We believe in transparency. Rather than making vague claims, we publish real accuracy data — tested against official examiner marks, and against a public benchmark anyone can check — so you can decide for yourself.
Last updated 17 July 2026
Most marking tools compare themselves to a single examiner. That tells you very little, because examiners disagree with each other — a lot. So we tested against a benchmark where every essay was marked twice, independently. That gives an honest yardstick: how close is a second human?
| On 60 double-marked GCSE essays | Graded Pro | A second examiner* |
|---|---|---|
| Agreement with the examiners (kappa) | 0.85 | 0.77 |
| Average difference from the examiner mark | 2.8 marks | 4.6 marks |
| Essays 8+ marks adrift | 1 in 60 | 12 in 60 |
| Marking bias | +0.7 marks | — |
*Essays are from the Medly marking benchmark (Fox et al., 2026), a public dataset of real GCSE mock responses each marked independently by two qualified examiners, released under CC BY 4.0. The 60 essays are two 40-mark extended writing tasks, 30 responses each, spanning the full mark range. “A second examiner” is the agreement between the two human examiners on the same scripts. Graded Pro marked every essay from the question and mark scheme alone, with no examiner marks visible, using our standard production settings. Our marking bias of +0.7 marks means we are very slightly more generous than the examiner average. Anyone can download the dataset and repeat this.
All results are from real examination papers, compared against the actual marks awarded by the official examiner. The only inputs were the students' work and the official mark scheme — nothing was adjusted or modified.
A note on kappa. It is the measure exam boards use for marker agreement, and it can be inflated by pooling questions of very different sizes together — mixing 1-mark answers with 40-mark essays makes any marker look better than it is. We calculate kappa per question and then average, which is the stricter and more standard method.
Our system is at its strongest on questions with defined correct answers — the kind that make up the majority of assessments. Across 609 structured questions in maths and English:
Whether it's a 1-mark calculation or an 11-mark multi-step problem, the AI consistently matches professional marking standards — and the maths papers above were marked from handwritten scripts, not typed answers.
Levelled questions — where markers use band descriptors to judge quality rather than tick off correct answers — are the hardest thing in marking, for anyone. Here is what the double-marked data actually shows.
AI marking is not a replacement for your professional judgement — it's a tool that handles the heavy lifting so you can focus on what matters.
Short-answer questions, calculations, retrieval tasks, and structured responses across all subjects. On these the AI matches the examiner on roughly 85% of questions and lands within a mark on 97% — reliable enough to use as a first pass and review by exception.
Extended writing. Not because the AI is unreliable — on double-marked essays it sits closer to the examiners than a second examiner does — but because essay marks are genuinely contested and carry the most weight for your students. Moderate a sample, as you would with any marker, and look at anything close to a grade boundary.
Use AI marking to get a fast, accurate first pass across a full class set. Moderate a sample — just as you would with your own marking or a colleague's — and adjust where your professional judgement says so. Teachers who use this approach typically report saving 50–70% of their marking time.
Our accuracy benchmarks are based on formal examination papers, but Graded Pro is built for everyday marking across all types of student work. The same AI that matches examiner standards on exam scripts delivers consistent, rubric-linked feedback on:
Wherever there's a rubric or mark scheme, Graded Pro delivers accurate, detailed feedback — whether the stakes are high or the goal is simply helping students learn from their work.
We continuously test and improve our marking accuracy. We don't claim perfection — no marker, human or AI, achieves that, and the data on this page shows just how far from perfect human marking is on the hardest questions.
What we promise is that we publish what we find, including when it is unflattering. We revised this page downwards in July 2026: our previously published kappa of 0.97 used a calculation that pooled 1-mark answers with 40-mark essays and flattered the result. The honest figure is around 0.85. We also previously warned that we under-marked the strongest essays; when we tested that against double-marked data rather than a single examiner, it did not hold, and we removed it.
Every time we change the underlying model, we re-test every paper. Where our sample is small, we say so. Where you can check our work yourself, we tell you where to find the data.
Sign up for a free trial with 150 free credits and test it on your own papers.
Start Free TrialNo credit card required
Teacher Free
Try Graded Pro risk free
$0
150 Credits Free
50 FREE Credits / Month - Max 50
Full Graded Pro Toolkit
Email Support
Single User
Teacher Pro
Best Choice for Busy Teachers
$25
6000 Credits
No Expiration Date
Access to all Features
Priority Email Support
Single User
School Account
Whole School or Departments
Custom Quote
Pooled Credits
No Expiration Date
All Features + User Dashboard
Dedicated Manager
Multiple Users
The number of credits needed to mark each student's work depends on the amount of text and images submitted, whether the work is handwritten or typed, and the level of feedback required.
Basic task
1¢
~3 credits
In-depth task
~3¢
8–10 credits
Full exam marking
~10¢
~30 credits
Weekly grading for 125 students (e.g. 5 classes) requires around 4,500 credits per month.