The criteria

What Makes an AI Math Tutor Good

Seven standards that separate a tutor that teaches from one that hands out answers. Each comes with what good looks like, what bad looks like, and a check you can run yourself.

Last reviewed November 10, 2026.

1. Socratic questioning, not answer-giving

The standard: It asks the next useful question and lets the student do the thinking, even when the student asks for the answer outright.

Why it matters: Students given a standard chatbot during practice did better while they had it and 17% worse on exams afterwards; a tutor built to guide instead of tell largely avoided that harm.[1] Withholding answers is the single most important thing to test.

Looks like

“What do you notice about the two denominators?” One question, aimed at the student's own step.

Doesn’t look like

“Here's how to solve it: Step 1…” or a “show solution” button that is always available.

Check it: Ask it for the answer three times, the third time with a reason ("my teacher said it's fine"). It should still not hand it over.

What each tool states
  • Glimmer Stated Glimmer states that Glo “never gives away the answer, and never has a student blindly follow a procedure,” using Socratic questioning strategies.[2]
  • Khanmigo Partly Its learner page says it “never gives you the answer,” but Khan Academy wrote in October 2026 that the rule was “too rigid”: after an incorrect attempt it can now become more direct.[3,4]
  • MATHia Partly Just-in-time feedback and contextual hints; no stated policy on withholding answers.[5]
  • Eedi Stated “It never gives answers, it only asks students questions,” and activates after a specific error.[6]
  • Smartschool Stated Uses guided questioning and hints that “build thinking, not just reveal answers.”[7]
  • DreamBox Math Partly Immediate feedback that guides strategic thinking, with scaffolded lessons; no stated answer-withholding policy.[8]
  • Mathspace Partly Hints, videos, worked examples and step-level feedback; no stated answer-withholding policy.[9,10]
  • Frizzle No A teacher-side grading tool, not a student-facing tutor.[11]
  • Ed.ai No Grades and explains where logic breaks down; teachers review every comment first. No student-facing tutor stated.[12]

2. Reads written work

The standard: It looks at the student's steps (handwritten, photographed or typed line by line), not only the final answer or a multiple-choice pick.

Why it matters: Feedback works when it addresses how the student did the task, not just the result.[13] A tool that sees only the final answer can’t tell a careless slip from a misconception, or catch a right answer reached the wrong way.

Looks like

It quotes or points to the line where the work went wrong.

Doesn’t look like

It marks the final answer right or wrong and offers a generic hint.

Check it: Submit a solution with the right answer reached by a wrong method. A tool that only checks answers will call it correct.

What each tool states
  • Glimmer Stated Identifies misconceptions from students' written work. Ask in a demo which formats your students will use (photo, stylus, typed).[2]
  • Khanmigo Partly Students type questions; it follows students' steps on Khan Academy exercises. Photo or handwriting input is not stated.[14,15]
  • MATHia Partly Students work multi-step problems inside the software. Handwriting or photo input is not stated.[5]
  • Eedi No Students answer multiple-choice diagnostic questions (or hold up A/B/C/D cards in class). Handwriting input is not stated.[16,17]
  • Smartschool Stated Students type solutions or upload photos of handwritten work; it assesses step-by-step work.[7]
  • DreamBox Math Partly Adapts by assessing both strategies and answers, through virtual manipulatives. Handwriting input is not stated.[22]
  • Mathspace Stated Students enter each step; reports keep a record of every step of a student's working (typed, not photographed).[28]
  • Frizzle Stated Reads every step of handwritten paper work from a phone, document camera or scanner, and flags pages it is unsure of for a person.[11]
  • Ed.ai Stated Reads handwritten work, follows multi-step reasoning and awards partial credit.[12]

3. Misconception-level feedback

The standard: Its feedback names the specific wrong idea behind the error ("you added the denominators") rather than "try again" or a generic hint.

Why it matters: Feedback that only says “incorrect” has little effect; feedback about the specific error and how to fix it does far more.[13] Studying an error with its cause explained helps students learn.[18]

Looks like

“You added the denominators. What does the bottom number of a fraction tell you?”

Doesn’t look like

“Try again!” or “Remember to find a common denominator” with no reference to what the student did.

Check it: Submit a classic error with a known cause. The feedback should name that cause, in words a student can act on.

What each tool states
  • Glimmer Stated Pinpoints the misconception in a student's work and gives teachers misconception-level insights.[2]
  • Khanmigo Partly Khan Academy says it is better at catching and pointing out mistakes; misconception diagnosis is not described as a feature.[15]
  • MATHia Partly Adapts skill by skill; naming the misconception to students is not described in those words.[5]
  • Eedi Stated The core of the product: every distractor is written by learning designers, and teachers see misconceptions broken down by wrong answer.[16,17]
  • Smartschool Partly Identifies learning gaps as they arise; naming misconceptions is not described in those words.[7]
  • DreamBox Math Partly Continuous formative assessment; naming the misconception is not described in those words.[8]
  • Mathspace Partly Diagnostic check-ins with AI-powered recommendations; naming the misconception is not described in those words.[9]
  • Frizzle Stated Tags named misconceptions, mapped to standards.[11]
  • Ed.ai Stated Error pattern detection; practice generated from each student's misconceptions.[12]

4. Prerequisite gap filling

The standard: When a grade-level error comes from an earlier gap, it drops back to the prerequisite skill, then returns to the grade-level task.

Why it matters: Students who struggle often lack earlier skills, and intervention guidance calls for systematic instruction that builds understanding step by step.[19]

Looks like

It spots that a negative-number error is blocking an equation, practices integers briefly, then returns to the equation.

Doesn’t look like

It repeats the same grade-level problem with a different number.

Check it: Give a 7th-grade equation that fails only because of a 4th-grade fraction error. It should notice the fraction gap.

What each tool states
  • Glimmer Stated Fills in prerequisite gaps and adds scaffolds for students who need them.[2]
  • Khanmigo Partly Paired with Khan Academy's mastery-based content designed to fill knowledge gaps.[15]
  • MATHia Partly Adjusts to every action a student takes to meet them where they are.[5]
  • Eedi Partly Predicts and helps resolve misconceptions, and suggests a lesson when a student is struggling.[6,20]
  • Smartschool Stated Immediate remediation, including reteaching last year's content.[21]
  • DreamBox Math Stated Intelligent adaptive learning that differentiates within and between lessons to fill gaps.[8,22]
  • Mathspace Partly Step-level adaptivity; adaptive tasks adjust difficulty in real time.[9,28]
  • Frizzle Partly Prerequisite tracing (“knows when a 7th-grade error is really a 4th-grade gap”), for the teacher to act on.[11]
  • Ed.ai Partly Personalized remediation activities built from each student's mistakes, editable by the teacher.[12]

5. Curriculum fit

The standard: It works from the materials, vocabulary and sequence your classes already use, instead of pulling students into a parallel program.

Why it matters: If the tutor uses its own sequence and vocabulary, students get two versions of math and teachers lose time translating. A tutor that works from your materials reinforces today’s lesson.

Looks like

It serves problems from your lesson and uses your curriculum's terms and models.

Doesn’t look like

Students leave your unit to work through the vendor's own scope and sequence.

Check it: Load tomorrow's lesson or problem set. Does the tutor use your materials and your terms?

What each tool states
  • Glimmer Stated Works with any curriculum, serving questions directly from your materials; aligned to high-quality instructional materials.[2]
  • Khanmigo Partly Standards-aligned, tied to Khan Academy's content library rather than your own materials.[14]
  • MATHia Partly Part of Carnegie Learning's middle and high school solutions; also sold as a supplement with a standards report.[5]
  • Eedi Partly Covers multiple curricula, with a curriculum mapping page.[20]
  • Smartschool Stated Assignments aligned with your standards and curriculum, using its terminology and scope.[7,21]
  • DreamBox Math Partly Its own lessons, aligned to every US state's standards; teachers can assign by grade, topic or standard.[8,22]
  • Mathspace Partly Its own standards-aligned textbooks and core curriculum, which can replace rather than follow yours.[9]
  • Frizzle Stated Curriculum-agnostic; aligns to CCSS, TEKS and 30+ state frameworks.[11,29]
  • Ed.ai Stated Works with your existing assessments; integrates with common learning management systems.[12]

6. Language support

The standard: Multilingual learners can get help in their home language while still building academic math language in English.

Why it matters: Guidance for teaching English learners stresses teaching academic vocabulary and building language alongside content.[23] Home-language help keeps students working while they learn math terms in English.

Looks like

It explains in Spanish but keeps the key math term in English beside it (“el denominador, denominator”).

Doesn’t look like

English only, or machine translation that mistranslates math terms.

Check it: Ask a question in Spanish (or your students' most common home language) and check the math vocabulary it uses back.

What each tool states
  • Glimmer Stated Personalized support in 100+ languages.[2]
  • Khanmigo Not stated Not stated on the pages we could read. Check khanmigo.ai.[14]
  • MATHia Not stated Not stated on the MATHia page. Ask Carnegie Learning.[5]
  • Eedi Not stated Not stated. Ask Eedi.[16]
  • Smartschool Stated Support for learners in 200+ languages.[7]
  • DreamBox Math Partly Teacher tools available in Spanish; Spanish for US students is not confirmed on the pages we read.[24]
  • Mathspace Not stated Not stated. Ask Mathspace.[9]
  • Frizzle Not stated Not stated. Ask Frizzle.[11]
  • Ed.ai Not stated Not stated. Ask Ed.ai.[12]

7. Teacher reporting you can use tomorrow

The standard: It tells you which misconceptions your class holds, who shares them and what to do next: groups, reteach points, check-ins.

Why it matters: Formative assessment improves learning when teachers use the evidence to adjust teaching.[25] Reports that show only minutes and percent correct don’t tell you what to reteach.

Looks like

“9 students added denominators; here is a suggested small group and a reteach question.”

Doesn’t look like

A leaderboard, time on task and a percent-correct column.

Check it: After your test session, open the teacher view. Could you plan tomorrow's small groups from it in five minutes?

What each tool states
  • Glimmer Stated Real-time misconceptions, whole-class strategies, suggested groupings and one-on-one check-ins, plus building and district reporting on standards and benchmark readiness.[2]
  • Khanmigo Partly An on-demand summary of recent student work for teachers.[14]
  • MATHia Stated Session, skills, student and standards reports, a year-end prediction report, a leadership report and LiveLab real-time alerts.[5]
  • Eedi Partly Whole-class data and misconception breakdowns from each quiz.[17]
  • Smartschool Partly Real-time data and student progress insights; admin accounts on school plans.[7,21]
  • DreamBox Math Stated Real-time dashboard of progress against standards at class, school and district level.[22]
  • Mathspace Stated Task, student, class, skills, mastery and live reports, plus an administrator dashboard.[28]
  • Frizzle Stated Live class dashboards; school and district standards dashboards on institution plans.[11,29]
  • Ed.ai Stated Student reports mapped to standards, a class mastery view and growth over the year.[12]

Plus: evidence of impact

The seven criteria describe how a tutor teaches. Evidence of impact tells you whether that teaching worked somewhere else. Ask for a study you can read, and check it against the ESSA evidence tiers: strong, moderate, promising, or demonstrates a rationale.[26] Human tutoring has strong experimental evidence across many studies.[27] For AI tools, design matters: the same model helped or hurt depending on whether it gave answers.[1]

Many newer tools haven’t published outcome studies. That doesn’t rule them out; it means you should run a pilot with a pre- and post-measure. Questions to ask about evidence.

How we read vendor claims

  • Stated The vendor’s own pages say the product does this, in words that match the standard.
  • Partly The pages describe something related but not the full standard (for example, “hints” without saying answers are withheld).
  • Not stated We couldn’t find it. Ask the vendor; it may exist.
  • No The product is designed differently (for example, a grading tool with no student-facing tutor).

A vendor saying it doesn’t prove it. That’s why every criterion has a check: run the 20-minute test before you commit.

Sources

Every claim above links to one of these. Vendor pages were read on November 10, 2026; check them again before you buy, since products change.

  1. Generative AI without guardrails can harm learning: Evidence from high school mathematics (Bastani et al., 2025). Proceedings of the National Academy of Sciences.
  2. Glimmer: real-time AI math coaching (Glo). learnwithglimmer.com.
  3. Khanmigo for learners. Khan Academy.
  4. Four things to look for in edtech, and how Khan Academy measures up. Khan Academy Blog.
  5. MATHia: the 1-to-1 math coach. Carnegie Learning.
  6. Eedi Tutor. Eedi.
  7. Smartschool home page. Smartschool.
  8. DreamBox Math. Discovery Education.
  9. Mathspace home page. Mathspace.
  10. Mathspace for learners. Mathspace.
  11. Frizzle home page. Frizzle.
  12. Ed.ai home page. Ed.ai.
  13. The Power of Feedback (Hattie & Timperley, 2007). Review of Educational Research.
  14. Khanmigo home page. Khan Academy.
  15. Khanmigo math computation and tutoring updates. Khan Academy Blog.
  16. Eedi home page. Eedi.
  17. DQR whole-class diagnostic quizzes. Eedi.
  18. Using example problems to improve student learning in algebra: Differentiating between correct and incorrect examples (Booth et al., 2013). Learning and Instruction.
  19. Assisting Students Struggling with Mathematics: Intervention in the Elementary Grades (2021). What Works Clearinghouse, Institute of Education Sciences.
  20. Eedi for schools. Eedi.
  21. Smartschool for schools and districts. Smartschool.
  22. DreamBox Math standards alignment. Discovery Education.
  23. Teaching Academic Content and Literacy to English Learners in Elementary and Middle School. What Works Clearinghouse, Institute of Education Sciences.
  24. DreamBox Math teacher tools. Discovery Education.
  25. Assessment and Classroom Learning (Black & Wiliam, 1998). Assessment in Education: Principles, Policy & Practice.
  26. Evidence for ESSA: programs that meet ESSA evidence standards. Center for Research and Reform in Education, Johns Hopkins University.
  27. The Impressive Effects of Tutoring on PreK-12 Learning: A Systematic Review and Meta-Analysis of the Experimental Evidence (Nickow, Oreopoulos & Quan, 2020). National Bureau of Economic Research, Working Paper 27476.