The research behind Acrux.
Four things we believe about marking, and the published studies, surveys and guidance behind each one. Every entry links to its source.
Planning and marking take 42% of a secondary teacher's time.
Workload surveys in Australia and the UK return the same picture: long weeks, marking and planning the biggest blocks outside the classroom, and workload the main reason teachers think about leaving.
2025Australian Institute for Teaching and School LeadershipPlanning and marking are the two largest non-teaching duties for secondary teachers, inside a median 50-hour week.Details ▾
What it found
Lesson planning and marking are the two largest non-teaching duties reported by secondary teachers in Australia, inside a median 50-hour week for full-time classroom teachers. Across planning and marking, secondary teachers give up 42% of their time.
Why it matters
Marking is one of the largest single blocks of a secondary teacher's non-teaching time, inside a 50-hour week. Time saved on marking comes straight back out of that week.
ATWD National Trends: Teacher Workforce ↗Australian Teacher Workforce Data, June 2025 edition (2023 data) · aitsl.edu.au
2025Australian Institute for Teaching and School Leadership24% of secondary teachers mark for 10 or more hours a week, against 11% of primary teachers.Details ▾
What it found
Nearly a quarter of secondary classroom teachers (24%) reported spending 10 or more hours a week marking student work, compared with 11% of primary teachers. Most primary teachers finished their marking in under 5 hours a week.
Why it matters
The marking load falls hardest on secondary teachers. A maths teacher with several classes of written work is likely to sit in that 10-hours-plus group.
In Brief: Teacher duties ↗Australian Teacher Workforce Data (2023 data) · aitsl.edu.au
2025IFF Research and UCL Institute of Education, for the Department for Education (England)43% of classroom teachers say they spend too much time marking. 90% of those considering leaving give workload as a reason.Details ▾
What it found
Of 10,020 teachers and leaders surveyed in English state schools in 2024, 43% of classroom teachers and middle leaders said they spent too much time marking pupils' work. About a third (34%) were considering leaving the sector in the next 12 months for reasons other than retirement, and 90% of that group gave high workload as a reason.
Why it matters
Marking is one of the tasks teachers most often say takes too much of their time, and workload is the reason they most often give for thinking about leaving.
Working lives of teachers and leaders: wave 3 summary report ↗Department for Education, published January 2025 · assets.publishing.service.gov.uk
2024Roy, Poet, Staunton, Aston and Thomas (NFER), for the Education Endowment FoundationTeachers using AI for lesson preparation spent 31% less time on it, with no difference in quality, in a randomised trial of 259 teachers.Details ▾
What it found
In a randomised trial of 259 Year 7 and 8 science teachers in England, teachers who used ChatGPT for lesson preparation spent 56.2 minutes a week on it, against 81.5 minutes for the comparison group, a 31% saving. An expert panel that did not know how the resources had been made found no evidence that their quality differed.
Why it matters
This is the best current evidence that AI can take time out of a teacher’s week without lowering quality. It measured planning in science, not marking in maths, so we cite it for the principle and not for our own numbers.
ChatGPT in Lesson Preparation: A Teacher Choices Trial ↗Education Endowment Foundation evaluation report · educationendowmentfoundation.org.uk
2019Walker, Worth and Van den Brande37% of secondary teachers call workload a very serious problem and another 50% fairly serious. Marking is still singled out.Details ▾
What it found
In a 2019 survey of 7,287 teachers and leaders in English state schools, 37% of secondary respondents called workload a 'very serious problem' and a further 50% called it 'fairly serious'. Most secondary teachers and middle leaders still said they spent too much time on marking, even though marking hours had fallen by 1.7 hours a week since 2016.
Why it matters
Workload reduction efforts did move the hours, and marking was still singled out. Marking time is where the hours are.
Teacher Workload Survey 2019: Research brief ↗Department for Education (England), with the National Foundation for Educational Research · dera.ioe.ac.uk
2019Heffernan, Longmuir, Bright and KimSelf-selected sample75% of surveyed Australian teachers said their workload was not manageable; 58% intended to leave, 62% of them citing workload.Details ▾
What it found
Of 2,444 Australian teachers who answered a 2019 Monash University survey, 75% disagreed that their workload was manageable and 58% said they intended to leave the profession. Among those intending to leave, 62% gave workload as a reason. Respondents were recruited through networks and social media, so the figures describe those who took part rather than all teachers.
Why it matters
Teachers who took part drew a straight line from workload to leaving. It is a signal from a self-selected group, not a national rate, and we quote it that way.
Perceptions of Teachers and Teaching in Australia ↗Monash University, Faculty of Education · researchmgt.monash.edu
2017Higton, Leonardi, Richards, Choudhoury, Sofroniou and OwenSecondary teachers reported a 53.5-hour week, with 8.0 hours of marking and 8.4 hours of planning.Details ▾
What it found
In a 2016 survey of 3,186 teachers in England, secondary classroom teachers and middle leaders reported an average working week of 53.5 hours, including 8.0 hours of marking and 8.4 hours of planning. Most said they spent too long on both.
Why it matters
A decade-old baseline of around eight hours a week on marking for secondary teachers. It is history rather than a current figure, but it shows how long the problem has been measured.
Teacher Workload Survey 2016: Research brief ↗Department for Education (England), DFE-RB633 · dera.ioe.ac.uk
Prompt, specific feedback moves learning. A score on its own does not.
Feedback is one of the better-evidenced influences on learning, and the research is clear about which kind works: timely, specific and explanatory, rather than a score on its own.
2021Education Endowment FoundationAround six months of additional progress, five at secondary, from 155 studies. High impact for very low cost.Details ▾
What it found
The Toolkit rates feedback as high impact for very low cost, on the strength of 155 studies. Its estimate is around six months of additional progress on average, five months at secondary, with slightly higher effects in maths and science. It also warns that the effects range widely and that some studies show feedback can make things worse.
Why it matters
Feedback is one of the strongest levers a maths department has, but only when it is specific and the student can act on it. The Toolkit’s warning to watch the workload cost of written marking is worth noting too.
Teaching and Learning Toolkit: Feedback ↗EEF, review last updated June 2021 · educationendowmentfoundation.org.uk
2020Wisniewski, Zierer and HattieAverage effect of feedback d = 0.48 across 435 studies. High-information feedback reached d = 0.99.Details ▾
What it found
A meta-analysis of 435 studies and more than 61,000 students put the average effect of feedback on learning at d = 0.48. The effect depended heavily on how much information the feedback carried: high-information feedback reached d = 0.99. Feedback had more effect on cognitive and motor skill outcomes than on motivation or behaviour.
Why it matters
The useful feedback is the kind that tells a student what went wrong and what to do about it. A score or a tick carries little of that information.
The Power of Feedback Revisited: A Meta-Analysis of Educational Feedback Research ↗Frontiers in Psychology, 10, 3087 · doi.org
2015Van der Kleij, Feskens and EggenAbstract readExplanatory feedback: effect size 0.49. Right-or-wrong only: 0.05. Effects larger in mathematics.Details ▾
What it found
A meta-analysis of 40 studies of feedback in computer-based learning found that elaborated feedback, such as an explanation, had an average effect size of 0.49. Feedback that only said whether an answer was right or wrong scored 0.05, and giving the correct answer 0.32. Effect sizes were larger for mathematics than for the other subjects studied, and smaller for primary and high school students than for older learners.
Why it matters
A right-or-wrong mark on its own does almost nothing. The gain is in the explanation. The maths result is encouraging, and the smaller effects at school level are a reason not to oversell it.
Effects of Feedback in a Computer-Based Learning Environment on Students' Learning Outcomes: A Meta-Analysis ↗Review of Educational Research, 85(4), 475–511 · doi.org
1988Kulik and KulikAbstract readIn real classroom studies, immediate feedback usually beat delayed feedback (53 studies).Details ▾
What it found
A meta-analysis of 53 studies on feedback timing found that applied studies using real classroom quizzes usually found immediate feedback more effective than delayed. Experimental studies of acquiring test content usually found the opposite, and laboratory list-learning studies varied with how they were designed.
Why it matters
In real classrooms, sooner has generally beaten later. We read it as a tendency rather than a rule, because the laboratory results point both ways.
Timing of Feedback and Verbal Learning ↗Review of Educational Research, 58(1), 79–97 · doi.org
The errors that matter live in the written working, not the final answer.
A student's written working shows where their thinking went right and where it went wrong. Exam boards award marks for method, and the errors in that working are the best guide to what to reteach.
2022Henderson, Hodgen, Foster and Kuchemann, for the Education Endowment FoundationAssessment should tell teachers what pupils do and do not know, and knowledge of common misconceptions should shape planning.Details ▾
What it found
The guidance recommends that assessment be used 'not only to track pupils' learning but also to provide teachers with information about what pupils do and do not know', and that it should inform planning and targeted support. It adds that 'knowledge of common misconceptions can be invaluable in planning lessons to address errors before they arise'.
Why it matters
Assessment is meant to tell the teacher what to do next. Marking that shows where each student’s working went wrong supplies exactly that information.
Improving Mathematics in Key Stages 2 and 3 ↗EEF guidance report, 2022 update (first published 2017) · educationendowmentfoundation.org.uk
2015AQAMethod marks and follow-through marks are awarded for correct working even when the final answer is wrong.Details ▾
What it found
AQA's mark scheme awards method marks 'for a correct method which could lead to a correct answer' and follow-through marks 'for correct working following a mistake in an earlier step'. Examiners are told that 'credit should be given to any valid methods' and that marks are usually not awarded to students who show no working.
Why it matters
In this exam a sound method earns credit even when the final answer is wrong. Marking that reads only the answer gives students less credit than the exam board does. One board’s scheme, cited as a concrete example of method marking.
GCSE Mathematics specimen mark scheme, Paper 3 Higher (8300/3H) ↗AQA, England · filestore.aqa.org.uk
2014Booth, Barbieri, Eyer and Paré-BlagoevSpecific errors in students' written algebra working predicted lower end-of-year scores (37 classes, 13 schools).Details ▾
What it found
Researchers coded the written working of students in 37 Algebra I classes across 13 US schools, on practice assignments that asked them to show all their work, and tracked seven types of error through the school year. Sixteen error-and-timing combinations were negatively correlated with end-of-year test scores. Equals-sign errors in the middle or at the end of the year, and negative-sign and arithmetic errors at the end of the year, all predicted lower scores.
Why it matters
The errors that predict later difficulty live in the working, not the final answer. A teacher who can see which error a student is making, and when, has something specific to reteach.
Persistent and Pernicious Errors in Algebraic Problem Solving ↗The Journal of Problem Solving, 7(1) · doi.org
2011Anthony, Yang and KoedingerPosition paperHandwriting is more usable than typing for entering maths, with the same learning gains in less time.Details ▾
What it found
The team that built handwriting input for an algebra tutor reports that typing and mouse interfaces 'do not allow students to easily enter math expressions' such as fractions and exponents. They summarise their earlier studies as finding handwriting more usable than typing for entering maths, with students reaching the same learning gains in less time on a handwriting-based tutor. This paper summarises that earlier work and reports no new data.
Why it matters
Maths notation is spatial. Typing it adds effort that has nothing to do with the mathematics, and students writing on paper avoid that overhead.
Handwriting Interaction for Math Tutors: Lessons for HCI in Education ↗CHI 2011 workshop paper, Carnegie Mellon University · lisa-anthony.com
AI marking works when a teacher reviews the result and keeps the final say.
Human marking is less consistent than most people assume, and the early evidence on automated marking of handwritten maths points the same way as government guidance: useful when a teacher reviews the result and keeps the final say.
2026Vanhoyweghen, Holst, Mobini and colleaguesPreprintAI-to-human marking agreement on handwritten maths (κ 0.50 to 0.91) was comparable to human-to-human (0.70 to 0.96).Details ▾
What it found
Researchers used an AI model to mark handwritten answers in two undergraduate maths courses, with every result checked by a person before release. Agreement between the AI and a marker (quadratically weighted kappa 0.50 to 0.91 across questions) was comparable to agreement between two human markers (0.70 to 0.96). The few larger AI errors were caught at the human verification step.
Why it matters
The right question is whether AI marking sits within the range of human variation, and here it did. The study is undergraduate maths and not yet peer reviewed, so we cite it as early evidence.
Human-in-the-Loop LLM Grading for Handwritten Mathematics Assessments ↗arXiv:2603.13083, Vrije Universiteit Brussel · arxiv.org
2026Levine, Aenlle, Zilles, West and SilvaPreprint87% to 99% agreement with the instructor on rubric items. Most errors came from misreading the handwriting, not the rubric.Details ▾
What it found
Vision-capable AI models checked photographed handwritten university STEM work against instructor rubrics. Agreement with the instructor on individual rubric items ranged from 87% to 99% across questions and models. In the best model, 87% of the errors came from misreading the handwriting or the photo rather than from misapplying the rubric, and the authors say high-stakes use 'demands greater reliability'.
Why it matters
The weak point is reading the page, not applying the marking criteria. That is the kind of error a teacher looking at the scan catches at a glance.
Automated Grading of Handwritten Mathematics Using Vision-Capable LLMs ↗arXiv:2605.19043, University of Illinois Urbana-Champaign · arxiv.org
2025Department for Education (England)AI can reduce the burden of feedback and marking, provided staff use their professional judgement.Details ▾
What it found
The guidance says research shows generative AI 'could also be used for tasks such as feedback and tailored support in schools', and describes investment that 'aims to help reduce the burden of feedback and marking on teachers'. It also says staff 'must use their professional judgement when using these tools' and that AI 'cannot replace the judgement and deep subject knowledge of a human expert'.
Why it matters
Government guidance welcomes AI that cuts marking workload, provided a professional checks the output. Teacher-reviewed AI marking is built on that position.
Generative artificial intelligence (AI) in education ↗GOV.UK policy paper, updated 12 August 2025 · gov.uk
2023National AI in Schools Taskforce, Australian Government Department of EducationTeachers retain control of decision making and remain accountable for decisions supported by AI (principle 5.1).Details ▾
What it found
The national framework says teachers and school leaders 'retain control of decision making and remain accountable for decisions that are supported by the use of generative AI tools' (5.1), that tools should be used in ways that support teacher expertise (1.3), and that people affected by AI tools should be able to question their outputs and any decisions informed by them (5.4).
Why it matters
A mark is a decision about a student. The framework expects a teacher to stay in control of that decision and accountable for it, which is why the teacher reviews and can override every mark.
Australian Framework for Generative Artificial Intelligence in Schools ↗Commonwealth of Australia · education.gov.au
2018Rhead, Black and Pinot de MoiraExperienced human markers agree most in maths: 0.96 probability of the definitive grade, against 0.52 in English.Details ▾
What it found
Ofqual compared examiners' marks on seeded responses with the senior examiner's definitive mark across GCSE and A level subjects. The probability of a student receiving the definitive qualification grade ranged from 0.96 for a mathematics qualification to 0.52 for English language and literature. For nearly every qualification the probability of the definitive or adjacent grade was above 0.95.
Why it matters
Experienced human markers do not always agree, and agreement is highest in maths because the answers are more objective. That is the realistic benchmark for any marking, human or AI, and the reason a teacher reviews each mark.
Marking consistency metrics: An update ↗Ofqual (England) · assets.publishing.service.gov.uk
See it mark real handwritten work.
A short demo with Leanne, our co-founder and a maths teacher of 25 years. Bring a class set if you like.