COMPLIANCE · OFQUAL 2026
AI in Assessment: What Ofqual’s 2026 Guidance Means for Awarding Organisations and Centres
Ofqual ran workshops with awarding organisations through 2025 and 2026 to find out what they were actually doing with AI. The picture is more textured than one rule about marking.
By Eduface · July 2026 · 11 min read
Most regulatory guidance gets written the way most contracts get written: cautiously, generically, designed to cover every case at once. Ofqual’s engagement with awarding organisations on AI has actually been the opposite. Through 2025 and into 2026, the regulator ran a series of workshops with awarding organisations specifically to find out what they were doing with AI in practice, not what they might theoretically want to do, and to work out whether the existing rules were getting in the way. That’s a genuinely useful exercise, and the picture it produced is more textured than a single rule about marking. If you sit between an awarding organisation and your learners, the detail matters.
The short version
Awarding organisations are cautiously exploring AI, and Ofqual’s framework leaves room for each to set its own rules. Appetite is strongest for generating assessment material and mapping qualifications against frameworks. Marking gets more caution, because a marking error changes a learner’s result directly and immediately. Expect real variation between awarding organisations, and don’t read silence as permission.
Cautious interest, not resistance
The tone that comes through from Ofqual’s account of these workshops is worth sitting with. Awarding organisations weren’t dragging their feet on AI out of institutional conservatism. They were, in Ofqual’s own description, showing a cautious but genuinely growing interest, and actively exploring where it could help. That’s a different starting point than a lot of commentary on this topic assumes. The caution isn’t coming from people who don’t understand the technology. It’s coming from people who understand the stakes.
Where the appetite is strongest, and where it isn’t
Three priority use cases came out of that engagement clearly enough to be worth naming individually.
Item and assessment generation, using AI to help draft exam questions and assessment material, with a human editor in the loop before anything gets used.
Stimulus material generation, producing supporting content like images, audio, or video that accompanies an assessment.
Mapping and classification, using AI to map qualifications against frameworks and standards, which is largely administrative work.
Marking sits apart from all three of these, and it’s worth asking why. The honest answer is that generating a practice question or classifying a qualification against a framework doesn’t change an individual learner’s result if something goes slightly wrong. A marking error does, directly and immediately. Ofqual’s caution scales with the stakes, which is a defensible position rather than a bureaucratic reflex, and it echoes a point that recurs throughout the human factors literature on automation more broadly: the acceptable error tolerance for a decision support tool depends entirely on what happens when it’s wrong (Parasuraman & Manzey, 2010).
What awarding organisations themselves are worried about
It’s worth being clear that the caution here isn’t purely regulatory box ticking imposed from outside. In the same workshops, awarding organisations identified their own risks and constraints around AI use, centred on the tension between validity, reliability, and fairness on one side, and the genuine appeal of speed and lower cost on the other. Ofqual’s summary of that engagement notes something else worth flagging: awarding organisations mostly weren’t asking for fewer rules. They were asking for clearer guidance and more confidence in how to proceed within the rules that already exist. That’s a meaningfully different ask than get out of our way, and it should shape how a training provider reads the whole area. This isn’t a fight between innovation and regulation. It’s a shared attempt to work out where the actual risk sits.
What this means if you work with more than one awarding organisation
Expect real variation between awarding organisations, and expect it to stay that way for a while. Ofqual’s framework leaves room for individual awarding organisations to set their own detailed policies, as long as those policies are communicated clearly to the centres delivering their qualifications. A rule that applies to one qualification you deliver will not automatically apply to another, even if both sit under the same broad AI conversation.
Don’t read silence as permission. Because differences between awarding organisations are expected and legitimate, the absence of a published position from yours doesn’t tell you anything either way. Ask directly, and get the answer in writing if you can, before you build a workflow around an assumption that might not hold.
Treat this as evolving rather than settled. Ofqual has said explicitly that it plans to keep gathering evidence, including evidence on how AI related malpractice is actually handled in practice, and to introduce further guidance as that evidence base grows. A policy read in early 2026 is a snapshot of where things stood then, not a permanent rule.
Keep the low stakes and high stakes uses genuinely separate in your own thinking. If you’re experimenting with AI for things like drafting practice material, you’re working in the part of this landscape where everyone, Ofqual included, is most comfortable. Marking is a different category, and it deserves the more careful standard described in our companion piece on human in the loop marking.
The commercial argument for getting ahead of this
There’s a temptation to treat all of this purely as compliance homework, something to tick off once and move on from. That undersells it. An awarding organisation that gets its AI policy wrong in either direction, too permissive and it ends up dealing with a fairness or malpractice incident, or too restrictive and it pushes centres toward tools with no real oversight at all, ends up rewriting that policy later under pressure, usually after something has already gone wrong. Getting a clear-eyed read on where things stand now, and where they’re likely to move next, is cheaper than that correction.
Frequently asked questions
Does Ofqual’s guidance apply the same way to every qualification and every awarding organisation?
No. Ofqual sets the overall regulatory principles, but individual awarding organisations have room to set their own specific rules within that framework, and real differences between them are expected. Always check directly with the ones you work with.
What has Ofqual said it will do next?
It has committed to continuing to gather evidence on how AI is used, and misused, in assessment, and to introduce further guidance as that evidence base develops. Current guidance is a current position, not a fixed endpoint.
Which AI use cases are awarding organisations most comfortable with right now?
Generating assessment material, generating supporting content, and mapping qualifications against frameworks. Marking gets more caution because the stakes for an individual learner are direct and immediate.
What should a training provider do if its awarding organisation hasn’t published a position on AI yet?
Ask directly, in writing, rather than assuming either that it’s allowed or that it isn’t.
Sources
Ofqual. Ofqual’s Approach to Regulating the Use of Artificial Intelligence in the Qualifications Sector. gov.uk.
Ofqual (2026). Using AI in Marking: Why Technical Capability, Fairness, and Transparency All Matter. Ofqual blog, gov.uk.
Parasuraman, R., & Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation: An Attentional Integration. Human Factors, 52(3), 381-410.
Deploy AI in the part of assessment everyone trusts
Eduface keeps a human decision on every grade, with an audit trail your awarding organisation can inspect. Book a demo or start free.