COMPLIANCE · OFQUAL
Is AI Allowed to Mark Assignments? Ofqual’s Rules Explained for Training Providers
The short answer is no, not on its own. What Ofqual actually requires, and what compliant AI use in marking looks like before you sign anything.
By Eduface · July 2026 · 9 min read
Here’s a question we get asked in almost every sales call: can we just let the AI mark the assignments? Not help with marking. Actually mark them, start to finish, no assessor required.
Can AI mark assignments on its own?
No. Not because the technology can’t produce a grade (it can, and often a plausible one) but because the regulator has looked at exactly this scenario and said, quite plainly, that it doesn’t count as valid marking on its own. If you run a training provider and you’re weighing up AI tools right now, this is the single most important piece of context you need before you sign anything.
What Ofqual is actually regulating here
Ofqual is the body that regulates qualifications in England, and in early 2026 it published a working paper laying out its current thinking on AI in marking. The tone of that paper is worth noting: it isn’t hostile to AI. It calls the technology promising, particularly for quality assurance and for training human markers. But it draws a hard line at using AI as the sole decision maker in a high stakes marking judgement. Ofqual’s own words are that this doesn’t satisfy the requirement for human based judgement in marking, full stop.
That’s a regulatory conclusion, but it maps neatly onto a much older academic argument about automated decisions generally. Danielle Citron and Frank Pasquale, in a widely cited 2014 paper in the Washington Law Review, made the case that automated scoring systems tend to be opaque and under scrutinised precisely because they look objective on the surface. Their argument was about credit scoring and employment screening, not education, but the underlying worry, that a number can hide the reasoning behind it and become very hard to challenge, applies just as well to a grade. Ofqual’s insistence on human judgement is, in effect, a due process safeguard for exactly that reason.
The three problems Ofqual keeps coming back to
If you read through Ofqual’s public statements on this, three concerns show up again and again.
Bias. Models can end up rewarding things that have nothing to do with the quality of the work: sentence length, vocabulary, how confident the tone sounds. None of that is a stand in for whether the argument is actually good.
Inaccuracy. Even a well built model gets things wrong sometimes, and the errors aren’t always small. A model can be right on average and badly wrong on the paper that matters most, which in an assessment context is usually the borderline one.
Transparency. If a learner wants to know why they got the mark they got, or wants to appeal it, the algorithm said so is not an answer that holds up. Someone has to be able to explain the decision in terms a human being can actually evaluate.
Ofqual’s own chief regulator has been unusually candid in public about this, pointing out that AI in its current form still makes mistakes, still invents plausible sounding detail that isn’t true, and still reflects the biases baked into whatever it was trained on. That’s a useful corrective if you’ve been reading vendor marketing that implies the accuracy problem is basically solved.
The other half of the picture: what learners can and can’t do
Ofqual’s marking guidance is one side of this. JCQ guidance covers the other: what students themselves are allowed to do with AI while producing their work. Learners aren’t permitted to generate their submission with AI and then hand it in as their own unaided effort. Centres carry responsibility for the authenticity of what gets submitted, which means this isn’t purely a how do we mark question. It’s also a what are we actually checking for question, and the two are more connected than they first appear. A marking tool that can’t tell polished, well supported human reasoning apart from confident sounding AI output is a weak link in both directions at once.
So what does compliant AI use in marking actually look like
The test that matters in practice is simple, even if the technology behind it isn’t: does an assessor see the AI’s proposed grade, understand the reasoning behind it, and make an explicit decision to approve it before it goes anywhere near the learner? If yes, you’re on solid ground. If a grade can reach a student without a human ever making that decision, you have a problem, regardless of how the tool is marketed.
Ask a vendor exactly how a grade becomes final, not whether review is possible in theory. There’s a real difference between a step that’s technically available and one that’s structurally required.
Keep evidence of what the human assessor actually did with each submission. If a grade is ever challenged, you want a record showing a person engaged with that specific piece of work, not just that the software was switched on that day.
Check what your own awarding organisation has said, in writing, about AI use. Ofqual leaves room for awarding organisations to set their own detailed rules within the wider framework, and not all of them have published a position yet. If yours hasn’t, don’t assume that means it’s fine. Ask.
This is a live area, not a settled one
Ofqual has said it intends to keep gathering evidence, including evidence on malpractice involving AI, and to update its guidance as that evidence grows. There’s also been recent signalling from the regulator about tighter scrutiny of coursework generally, as capable AI becomes something every student has access to. None of that changes today’s answer. It does mean today’s answer has a shelf life, and it’s worth checking back on this rather than treating one policy reading as permanent.
Frequently asked questions
Can AI mark my learners’ work without a human checking it first?
No. Ofqual’s current position is clear that AI cannot be the sole mechanism for awarding a mark. A human decision has to be part of it.
Does that mean AI has no place in marking at all?
Not at all. It means AI needs to support the decision rather than make it. A tool that drafts a grade and holds it for an assessor’s explicit sign off fits comfortably within the current rules.
Is the rule the same everywhere, for every qualification?
The principle that human judgement must be involved comes from Ofqual and applies broadly. Individual awarding organisations can add their own specific guidance on top, and not all have done so yet, so it’s worth checking directly rather than guessing.
What’s the one question worth asking a vendor before anything else?
How, specifically, does a grade move from the AI’s output to being final. If there’s any path where that can happen without an assessor’s explicit approval, treat that as a compliance risk.
Sources
Ofqual (2026). Using AI in Marking: Why Technical Capability, Fairness, and Transparency All Matter. Ofqual blog, gov.uk.
Ofqual. Ofqual’s Approach to Regulating the Use of Artificial Intelligence in the Qualifications Sector. gov.uk.
JCQ. AI Use in Assessments: Protecting the Integrity of Qualifications.
Citron, D. K., & Pasquale, F. (2014). The Scored Society: Due Process for Automated Predictions. Washington Law Review, 89, 1-33.
See what compliant AI-assisted marking looks like
Every Eduface grade is a draft until an assessor explicitly approves it, with the reasoning and an audit trail. Book a demo or start free.