AI math grader for handwritten work

Photograph a finished math page and it comes back marked, with the step that went wrong named beside each mistake. It works every problem itself rather than matching a key, so ½, 0.5 and 50% all count, and a right answer reached by an unexpected route is still right. Pencil, cramped working, answers in the margin. What it cannot read, it flags rather than guesses.

Grade a math page free Every subject, not just math

It works the problem, it does not match a key

Before it judges a single answer, it works the problem out for itself. That is the whole difference between this and an answer key, and it matters most on exactly the pages that take longest to mark by hand: a page of algebra where the student lost a sign in the second line and recovered by accident, or a word problem answered with the right number and the wrong reasoning.

A key says right. A teacher says otherwise. The grader agrees with the teacher: it can tell you the final answer is correct and line three is still wrong.

If you give it an answer key it grades to your key, accepting the same equivalent forms, and it still works out any problem your key does not cover.

It marks the step that went wrong

A red mark on its own teaches nothing. For every wrong answer the grader traces the student’s answer back to the slip that would produce it and names that slip in one sentence: solved for x when the question asked for 2x, counted every number shown instead of only the different ones, used the area of the whole square instead of the triangle. If the answer does not trace back to a single slip, it says which step the working breaks down at.

Under the mark, the student gets a short worked solution: two to four lines, each carrying its own numbers, ending at the right answer. It appears only when a problem did not earn full credit. A student who got it right already knows how.

Equivalent forms it accepts

Because it works the problem rather than string-matching a key, any mathematically equivalent form counts:

Unsimplified fractions: 2/4 counts as 1/2
Improper fractions, mixed numbers and decimals: 3/2, 1½ and 1.5 all count
Fractions, decimals and percents: 0.5, 1/2 and 50% all count
Leading and trailing zeros: .5, 0.5 and 0.50
Reordered terms in a sum or product: 3 + 4 and 4 + 3
Ratios either way: 3:4 or "3 to 4"
Units written or left off: 140 seconds, 140 s, or 140
Exact forms: a radical or a multiple of π left as written, degree signs present or absent
Reasonable rounding

A multi-part problem is different. “Simplify and state the restriction” is not correct until both parts are there: a student who simplifies correctly and never writes the restriction is marked partially right, with the missing part named. A plain single-answer computation stays all or nothing unless your own rubric says otherwise.

What it reads

A phone photo of the finished page. Pencil or pen, working squeezed into the margin, an answer circled somewhere other than the answer line, a problem that starts on one page and finishes on the next. There is no template to print and no answer box a student has to write inside.

A blank problem is marked blank, never guessed at, and a lone question mark or “idk” is read as a blank too. A page with a diagram, a chart or a construction beside the working is not turned away: the working is read and marked as usual. What happens to the figure itself is in the limits below.

What it does when it cannot read something

It says so. The answer is flagged as unread and handed back to you with its best reading, so confirming or correcting it is one tap.

This is the behaviour worth testing on any math grader, including this one. The usual failure is not arithmetic, it is transcription: a faint minus sign, a decimal point read as a smudge, a 7 read as a 1. A tool that guesses marks a right answer wrong and nothing looks broken, so you never find out. The flag is rare on a clearly photographed page and it is meant to be, because a grader that flags every untidy digit is one you stop reading.

How this differs from our all-subjects page

AI grading for handwritten work covers every subject the grader reads: math, science, English, world languages and social studies, plus rubric-scored essays. This page is the math case only, and it is the case where working the problem, naming the wrong step and accepting equivalent forms do the most. If your students hand in a mix of subjects, start there. If they hand in math, start here.

Where this is the wrong tool, and what it will not do

It does not score figures. A freehand graph, a diagram or a geometry construction is read far less reliably than working and text. The page is accepted and the working beside the figure is marked; the figure itself you check yourself.

There is no published accuracy number for math. There is no equivalent published benchmark for math yet, and the page says so rather than implying one. The measured figures we publish are for essays: On 78 real student essays that Texas released with official examiner scores, nine in ten came back within one point of the examiner and quadratic weighted kappa was 0.78, against 0.70 as the commonly accepted bar for an automated marking system. Treat any math accuracy percentage, ours or anyone’s, that has no released scored set behind it as marketing.

A written proof goes to you. Anything open-ended with many valid answers is handed back for your judgement rather than forced into right or wrong, unless you give it a rubric to score against.

Fixed-template exams at cohort scale. If every student answers in the same printed box and you are grading hundreds, a scanning platform built for that will beat us.

District procurement with a SOC 2 requirement. We do not have that certification today.

Test it on five pages before you trust it

Take five math pages you have already graded yourself, including your two messiest, and run them through. You are not checking whether the totals match. You are checking three things: did it read the handwriting, did it name the same slip you would have named, and when it was unsure did it say so or did it bluff.

Do that before a class set, not after, and do it to every tool you are weighing up.

Common questions

Can AI grade handwritten math from a photo?

Yes. Photograph the finished page with a phone and each problem comes back marked on the page itself. It reads pencil and pen, working crammed into the margin, and an answer written somewhere other than the answer line. What it cannot read, it flags for you rather than guessing at.

Does it need an answer key?

No. It works each problem out itself before judging, which is what lets it accept an equivalent form and mark a student right who arrived by an unexpected route. If you do give it a key, it grades to the key and still works out any problem the key does not cover.

Does it mark the step that went wrong, or just the final answer?

The step. For every wrong answer it traces the student’s answer back to the slip that would produce it and names that slip in one sentence: solved for x when the question asked for 2x, used the area of the whole square instead of the triangle. Under the mark the student gets a short worked solution, two to four lines, only when they did not get full marks.

Which equivalent answers does it accept?

Any mathematically equivalent form: unsimplified fractions, mixed numbers and decimals, fractions against percents, leading and trailing zeros, reordered terms, ratios written either way, units written or left off, exact radical and π forms, and reasonable rounding.

Does it give partial credit?

On a multi-part problem, yes: if a student simplifies correctly but never states the restriction, or finds x but not the angles, the mark records the fraction of parts earned and says which part was missed. A plain single-answer computation stays all or nothing unless your own rubric says otherwise.

What about diagrams, graphs and geometry constructions?

A page with a figure on it is not turned away: the working beside the figure is read and marked as usual. The figure itself is not scored. A freehand graph or construction is read far less reliably than working and text, so check those yourself.

How accurate is it on math?

There is no equivalent published benchmark for math yet, and the page says so rather than implying one. On 78 real student essays that Texas released with official examiner scores, nine in ten came back within one point of the examiner and quadratic weighted kappa was 0.78, against 0.70 as the commonly accepted bar for an automated marking system. The method and its limits are published, and the same rule applies to math: every mark is yours to change before a student sees it.

Can I change a mark?

Any mark. Set an answer right, wrong or blank, edit the reason, adjust the credit, and nothing reaches a student until you release it. It grades; the grade stays yours.