Pick the grader by what is on the page. Gradescope and GradingPal read short answers and math, EssayGrader and Marking.ai read handwritten essays, and Top Marks AI and Running Paper read exam-style scripts marked against a UK mark scheme (the UK term for a rubric plus answer key). Research finds that AI reads neat numbers well and struggles with drawings, so grade a sample by hand before you trust any of them with a class set.
Why handwritten work is back on teachers' desks
Teachers who want to know what a student can do without a chatbot have been moving the evidence onto paper. The Chronicle of Higher Education reported on September 4, 2026 that exam booklet sales at Follett's stores are "up nearly 40 percent since the fall of 2023" as educators turn to in-person testing to combat AI cheating (The Chronicle). That is a college story, but a high school teacher with 150 students and a stack of handwritten quizzes has the same problem on a smaller floor: the thinking is on paper and the grading is still done by hand. Our guide to AI-resistant assessment tools covers the assessment side. This post covers the pile.
How this list was built
The AI Educator Tools directory lists 138 tools, and most graders sit in its assessment and grading category. Fourteen of the 138 mention handwriting somewhere in their listing, and twelve of those are marking tools (the other two are a worksheet cleaner and a writing assistant). Eight of the twelve state UK or EU data hosting, two state US hosting (EssayGrader and GradingPal), and two do not say (Gradescope and AI Mentor by Novalearn). Only one of the twelve, Gradescope, carries an independent 9ine security assessment.
I chose six on three tests: the vendor's own pages say it reads handwritten work, the pages give enough on pricing and data to compare, and each fits a different kind of paper. Nobody paid for a place. I left out EduTestPro because the terms and privacy notice on its website were marked "not yet adopted" and "not yet recorded" when I read them on November 10, 2026, and I left out AI Mentor by Novalearn because I could not check its handwriting claims at the vendor's site. Everything below was checked on vendor pages on November 10, 2026. Where a page was silent or unreadable, the text says so.
What the research says AI can and cannot read
Three recent studies tested multimodal AI models, the kind that accept a photo, on real handwritten student work.
- Arithmetic and short answers, strong. Henkel and colleagues graded 288 handwritten arithmetic responses from Ghanaian middle school students and report "near-human accuracy (95%, k = 0.90)," with "occasional errors that human educators would be unlikely to make" (arXiv, October 2025).
- Drawings, weak. In the same paper, 150 illustrations from American elementary students scored only k = 0.20 against human scores when the model graded the picture directly, and k = 0.47 when a human described the picture first (arXiv). The authors say the models "struggle to 'see'" student illustrations.
- Mixed formulas and diagrams, not yet dependable for high stakes. The EDU-CIRCUIT-HW benchmark, built from more than 1,300 handwritten university STEM solutions, concludes that current models show "insufficient reliability for auto-grading" in high-stakes settings (arXiv, accepted to Findings of ACL 2026).
- Prompting matters. A study of one handwritten engineering quiz, graded with GPT-5.2 and Gemini 3 Pro, estimates that about 17 percent of papers would need manual review, and found that removing the reference solution introduced "systematic over-grading" (arXiv, January 2026). It covers a single quiz, so treat it as a warning, not a rate.
The pattern is useful. Neat numbers and short text are in range. Diagrams, graphs and sketches are where you should keep your own pen. A vendor's accuracy figure is always a figure about one kind of paper, and the next section shows how differently each vendor measures it.
Comparison table: six AI graders that read handwriting
| Tool | Best use | Strengths | Limitations | Price / free tier (checked Nov 10, 2026) | Privacy considerations |
|---|---|---|---|---|---|
| Gradescope | Paper exams graded question by question | Groups similar answers; rubric edits apply retroactively | AI-assisted grading is Institutional-only; no public Institutional price | Basic free; Institutional by quote | Vendor says data stays in a region your institution chooses; privacy policy last updated October 2022 |
| GradingPal | Worksheets, quizzes and mixed K-12 assignments | Handwritten work included on the free plan | Terms warn of "incorrect output"; no measured accuracy published | Free: 50 submissions a month; Lite $14; Pro $26 per teacher per month | Vendor says student work is not used to train AI; US storage by default |
| EssayGrader | Handwritten essays against a rubric | Class set scanned as one PDF; syncs to Google Classroom and Canvas | Its pages disagree on the free-tier cap; adults only (18+) | Free tier; Pro $14.99, Premium $34.99 per month | Homepage says student work is never used to train AI; data stored in the US |
| Marking.ai | Handwritten work marked against your uploaded mark scheme | Quotes the words that earned each mark; teacher approves every mark | Scanned scripts end to end are listed as "Not started" in its accuracy studies | 20 free submissions, one time; from £21 a month | Student work stored in Frankfurt; AI processing in the US |
| Top Marks AI | Batch marking of exam scripts in the humanities | Large batches; per-subject tools aligned to UK boards | Pricing by quote; training on student work not addressed in its policy | Contact for school pricing | School data hosted in the UK and EU; anonymized content goes to US AI providers |
| Running Paper | Pen-and-paper assessments, ages 11 to 18 (UK Key Stages 3 to 5) | Two AI models mark each question; disagreements go to the teacher | Accuracy results are vendor case studies; no per-pupil price published | Trial up to 3 months for £500; per-pupil pricing by quote | Says neither it nor its sub-processors use student work to train AI |
The privacy column records what each vendor says, not an audit. Your district's own data agreement is what counts.
Gradescope: the best fit for paper exams graded question by question
Gradescope is the best fit when a department gives the same paper exam to many students and wants to grade one question across all of them at once. Students' work is scanned or photographed, and the homepage says "For some question types, Gradescope AI automatically forms groups for you to review." You grade a group of similar answers in one click instead of reading each paper top to bottom (Gradescope).
It differs from the rest of this list in what it reads. I found no wording about reading handwritten text for essays. The privacy page mentions text recognition performed on handwritten formulas, which points to math as its strength (privacy policy). Choose something else for handwritten essays.
Watch out for: AI-assisted grading and answer groups are "part of our Institutional license," so the free Basic plan gives you the scanning and rubric workflow without the AI features (pricing).
Pricing: Basic is free. Institutional has no published price, only a contact form (checked November 10, 2026, pricing).
Safety: Turnitin's terms say it accesses student data as an outsourced institutional function under FERPA's school official provision (34 CFR 99.31(a)(1)), and the privacy page says data is processed in Europe, Canada, Australia or the United States depending on the institution's preference. The policy page I read was last updated in October 2022, and I found no minimum age and no statement on AI training (terms, privacy). Ask your district for the current data agreement. Instead of Gradescope, a teacher grading essays should look at EssayGrader.
GradingPal: the best fit for K-12 worksheets and mixed assignments
GradingPal is the best fit for a K-12 teacher whose weekly pile mixes math worksheets, short-answer quizzes and the occasional lab report. Its pricing table lists "Handwritten and photographed work" as included on every plan, free one included (pricing). Its homepage describes splitting a scanned class set into separate student submissions.
Next to Gradescope, GradingPal covers more kinds of assignment and fewer institutions: it is priced per teacher and the free plan works without an institutional contract.
Watch out for: the terms state that "use of the services may result in incorrect output," and I found no published accuracy figure for handwriting (terms). Its terms also license the company to use user contributions "to improve our products and services," which sits awkwardly beside the privacy page's promise not to train AI on student data. Read both before a pilot.
Pricing: Free is $0 for 50 submissions a month, Lite is $14 per teacher per month (250 submissions) and Pro is $26 per teacher per month with unlimited submissions. School and district pricing is by quote (checked November 10, 2026, pricing).
Safety: The vendor displays FERPA, COPPA and GDPR badges, says "Student work is not used to train AI models," and stores data in the United States by default (security page, privacy notice). It names five AI providers: Anthropic, OpenAI, Google, Microsoft and Mistral. Under 13, accounts are allowed only in a school context. Your district still decides whether a signed agreement is required.
EssayGrader: the best fit for handwritten essays graded against a rubric
EssayGrader is the best fit for an English or social studies teacher with a class set of handwritten essays and a rubric. The vendor says to "Upload one PDF with the entire class's essays," and says it "recognizes cursive, print & messy handwriting, then transcribes it for accurate grading" (homepage). It syncs grades to Google Classroom, Canvas and Schoology, and the vendor says nothing reaches the student without the teacher's approval (Paper Grader page).
Next to Marking.ai, EssayGrader is built around US state rubrics rather than UK exam boards.
Watch out for: the vendor's own pages disagree. The pricing page says free users get up to 100 essays a month (pricing), the Paper Grader FAQ says 100 papers in the first 14 days and then 50 a month, and a help-center excerpt says 50. The pages also disagree on whether to combine essays in one PDF. I found no handwriting accuracy figure. The 94 percent on the homepage refers to agreement with teacher scores, not to reading handwriting.
Pricing: Basic is free, Pro is $14.99 a month and Premium is $34.99 a month, with district pricing by quote (checked November 10, 2026, pricing).
Safety: The homepage says "Student work is never used to train AI models" and that data is stored in the US. The terms require users to be 18 or older, which makes it a teacher tool, not a student-facing one, and name OpenAI as the AI provider (privacy policy, terms). The vendor says it acts as a FERPA "school official," which is the arrangement your district's agreement should spell out.
Marking.ai: the best fit for handwritten work marked against your own mark scheme
Marking.ai is the best fit when you already have a mark scheme or rubric and want each mark traced to the student's own words. It reads typed and handwritten work, marks "criterion by criterion," quotes the words that earned each mark, and a teacher approves every mark before a student sees it (homepage). It is built for the British curriculum (GCSE and A-Level) and the South African one (CAPS and IEB).
Next to Top Marks AI, Marking.ai marks against whatever scheme you upload, instead of tools built per exam-board specification.
Watch out for: its accuracy page is unusually frank. It reports vendor-run studies on typed English answers and handwritten math answers, lists scanned scripts end to end as "Not started," and says it has "not measured it" (accuracy page). It also reports how often two human examiners agree: the same mark on a 40-mark English essay 13 percent of the time, and on GCSE math questions 72 percent. That is a fair baseline any tool should be compared with.
Pricing: 20 free submissions, one time, with the full Pro feature set. Starter is £21 a month, Pro £36 and Plus £65, each with an allowance it does not publish. Its FAQ says payments are processed in USD even though plans are shown in pounds (checked November 10, 2026, pricing).
Safety: Student work is stored in Frankfurt and AI processing happens in the US under standard contractual clauses. The vendor says it does not use student work to train any model, though its privacy page says test-set copies of student work can remain after the original is deleted. It states no minimum age and says the school sets the lawful basis (security, privacy). That is a UK and EU frame, so a US district needs to ask about FERPA terms directly.
Top Marks AI: the best fit for batch-marking exam scripts in the humanities
Top Marks AI is the best fit for a department marking hundreds of handwritten exam scripts in subjects like history, geography and economics against UK exam boards. The homepage offers to "Batch mark handwritten scripts" across GCSE, A Level, IB and other qualifications (homepage).
Next to Running Paper, Top Marks AI builds a separate tool per qualification and subject, and it is the one that publishes a headline correlation claim.
Watch out for: the claim of 90 percent or more correlation with chief examiners applies to its AQA English tools. The homepage's own news list includes an Edexcel poetry result at 0.87. The method it cites is calibration against board standardization materials, and the page does not say who ran the comparison (homepage). It is a UK product, and I found nothing about US customers.
Pricing: No public pricing. The site says to contact the company or book a demo (checked November 10, 2026).
Safety: The privacy policy says school data is hosted in the UK and EU, with anonymized assessment content sent to AI providers in the United States. The company acts as a processor for schools, and school data is deleted within 90 days of contract end. The policy does not say whether student work trains models. Its terms and sub-processor pages refused automated reading, so hosting details beyond the policy are unverified (privacy policy).
Running Paper: the best fit for pen-and-paper assessments from ages 11 to 18
Running Paper is the best fit for a UK-style secondary school that runs regular pen-and-paper tests and wants teachers to review only the marks the system doubts. The homepage describes "AI marking for pen and paper assessments" from Key Stage 3 to Key Stage 5. Each question is marked by two AI models, a primary and a validator that never sees the primary mark, and questions where they disagree beyond a set tolerance are flagged for the teacher (how accuracy works).
Next to Top Marks AI, Running Paper centers on the review queue: it tells you which answers to read.
Watch out for: the accuracy results are the vendor's own case studies. One example: for GCSE Physics Higher, 100 scripts, "After low-confidence review, 95% of marks were within one mark of the examiner" (accuracy). That figure includes the teacher's review of flagged answers, so it describes the whole process, not the AI alone. I did not find IB coverage on the pages I read.
Pricing: A trial of up to 3 months costs £500 with full access. Schools then pay per pupil for unlimited marking or pay as they go, and no per-pupil rate is published (checked November 10, 2026, pricing).
Safety: The privacy policy says school data is hosted in the UK with some processing in the EEA and "a limited amount outside the UK and EEA," and that "neither we nor our sub-processors use student work to train AI models." Schools are the controllers and Running Paper is the processor under a data processing agreement, which I could not read (privacy policy).
Other tools that read handwriting, and the chatbots
Excelas Learn marks GCSE-style science and math scripts, publishes prices of £995, £1,495 and £1,995 per department per year excluding VAT, and says its accuracy figures come from internal benchmarks (pricing, directory listing). The directory also lists Olex.AI (English, handwritten or digital essays), TeachEdge (students submit typed or handwritten work through a portal) and GradeitAI (typed and handwritten English work, including multi-page scanned scripts). Those three descriptions come from the vendors' own listings, and I did not test them.
General chatbots such as ChatGPT, Gemini and Claude can read a photographed page, and the studies above tested GPT-5.1, GPT-5.2 and Gemini 3 Pro. What they lack is a batch workflow, a class roster and a review queue. I did not check whether MagicSchool, Brisk Teaching or Khanmigo read handwritten work, so treat that as unverified.
How to choose: four questions
- What is on the page? Numbers and short text suit Gradescope and GradingPal. Essays suit EssayGrader and Marking.ai. Exam scripts against a UK board suit Top Marks AI and Running Paper. If the answer is a drawing or a graph, plan to grade that question yourself.
- Whose mark scheme is it? If you write your own rubric, choose a tool that takes it as input (Marking.ai, EssayGrader, GradingPal). If an exam board writes it, choose a tool built around the board.
- Who signs for the student data, and where does the page image go? A photo of a handwritten quiz carries a name and a handwriting sample. Ask your district which tools have an agreement, and read the privacy page for the vendor's training position. Our guide to AI tools that keep pupil data in the UK shows how to read hosting claims.
- What does a wrong mark cost? A mistake on a weekly formative quiz is cheap. A mistake on a report card grade is not. The higher the stakes, the larger the share of papers you should read yourself.
For the grading side without handwriting, see AI grading tools that use your own rubric and how to grade faster with AI. For math specifically, see AI tools for math teachers.
A worked plan: a Grade 8 science unit test
This is a plan, not a report from a classroom. Take a unit test for 28 students with twelve questions: eight short answers, three calculations and one labeled diagram of a food web.
- Scan the class set into one PDF.
- Before any AI, grade six papers by hand, choosing the strongest student, the weakest and the three messiest handwriters.
- Run all 28 through the tool and compare its marks with yours on those six, question by question.
- If disagreements cluster on the diagram question, which is where the research predicts trouble, grade that question by hand for everyone. Keep the AI marks only for question types where it matched you on all six papers.
- Read every answer the tool flags as low confidence, plus one paper in five from the rest.
- Write your own comment at the top of each paper before it goes home.
If the tool matches you on the six papers, you have saved the time. If it does not, you have found out on six papers, not twenty-eight.
The bottom line
For handwritten work, no single tool wins. Gradescope and GradingPal suit short answers and math, EssayGrader and Marking.ai suit essays, and Top Marks AI and Running Paper suit UK-style exam scripts. All of them need a human check, and the research says the check matters most for drawings and diagrams. Compare the rest of the field in the assessment and grading category or browse all 138 tools.
Dan Fitzpatrick is a Forbes contributor, four-time #1 bestselling author and founder of The AI Educator, who has trained more than 150,000 educators across more than 30 countries.
