AI is uncannily good at judging writing

Summary

A large vendor report concerns judging agreement, not student learning.

No More Marking reports roughly 70,000 scripts from 177 secondary schools, with AI agreeing with 83% of 133,983 human comparisons. Most schools used approximately 90% AI judging.

Pairwise comparison replaces direct marking while retaining teachers. Agreement on a jointly constructed scale still needs independent validation; reliability, bias, susceptibility to gaming and learning are separate questions.

Vendor account, not independent validity evidence or a delayed, unaided writing test.