Skip to main content

Guide - regulation and AI marking

AI-assisted marking when a person must make the decision

In January 2026 Ofqual, the regulator for qualifications in England, said that AI cannot be the sole mechanism for awarding marks, because its rules require a human-based judgement in marking decisions. That does not rule out AI. It rules out AI deciding. This guide summarises what Ofqual said and sets out how to design AI-assisted marking so a trained person makes every decision and you can show that they did.

What Ofqual said

In a blog post on 14 January 2026, Jo Handford and Joanna Williamson set out Ofqual’s current position. The key sentence: "the use of AI as the sole mechanism for awarding marks does not comply with our current regulations." The reason given is that such use "does not meet our requirements for a human based judgement to be used in marking decisions." The same post describes AI as promising for quality assurance and marker training, and names three things that matter when AI is involved in marking: technical capability, fairness and transparency (Ofqual blog, 14 January 2026).

This guide summarises a public blog post. It is not legal or regulatory advice. If you are an awarding organisation, check your obligations against Ofqual’s conditions and guidance directly.

Assisting is allowed. Deciding is not.

The line Ofqual draws is between AI that helps a person mark and AI that awards the mark. A design that keeps you on the right side of that line has four properties:

  • A trained, named person makes the marking decision for each candidate, not a person who simply clicks approve on a batch.
  • The person sees the evidence behind each proposed mark, so they are judging the answer, not the AI’s number.
  • Nothing reaches the candidate until that person has released it.
  • Every change the person makes to a proposed mark is recorded, with who made it and when.

Technical capability: make the AI show its working

Ofqual’s post notes that AI systems "lack true semantic understanding and the capacity for human-like judgement." The practical answer is not to hide the AI’s reasoning but to expose it. Each proposed mark should name the criterion and level and quote the part of the answer it relied on. Where the AI is less sure, it should say so, so the assessor checks those marks first. A mark the assessor cannot trace back to the answer is a mark they cannot responsibly accept.

Fairness: one rubric, applied the same way

Ofqual’s post refers to ensuring criteria are applied consistently and fairly, regardless of who the candidate or the marker is. That starts with a written rubric with levels, used for every candidate. Mark recorded answers from a transcript so appearance and setting do not affect the score. Then measure: compare the AI’s proposed marks with your assessors’ final marks, criterion by criterion, and look for patterns in the overrides, such as one question or one assessor moving marks in the same direction every time.

Transparency: keep a record someone else can follow

"We must be able to understand how marks are determined," the post says. For every released result, keep: the rubric and its version, the mark the AI proposed and the evidence it cited, every change the assessor made and why, who released it and when. That record is what lets a moderator, an appeals panel or a regulator follow the decision after the fact. The audit trail guide lists what it should contain.

Where Scorafy fits

Scorafy is built around the same line. The AI drafts a mark for each criterion with the evidence quoted and a confidence figure, a person on your side reviews and overrides, and no score reaches a candidate until that person releases it. Every override is recorded, and each assessment has a one-click audit pack. Scorafy is not exam-board marking infrastructure for GCSEs or A levels. It suits training providers, awarding and certification bodies, and organisations certifying their own staff that want AI assistance with a person accountable for each result.

Common questions

Does Ofqual allow AI to mark regulated qualifications?
Not as the only marker. In January 2026 Ofqual wrote that "the use of AI as the sole mechanism for awarding marks does not comply with our current regulations", because its rules require a human-based judgement in marking decisions. It described AI as promising for quality assurance and marker training.
What does "human in the loop" need to mean to respect that position?
A trained person makes the marking decision, not just a person who can see the AI’s output. In practice: the AI drafts, a named assessor reviews each mark against the evidence, changes what they disagree with, and only then is the result released, with every change recorded.
Which uses of AI in marking are lowest risk?
Uses that support human markers rather than replace them: checking human marking for consistency, flagging scripts for a second look, and training and standardising markers. Ofqual named quality assurance and marker training as promising applications.
What should an AI marking tool show the assessor?
The criterion and level it proposes, the exact part of the answer it relied on, and how confident it is, so the assessor can check the reasoning rather than accept a number. Ofqual’s point on transparency was that it must be possible to understand how marks are determined.

“Both respondents and staff have found the Scorafy recommendations highly actionable, and the platform itself very easy to navigate and use.”

Product and Digital Lead, online education non-profit

See where the decision sits in Scorafy

The AI drafts each mark with its evidence. Your assessor reviews, overrides and releases. No score reaches a candidate before that, and every change is recorded.