Reflections on today’s Office for Students and Advance HE report on generative AI, assessment and evidence of learning.
Today’s Office for Students and Advance HE report confirms what many academics already suspect. A significant share of university assessment can no longer prove what students know.
Based on conversations with more than 300 students and staff across over 120 higher education institutions, the research paints a stark picture of the challenge generative AI poses to assessment. Many unsupervised assessments can be completed to a high standard with AI, sometimes in minutes. Detection cannot reliably catch it.
Better detection won’t fix this. Better design will. Build learning around the moments where the human must stay visible: thinking, judgement, feedback, reflection and growth.
That means looking beyond the final artefact and making the learning process itself more visible. Because in the age of AI, the essay, assignment or finished product can no longer be expected to carry the full burden of proving what a student knows and can do. One essay can be faked. Three years of connected evidence is a different matter.
What the report found
Students were candid. Some told researchers that peers wrote whole essays with AI in minutes and got the same grade as those who did the work. One member of staff warned that “a number of our assessments may be unfit for purpose”.
Detection is not the fix. The report finds AI detectors do not reliably identify AI use. It warns against using misconduct processes to “retrospectively fix” assessments that are not resilient to AI. Many students said they would not declare even permitted AI use, for fear of suspicion.
Policy alone will not solve it. Without structural change, the report says, it is not possible to verify whether students followed the rules.
The real risk is separation
The report names one feature of generative AI that matters more than the rest. The final product can become separated from the learner’s own knowledge, reasoning and capability.
That is the heart of it. If we only assess the product, we cannot know who did the learning.
Process, not just product
The report points to secure assessments such as exams and vivas. They have their place. But it is honest about their limits. They are resource-intensive. They can disadvantage disabled students. Closed-book exams bear little resemblance to professional work.
It also highlights process-based assessment. Checking how students produce their work at multiple stage gates, not just what they hand in.
This is where the sector should look harder. Those stage gates create opportunities to make human involvement visible. When learning is captured as it happens, across modules and whole programmes, evidence builds over time. Drafts, reflection, feedback, practice and observed performance sit together. Growth becomes visible. AI can produce an essay. It can’t fake three years of growth.
Three kinds of evidence. One record.
At PebblePad we call this connected evidence. The same record serves three purposes.
- Evidence of learning. The stage gates the report calls for, captured across a whole programme. Longitudinal evidence of the process, not just the product.
- Evidence of readiness. Students told the researchers that AI skills matter for employability. Placements, work-integrated and experiential learning show what a graduate can actually do.
- Evidence of quality. For accredited and regulated programmes, connected evidence can demonstrate skills and competencies against defined frameworks, providing a stronger basis for institutions, professional bodies and regulators to have confidence in what graduates can actually do. In nursing, medicine, engineering and teaching, this isn’t an abstract question. Institutions need credible evidence that graduates are ready to practise.
The report notes that professional bodies can limit how quickly assessment changes. Connected evidence works within those frameworks, not against them.
Better evidence, not better detection
The report calls for strategic leadership. For deliberate choices about where students use AI. For assessment that assures achievement while staying authentic.
Even the detection industry now admits detection is not the endgame. This morning on BBC Radio 5 Live, a leading AI detection provider argued that the focus needs to shift from whether a student used AI towards how they used it.
The sector has spent years asking the wrong question: can we catch AI? The right question is simpler. Did learning happen, and can we prove it?
We need to be very clear about where humans must own the learning process and let technology take care of the rest. Think about the development and feedback loop: draft, notice, reflect, amend, submit, receive feedback, reflect, improve. These aren’t simply steps towards an assessment. They are evidence of learning.
The more deliberately those moments are scaffolded and evidenced, the more confident institutions can be. And the more granular the evidence gathered along the way, the less we have to rely on a single final artefact to tell us whether learning took place.
This isn’t about rejecting AI. It’s about ensuring technology supports learning without removing the human thinking, judgement and reflection we’re actually trying to develop.
The answer is not better detection. It is better evidence, designed into the learning process. Stop policing the essay. Start evidencing the learning.
