AI math tools can make practice easier and help students solve more problems while the tool is available. That is not the same as learning a method well enough to use it alone. The key difference is instructional design: a tool can coach with questions and hints, or it can give away the solution. Evidence so far suggests that distinction matters, while also falling short of proving that any one design reliably improves long-term achievement.
What changes when AI helps with math?
A conventional answer key shows a result; an AI assistant can respond to a student’s particular question, explain a step, or offer another example. That immediacy can make practice more accessible. But if a student copies a worked solution instead of reasoning through the problem, strong performance during assisted practice may not carry over to an exam without help.
As an Amazon Associate I earn from qualifying purchases.
It helps to separate three outcomes: solving a problem with AI available, answering a related question after the AI is gone, and retaining or applying the skill later. A correct answer in the first situation is evidence of assisted performance, not by itself evidence of independent learning or durable understanding.
What a controlled high-school study found
Bastani and coauthors studied nearly 1,000 students in grades 9–11 at one large high school in Turkey during the 2023–2024 school year. Classes were assigned to standard GPT-4 chat (“GPT Base”), a guarded version of the same model (“GPT Tutor”), or no AI access during four 90-minute math review sessions. The researchers compared assisted practice with a later exam taken without resources. The study was published in PNAS in 2025.
#1 Best Overall
- Full of different activities to help your child develop their skills
- Contains one sixty-four page workbook
- Available in a variety of different age groups
- Available in different themed activity books
- Made in USA
| Study condition | Assisted practice versus control | Later unassisted exam versus control |
|---|---|---|
| GPT Base, standard GPT-4 chat | 48% improvement | 17% lower score |
| GPT Tutor, teacher-informed guardrails | 127% improvement | Statistically indistinguishable from control; no significant improvement was demonstrated |
These percentages describe the study’s reported relative comparisons, not percentage-point increases in exam grades. The pattern is the important point: both AI groups performed better during assisted practice, but only the unguarded chat group had a measured decline on the later unassisted exam. The guarded tutor avoided that decline; it did not show a statistically significant achievement gain over students who had no AI access.
The authors reported that GPT Base students often copied solutions, while GPT Tutor students more often asked for help or tried answers themselves. The tutor was instructed to offer hints rather than the answer and was supplied with teacher-prepared correct solutions and common student mistakes for the specific problems. That design involved substantial, problem-specific teacher input, so it should not be mistaken for a setting that can automatically be applied to every math topic.
Bastani and coauthors summarized the result this way: “Our results suggest that while access to generative AI can improve performance, it can substantially inhibit learning without appropriate guardrails.” The experiment does not establish that AI harms math learning in general: it involved one school, one model setup, a particular curriculum, and four review sessions.
How coaching differs from answer-giving
A coaching-oriented tutor makes the student do some of the mathematical work. Instead of immediately completing a problem, it might ask what the student has tried, point to the next step, or identify a likely misconception. Answer-giving can be useful for checking work or seeing a worked example, but it also makes it easier to bypass the effort that independent recall and problem-solving require.
- Hints before solutions: The student gets a chance to attempt the next step before seeing a full method.
- Attention to prior work: Explanations can address an actual error or prerequisite gap rather than repeat a generic lesson.
- Accuracy checks: The student should be able to verify calculations and reasoning; fluent explanations are not proof that every step is correct.
- Unaided follow-up: After help, the learner tries a similar problem without AI and explains the reasoning.
These are design criteria, not proof that a tool meeting them will raise achievement. A tutor can still make mistakes, misread a student’s work, or give hints that effectively reveal the answer. Teacher or parent oversight remains relevant, particularly when students are young or using AI for graded work.
What Khan Academy’s product experiments add
Khan Academy describes Khanmigo product tests conducted from October 2025 through April 2026. The organization says the tests covered more than 15 million tutoring threads in total. Its reported outcome included whether a learner answered the next problem in the same skill correctly without AI assistance; it also examined response latency and signs of active or constructive engagement. Khan Academy’s account of the tests reports a 3.4% improvement in next-item correctness when the tutor received a summary of recent problem-solving history, a 2.7% improvement when it identified unmet prerequisites and offered a brief review, and a 6.1% combined improvement from those two changes.
Rank #4
Those are Khan Academy’s internal product metrics, not independent findings of test-score improvement or lasting course achievement. They indicate that context about recent work and prerequisite skills can improve a near-term measure within the product; they do not show that the same changes work equally well in other tools, curricula, or populations.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteWhat the current evidence does not settle
A 2024 randomized PLOS ONE study assigned 274 participants to ChatGPT-generated help, human tutor-authored help, or no hint across elementary algebra, intermediate algebra, college algebra, and statistics, using questions from OpenStax math textbooks and the Open Adaptive Tutor platform. The available source record describes this comparison, but it does not provide enough outcome detail to establish whether AI hints matched or exceeded human-authored hints. Its title alone is not a substitute for verified results. The study record is available at PLOS ONE.
Best Value
- Carefully Crafted Queries: Engaging and relevant math questions
- Diverse Fun Activities: A mix of enjoyable exercises
- Problem-Solving Techniques: Step-by-step strategies
- Vivid Color Illustrations: Bright, full-color visuals
An NBER page for a 2026 working paper describes a two-year school experiment involving Khanmigo in remedial math sessions and reports an achievement effect. The full paper’s methods, sample, estimates, and interpretation are not established here, so that result cannot responsibly be quantified or used to draw a conclusion about the size or reliability of the effect. The working-paper page is available from NBER.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to judge an AI math tool for learning
Rather than choosing by how polished an explanation sounds, ask whether the tool helps a learner build a skill that remains available without it. A useful comparison should cover how the system behaves, what it can verify, and what evidence exists for its outcomes.
Quick Recap
- Does it coach or complete? Check whether it prompts an attempt and offers staged hints, or readily supplies full solutions.
- Does it adapt to demonstrated skill? Look for use of the learner’s recent attempts and prerequisite gaps, not just personalization by age or topic.
- Can its math be checked? See whether it shows enough reasoning to inspect and acknowledges uncertainty instead of presenting every answer as certain.
- Is independent practice built in? A learner should solve a fresh, similar problem without assistance after receiving help.
- Are privacy and oversight suitable? Consider the student’s age, the data collected, teacher or parent controls, and whether use fits classroom rules.
- Does it fit the actual course? Curriculum alignment, accessibility, and cost matter alongside the tutor’s conversational quality.
- What kind of evidence supports it? Separate independent controlled studies from a company’s usage metrics, and distinguish immediate correctness from later unaided performance.
A practical way to use AI without outsourcing the learning
- Try first. Write down an approach or identify the step where you are stuck before asking for help.
- Request a hint, not a finished solution. Ask for one next step or a question that helps diagnose the mistake.
- Check the explanation. Compare the reasoning with class notes, a textbook, or a teacher’s method, especially if the answer seems surprising.
- Close the help and retry. Solve the original problem again, then attempt a new problem of the same type without assistance.
- Notice what still fails. If the method does not transfer, revisit the underlying concept or ask a teacher to diagnose the gap rather than counting the AI-assisted answer as mastery.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
