Tech developer CENTURY recently released details on its new AI tool designed to grade GCSE-level mathematics questions, claiming a 95% agreement rate with expert human markers. Schools are increasingly experimenting with automated grading, which raises questions about accuracy and fairness for parents and teachers. As schools try to reduce teacher workloads, understanding how these systems grade student work is important.
What Happened
CENTURY developed a system to evaluate student answers against GCSE math mark schemes using Large Language Models (LLMs). To handle the unpredictability of AI, they built a "marking harness" around the model. This software layer applies strict, predictable rules. For example, a student cannot receive a final answer mark unless they successfully completed the required method steps.
According to CENTURY's development team, the system relies on an "evaluation dataset" containing over 90 representative GCSE questions and more than 1,000 individually judged marks. Whenever engineers update the AI to fix an error, such as making it accept equivalent fractions instead of only decimals, they run the update against the entire dataset. This process prevents a fix for one math problem from breaking the grading on other types of questions. With this testing loop, CENTURY claims its tool agrees with human markers on over 95% of individual grading decisions.
The Bigger Picture
While a 95% accuracy rate is high, human grading is also inconsistent. A 2026 report by Top Marks AI found that experienced human GCSE markers typically achieve a 0.65 correlation with chief examiner scores. Only about 45% of their marks fall strictly within official exam board tolerances.
AI grading tools can approach these human benchmarks. A case study published in SSRN found that an AI marking system for GCSE Physics achieved a 77.5% exact agreement rate with official exam board marks, which aligned closely with human teacher evaluations.
Still, math accuracy depends heavily on how the AI is set up. A study published in the World Conference on Research in Teaching and Education showed that while a tailored AI grading tool achieved a 95.57% exact agreement rate in undergraduate math, its accuracy dropped to 43.75% when used without specialized prompting.
Parents and teachers should also consider the limits of these models. Researchers writing in arXiv found that LLMs rely on provided answer keys rather than actual mathematical understanding. This makes them prone to errors if a student uses an unexpected but valid method. At the same time, a systematic review in the Proceedings of Machine Learning Research warned that despite the growth of AI grading, only 11.2% of empirical studies evaluate these tools for demographic bias and fairness.
What This Means for Families
For families and teachers, the rise of tools like CENTURY's Maths Prep GENai shows that automated homework grading is becoming more precise. As we previously covered in our look at how schools use adaptive technology, instant feedback can help students learn faster and identify gaps immediately.
However, AI struggles with higher-order conceptual reasoning, as noted in Discover Education. Automated grading should remain a tool for homework and low-stakes practice rather than final, high-stakes exams. While AI can easily check if a final number is correct, it cannot replace a teacher's ability to understand a student’s unique thought process.
What You Can Do
First, ask if your child's school uses automated tools like CENTURY for homework or exam prep, and find out if a teacher reviews the final grades. Next, double-check the math. If your child receives an unexpected mark on an AI-graded assignment, review the work together. The AI may have missed a valid, non-standard approach to a multi-step problem. Finally, encourage your child to write down every step of their calculations. Structured answers are much easier for both human teachers and AI tools to grade accurately.