NZQA backs AI marking of NCEA writing, despite one-in-five scoring differences
NCEA writing assessments scored by artificial intelligence don’t match human markers all the time – but the New Zealand Qualifications Authority (NZQA) says students aren’t being disadvantaged.
An independent review by the New Zealand Council for Educational Research (NZCER) has found NZQA’s automated text scoring model is “fit for continued operational use” in marking NCEA co-requisite Literacy Writing assessments.
It said the system demonstrated “strong operational performance”, with AI scores exactly matching human markers 79.6% of the time.
But NZQA deputy chief executive assessment Jan Marshall said that figure did not mean 20% of students received a different final grade.
The system marked different elements of a written response, including content, structure and accuracy, which could differ from human assessments without changing the overall result.
She said any disagreement was taken seriously, with experienced human markers retaining oversight of the process.
“We don’t rely on AI alone.”
The system has been used since 2025 to help mark the Literacy Writing co-requisite, which students must pass before receiving an NCEA certificate.
NZQA introduced the technology in response to schools wanting quicker results, allowing students more time to prepare for another assessment if necessary.
It said AI marking, alongside other process improvements, helped to return results to schools and students up to four and a half weeks sooner in 2025 than in 2024.
The NZCER review found that AI scores either matched or were just one mark away from human scores in almost all cases.
Overall, 84.5% of AI total scores were within one point of human markers.
But the AI tended to score slightly lower on average, raising questions about whether students close to the pass mark could be disadvantaged.
Marshall said those students received additional human scrutiny.
About 30% of scripts were cross-checked by human markers, with particular attention paid to assessments near the Achieved/Not Achieved boundary.
Marshall said NZQA was not aware of any case in which a discrepancy between AI and human marking had resulted in a grade being changed.
But she said, “if there’s disagreement, the human always takes precedence”.
The NZCER review found “no substantial differences” in how closely AI and human scores aligned across gender, ethnicity or socio-economic groups.
However, it recommended improvements to quality checks and continued monitoring of the model’s performance.
Marshall said findings from four independent reviews – two by NZCER and two by technical consultancy ClearPoint – had reinforced NZQA’s confidence in the system.
But she said human markers would remain central to assessment for the foreseeable future.
NZQA is also exploring whether AI could be used to mark other assessment standards, including more complex work requiring it to distinguish between Achieved, Merit and Excellence standards.
Marshall said it was testing the technology using scripts from previous years, rather than current student assessments.
Any expansion would require careful testing, with human oversight remaining essential.
NZQA does not recommend that schools use AI to mark internal assessments, saying they generally don’t have enough scripts to adequately train a reliable model.
Jaime Cunningham is a Christchurch-based reporter with a focus on education, social issues and general news. Jaime joined Newstalk ZB in 2023, after working as a sports reporter at the Christchurch Star.