to the instructions separates a 17% loss from no loss at all.
That is the first thing to check before you learn with AI: whether the tool lets you skip the retrieval. A chat window that answers on demand removes the retrieval step, and it removes it at exactly the moment the effort would have paid.
A day of learning that follows the evidence
Forty-five minutes across two mornings carries one concept from placement to a card in a review queue. Each step measures something different.
Five minutes, placement. Answer three broad questions about the field, then a few targeted ones if real gaps remain. This measures what you already know, so the lesson can start from there and skip what you can already state without help.
Twenty minutes, one lesson. Read a single self-contained page on one concept, with a diagram in each part and one worked example carried through the whole page. This is the only step that hands you material, and it stays short on purpose.
Five minutes, quiz. Answer questions on the concept the same day you read it, with the lesson closed. This measures retrieval from memory rather than recognition on the page, and it is the practice testing that the 2013 review rated highest.
The next morning, five minutes, explain-back. Write the concept in your own words and have it graded against the lesson. Sleep in between matters: declarative memories are reactivated and stabilized during slow-wave sleep (Diekelmann and Born, 2010). This step is the Feynman technique run with AI, and a concept that fails it is taught again and tested again the next day.
Ten minutes, cards. Review the cards that are due, on a schedule that widens each time you recall an item correctly. This is distributed practice, and managing that schedule is the whole job of spaced repetition with AI. Retention is then read from the review history, from intervals and lapses, rather than from how confident you feel.
What to learn next, and in what order
The next thing to learn is the smallest step from what you already hold. Prerequisites set the order, and skipping one costs more time than learning it would have taken. A lesson on gradient descent lands only if you hold partial derivatives already. A lesson on database indexes lands only if you know what a full table scan costs.
Placement questions exist for this reason. Without them a tutor either repeats what you know, which wastes the session, or starts above you, which produces fluent reading and no memory.
For a technical field, order the path by dependency rather than by syllabus. Write down the concept you actually want. Ask what that concept needs. Ask the same question of each answer until you reach something you can state without help, then learn the path back up in that order. Interleave the review of finished concepts so each day of cards spans several of them.
The Learning Harness is my implementation of this routine. It asks what you know before it teaches, keeps every topic on one ordered path, writes each lesson for one learner, quizzes you the same day, grades an explain-back the next day after sleep, and writes cards into Anki only for concepts that passed both. You can read how the harness teaches in full.
Common questions about the best way to learn in 2026
What is the best way to learn in 2026?
Test yourself on the material and spread the sessions across days. Dunlosky and colleagues (2013) reviewed ten study techniques and rated only two as high utility: practice testing and distributed practice. In practice that means one short lesson, a quiz the same day without the page open, an explain-back the next morning after sleep, then cards on a widening schedule.
What is the best way to study in 2026?
Study by retrieving what you read rather than by looking at it again. Close the page, write what you remember, then check it against the source. Roediger and Karpicke (2006) found that restudying won on a test five minutes later and lost on tests two days and one week later. Put the second pass on a different day.
Is it better to learn with AI or with a course?
The one that tests you is the better choice. Kestin and colleagues (2025) found median learning gains more than doubled with an AI tutor designed around pedagogical best practices, measured against an active-learning physics class. Bastani and colleagues (2025) found that an AI which supplied answers left exam scores 17% below students with no AI. The design of the tool decides the outcome.
What is the best way to learn coding in 2026?
Write the code from memory before you read a solution, and space the attempts across days. Coding in 2026 has a specific failure mode: the assistant completes the function, the tests pass, and you retrieved nothing. Read the concept, close the tab, implement it, then compare. Keep every error you hit as a review item for later.
How long should a learning session be?
Twenty to forty minutes on one concept, then stop. Spacing matters more than length: two thirty-minute sessions on separate days beat one hour in a single sitting, because spaced study beats massed study across retention intervals (Cepeda and colleagues, 2006). End every session with a five-minute quiz and take a five-minute explain-back the next morning.
Sources
- Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving Students' Learning With Effective Learning Techniques. Psychological Science in the Public Interest, 14(1), 4–58. https://doi.org/10.1177/1529100612453266
- Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning. Scientific Reports, 15, 17458. https://doi.org/10.1038/s41598-025-97652-6
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning. Proceedings of the National Academy of Sciences. https://doi.org/10.1073/pnas.2422633122
- Roediger, H. L., & Karpicke, J. D. (2006). Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x
- Karpicke, J. D., & Blunt, J. R. (2011). Retrieval Practice Produces More Learning than Elaborative Studying with Concept Mapping. Science, 331(6018), 772–775. https://doi.org/10.1126/science.1199327
- Bloom, B. S. (1984). The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring. Educational Researcher, 13(6), 4–16. https://doi.org/10.3102/0013189X013006004
- Diekelmann, S., & Born, J. (2010). The memory function of sleep. Nature Reviews Neuroscience, 11, 114–126. https://doi.org/10.1038/nrn2762
- 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://doi.org/10.1037/0033-2909.132.3.354
The Learning Harness is in a closed beta and not yet available.