Teachers know better than anyone the sheer volume of marking that comes with the job. It’s no surprise, then, that the idea of using AI to mark students’ work feels appealing.

AI could free up precious time for teachers to focus on teaching, planning and supporting students. But there are trade-offs, such as being one step removed from the understanding of students’ work that comes from reading it first-hand.

When it comes to marking external exams, different questions arise. It might seem at first glance that AI could be well-suited to achieving the kind of national consistency and reliability we expect for large-scale national qualifications.

However, where the stakes are high for students and for public trust, we must pay as much attention to AI’s potential to erode trustworthiness as we do to its promise of improved efficiency.

At Ofqual, we’ve just published new research exploring AI’s potential role in marking and its current capabilities and constraints.

Our conclusion is clear: AI is promising for quality assurance and marker training, but for the moment it’s nowhere near ready to take over high-stakes marking. For AI to have a role in high-stakes marking, it would need to align fully with the values that make our exam system fair, transparent and trusted.

Could AI mark exams?

Current AI systems, particularly those based on large language models, function as “black boxes”. Even expert designers cannot straightforwardly explain why specific student responses generate particular marks. This opacity creates serious challenges for high-stakes assessment.

Empirical studies have found that AI’s performance in marking can vary across demographic groups. For example, a study by Johnson and Zhang found that a scoring algorithm “appeared to be treating different race/ethnicity groups differently”. A systemic problem that risks undermining the fairness of assessments can be hidden inside the black box.

The lack of transparency also raises a fundamental question about contestability. A basic principle underpinning public trust is that students can challenge their marks and request a review if they believe something is wrong.

When marks are generated by AI, it becomes much harder to explain or scrutinise how decisions were made, and what rights exist to challenge it.

Transparency, fairness and contestability aren’t merely ethical nice-to-haves for high-stakes marking - they’re a necessity.

They are what underpin public trust in qualifications. A marking system might be technically perfect, but if people do not understand how it works or who is responsible when something goes wrong, it does not pass the trust test.

Public trust in qualifications

When students sit GCSEs and A levels, they’re entering into an implied social contract.

They submit to standardised testing on the promise that their results will have real value in education and employment. This promise only holds if the public - students, parents, teachers, employers, universities - have confidence that the marking process is fair and transparent.

Any innovation must come with a cast-iron assurance that it will maintain trust in qualifications.

Ofqual’s priorities remain ensuring fairness for students; making sure assessments measure the right knowledge, skills and understanding; and maintaining public confidence. We are also committed to supporting responsible innovation.

AI has genuine potential in quality assurance and marker training. But for high-stakes marking itself, we must move beyond treating AI as simply technological progress and grip the ethical and social considerations.

Students’ legitimate expectations currently include that their exam work is marked by trained examiners intelligently applying expert and well-practised judgement.

Any implementation of AI in high-stakes marking must take these expectations seriously, not as obstacles to progress but as essential features of a system that works precisely because it is trusted.

Sir Ian Bauckham is chief regulator at Ofqual

You can now get the UK’s most-trusted source of education news in a mobile app. Get Tes magazine on iOS and on Android