Is AI Grading Actually Accurate? We Tested It on 500 Questions

We took 500 SC06 structured questions (calculus, trigonometry, sequences — a mix of types) and had them graded blind by both AI and by three math teachers with 5+ years of independent-school experience. A reference solution was written first, then teachers and AI graded student attempts independently.

Human grading accuracy: 97.2% AI grading accuracy: 96.8% A gap of just 0.4 percentage points. The AI errors fell into two buckets: messy student handwriting causing recognition errors (60% of mistakes), and non-standard but valid solution paths (40%). The second one is what we're working on now.

Human grading: ~2.3 minutes per question AI grading: ~0.8 seconds per question A teacher marking 30 questions needs about 70 minutes; the AI does the same job in 24 seconds. That means students get feedback the moment they submit, not the next day or the day after. Immediate feedback makes a big difference to how well things stick.

AI grading does not just judge the final answer — it walks each step of the working and points to where the student went wrong. Of the 500 questions, 312 had a wrong final answer; the AI correctly located the faulty step in 94% of them. That's hard to do by hand — teachers are pressed for time and usually just write "wrong" next to the answer, rarely marking up each step.

For neatly written structured questions, AI grading has reached near-human accuracy, with a clear edge in speed and step-level analysis. AI doesn't replace teachers — their value is in the classroom, in motivating students, in reviewing questions. But on the repetitive grind of marking, it frees up a lot of teacher time for the things that matter more.

Now go practise

Understanding the theory is not enough — practice is what makes it stick. DuckMath has a topic-organised, human-reviewed question bank waiting.