[Written in June 2026]
Every generation of educational technology arrives with the same promise: it will save teachers time.
Email would simplify communication. Interactive whiteboards would transform lessons. Management information systems would reduce administration. More recently, artificial intelligence has been presented as the next great solution to teacher workload, with marking frequently cited as one of its most obvious applications.
There is truth in that claim. AI can reduce the time spent marking.
My concern is that we are asking the wrong question.
The more important question is whether AI reduces workload at all – or simply changes where that workload sits.
I recently put this to the test while developing an AI-assisted workflow capable of marking more than 4,000 short-answer responses. As a secondary Computer Science teacher with a software development background, I was interested in whether AI could produce reliable marking at scale.
It could.
But not immediately, and not cleanly.
At one point early in testing, I ran a batch of responses through the system and found something puzzling: nearly identical student answers were being awarded different marks. At first I assumed the model was inconsistent. In fact, the issue was in my mark scheme. Human markers silently reconcile this kind of ambiguity. The AI did not. It exposed it.
That moment changed how I approached the entire project.
The surprise was that the biggest challenge had almost nothing to do with the AI itself.
Instead, the project exposed something far more familiar: the quality of our assessment design.
The AI had little difficulty applying a clear mark scheme to a well-written question. Where it struggled, it was usually for the same reasons students struggle. Questions relied on context from earlier parts of the assessment. Mark schemes left room for interpretation. Learning objectives were bundled together into a single question. Human markers compensate for these issues instinctively. AI does not.
The solution was not to search for a better AI model.
It was to write better assessments.
Questions became more precise. Mark schemes became more explicit. Acceptable alternative answers were identified. Common misconceptions were anticipated. Each refinement improved the consistency of the marking.
In practice, I ended up doing something slightly uncomfortable: rewriting questions I had previously been confident were “fine”. The AI forced a level of precision that manual marking had never required.
In other words, I wasn’t really improving the AI. I was improving the assessment.
That has implications well beyond artificial intelligence.
Good assessment has always depended on clear questions and carefully constructed success criteria. AI simply makes the consequences of getting those things wrong more visible. If anything, it reinforces the value of teachers’ professional expertise rather than diminishing it.
There were other positives. The quality of the formative feedback generated was better than I had expected, often identifying precisely what information a student had omitted and suggesting how an answer could be strengthened. In one case, a response that I would have marked as “nearly there” was consistently downgraded by the system until I realised I had been implicitly rewarding a misconception. I had to adjust my own judgement, not just the model.
Reviewing AI-generated marking also proved significantly quicker than marking every response from scratch.
However, “quicker” should not be confused with “automatic”. Every response still required professional oversight. AI can support marking, but it does not replace the need for teacher judgement, particularly where nuance, borderline responses or student expression may affect outcomes. Responsibility for final assessment decisions remained with the teacher.
The responsibility for assessment remained mine, not the software’s. Perhaps the greatest risk is that, once AI appears consistently reliable, teachers may feel tempted to stop checking its decisions with the same level of scrutiny.
But even that is not the issue that has stayed with me.
What has stayed with me is the amount of work required before the AI could be trusted.
Designing assessments suitable for automated marking took time. Refining mark schemes took time. Testing edge cases took time. So did designing the prompts that structured how the model interpreted responses. And even after all that, I still found myself making small adjustments during live marking because unexpected student phrasing would surface assumptions I hadn’t fully captured.
None of this work disappeared simply because another part of the process became automated.
This is a pattern education has seen before.
Technology rarely removes professional work altogether. More often, it redistributes it. Administrative tasks become data management. Preparing resources becomes designing digital learning. Marking becomes validating AI outputs.
The mechanical work decreases. The cognitive work often increases.
That distinction matters because cognitive workload is already one of the greatest pressures on teachers. Every lesson, assessment and interaction requires hundreds of professional judgements. If AI reduces repetitive marking but replaces it with new expectations around assessment design, quality assurance, data governance and AI literacy, we need to be honest about whether workload has genuinely fallen – or simply changed shape.
This is not an argument against AI.
Quite the opposite.
Used thoughtfully, AI has enormous potential to reduce repetitive tasks and improve access to high-quality feedback. My own experience convinced me that it has a valuable role to play in assessment.
But if we expect every teacher to become an expert in writing AI-ready assessments, validating automated marking and keeping pace with rapidly evolving technologies, we risk replacing one workload problem with another.
The better solution may be to think more collectively. High-quality assessment materials, validated mark schemes and robust AI workflows could be developed collaboratively across departments, trusts or subject communities, allowing teachers to benefit from AI without individually recreating the same work.
The debate around AI in education often focuses on what the technology can do.
Perhaps it is time to focus instead on what we are asking teachers to do because the technology exists.
Whether AI ultimately reduces workload will depend less on advances in artificial intelligence than on the choices school leaders and policymakers make about how its use is implemented.
Technology doesn’t simply reduce workload.
It redistributes it.
Our challenge is to ensure that, this time, it redistributes it wisely.
