New Zealand’s qualifications authority will keep using an AI system to help mark NCEA literacy assessments after an independent review found it fit for continued operational use. The review found AI scores exactly matched human markers 79.6% of the time, while NZQA says human markers retain oversight and take precedence when there is disagreement.
Watch Desk analysis
What happened
The New Zealand Council for Educational Research (NZCER) reviewed NZQA’s Automated Text Scoring model, which has helped mark Literacy Writing assessments since 2025. NZCER found AI scores either matched human scores or were one mark away in almost all cases; 84.5% of total scores were within one point of human markers.
NZQA says about 30% of scripts were cross-checked by human markers, with particular attention to students near the Achieved/Not Achieved boundary. Deputy chief executive Jan Marshall said the AI’s exact-match rate does not mean that 20% of students received a different final grade. NZQA said marking, alongside other process changes, helped return results up to four and a half weeks sooner in 2025 than in 2024.
The review found no substantial differences in how closely AI and human scores aligned across gender, ethnicity or socioeconomic groups. It recommended improvements to quality checks and continued monitoring. NZQA is testing whether the technology could mark more complex standards using scripts from previous years, not current student assessments. The authority says any expansion would require careful testing and continued human oversight.
Why it matters
This is AI being used in a consequential assessment, not just a classroom writing aid. The reported results show where the system agrees with human markers, while the human checks and review recommendations matter because a score near the pass boundary can affect a student’s next steps.
The figures do not mean every script gets a second human mark, nor do they establish that the system is error-free. NZQA says it is unaware of a case where a disagreement led to a grade needing correction, and says a human takes precedence when markers disagree.
Our read
The useful detail is the arrangement around the AI: human oversight, extra attention near the pass boundary and a review that still calls for better checks and monitoring. Faster results are a practical benefit, but they are not a substitute for showing how borderline work is handled. NZQA should publish clear evidence about that process before extending automated marking to harder assessments.
What to watch
- Whether NZQA publishes details of the review’s recommended quality-check improvements.
- How it handles borderline assessments and disagreements between AI and human markers.
- Whether testing on past scripts leads to a proposal to mark more complex standards.
Discussion spark: Should AI marking expand to more complex exams only after independent reviews, or is ongoing human oversight enough?
Sources and evidence
- Source update (10 November 2026, 00:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.