Is AI marking accurate?

The short answer

On the best-evidenced task — judging which of two pieces of writing is stronger — AI now agrees with teachers about as often as teachers agree with each other. The largest UK trial found 83% agreement across roughly 70,000 scripts. On exam-style essays with published mark schemes, vendor studies report agreement in the same range or higher.

The honest caveats: almost all of the published evidence is secondary-level; agreement on a ranking or a mark is not the same as writing the right next step for a nine-year-old; and AI is strongest where criteria are structured and weaker on nuance — a child's deliberate rule-breaking, humour, or a voice that doesn't fit the rubric. Which is why the most important accuracy question isn't the headline percentage. It is: who catches the errors?

What does the published evidence show?

The most substantial public data comes from No More Marking, who run comparative judgement assessments across thousands of schools. In their 2025 assessment of around 70,000 pieces of Year 7–9 writing across 177 UK secondary schools, AI judges agreed with 83% of the 133,983 decisions made by human teachers — the same level at which human judges typically agree with each other on their projects. Of the disagreements, 90% fell within 45 points on their 300–700 scale, and their review found the largest disagreements traced to human error rather than AI error.

An earlier, smaller trial reported by Schools Week (5,251 Year 7 pupils, 44 schools) found 81% AI–human agreement, against a human–human benchmark of 87% — encouraging, with the authors still recommending humans stay in the loop.

At GCSE, Top Marks AI publishes its own accuracy studies, including 93% agreement with experienced human markers on handwritten AQA English Literature Shakespeare mocks. Vendor-published numbers deserve the usual scrutiny, but publishing them at all is the right habit, and the pattern matches the independent data: AI marking agrees well with humans where the task is structured and the mark scheme is explicit. What none of these studies measure is the task primary teachers actually need — formative feedback on a child's everyday independent writing. That evidence base is still being built.

What does “accurate” mean for feedback rather than grading?

When AI is used for grading, accuracy is everything, because the output is a verdict. A wrong mark on an exam script gets recorded, feeds a data drop, and follows the student. Nobody downstream is checking each one; that is the point of automating it.

Formative feedback fails differently. A wrong suggestion — a next step pitched too high, a strength the teacher knows was a fluke — is a draft a professional reads before a child does. The error costs the teacher a moment's editing, not the child a result. The EEF's evidence on feedback (an average of six months' additional progress across 155 studies, one of the strongest effects in the Toolkit) also carries its caveat: not all feedback is equal, and feedback only works when pupils get information they can use. Specific, criteria-linked, same-lesson feedback — checked by the person who knows the child — is what that evidence describes.

So the accuracy bar depends on the job. A grading tool must be right on its own. A feedback tool must be right enough, specific enough, and transparent enough for a teacher to make it right quickly.

Why does teacher-in-the-loop design change the risk?

Every study above found some AI–human disagreement. The design question is what happens to those cases.

• Errors get surfaced, not buried. A criteria-level breakdown shows why the AI thinks a standard was met, so a wrong call is visible against the evidence rather than hidden inside a single score.
• The reviewer knows the child. The teacher can catch exactly the errors AI is weakest on — nuance, intent, voice, what this child did last week — because that context never reaches the model.
• Review is cheap; wrong verdicts are not. Editing a drafted comment takes seconds. Correcting a recorded judgement after the fact — or never noticing it — is where the real cost of AI error lives.

What does Howay do about accuracy?

Common questions

Is AI marking as accurate as a teacher?

In the largest published UK trial — No More Marking's AI-enhanced comparative judgement of around 70,000 pieces of writing across 177 secondary schools — AI agreed with teachers' judgements 83% of the time, which matched the level at which human judges typically agree with each other. That is agreement on which of two pieces is better, not on every detail of feedback. Accuracy is strongest where the criteria are structured and weaker on nuance, which is why a teacher should review AI output before it reaches a child.

How accurate is AI at judging children's writing?

The strongest published evidence comes from comparative judgement trials at secondary level: 83% agreement with teachers across roughly 70,000 Year 7–9 scripts in 2025, and 81% in an earlier trial where human judges agreed with each other 87% of the time. Vendor-published studies at GCSE report similar or higher agreement on exam-style mark schemes. Published evidence specific to primary-age writing is thinner, which is worth knowing before anyone quotes a single accuracy number at you.

Can AI give accurate feedback on primary children's writing?

It can give useful feedback, and its usefulness depends on design: whether it assesses against the curriculum criteria the teacher chose, whether it can read the child's actual handwriting, and whether a teacher reviews the output. For formative feedback the accuracy question changes shape — a wrong suggestion is caught and edited by the teacher, where a wrong grade would simply be recorded. That is why teacher-in-the-loop design matters more than any single headline percentage.

What should teachers check before trusting an AI marking tool?

Four things. Whether the tool publishes accuracy or agreement data, and on what age group and task type — secondary exam essays and primary independent writing are different jobs. Whether it assesses against the criteria you actually use. Whether you can see and edit everything before a child does. And whether the errors it makes are visible ones you would catch in review, rather than silent ones buried in a score.

Does Howay publish its own accuracy data?

Not yet — and we won't quote a number before we have one worth quoting. Howay will publish its own teacher-agreement data as it matures. In the meantime the design assumes imperfection: feedback is drafted against the criteria the teacher chose, broken down criterion by criterion so errors are visible rather than buried, and nothing reaches a child or parent until the teacher has reviewed and edited it.