ACADEMIC INTEGRITY · AI
Academic Integrity in the Age of AI: Assessment Design That Holds Up
Detection tools are a losing game. Four assessment design principles that genuinely resist AI-generated cheating, and how Eduface supports integrity-resilient assessment.
By Eduface · July 2026 · 9 min read
Every lecturer knows a plausible-sounding essay does not prove a student understands the subject. That gap has always existed. Generative AI has simply made it wider, faster, and harder to ignore. Students can now produce a coherent 2,000-word essay on any topic in seconds. The question is no longer whether this is happening. It is what institutions should do about it.
How should institutions protect integrity?
The most durable response to AI-enabled cheating is not better detection: it is better assessment design. Tasks that require personal reflection, disciplinary application to real contexts, or oral defence are significantly harder to outsource to generative AI. Institutions that redesign assessment around these principles, and use tools that create iterative, personalised learning records, are better placed to uphold integrity than those relying on detection software alone.
Why is AI detection not the answer?
Tools such as Turnitin AI and GPTZero have become the default institutional response to AI-generated submissions. The logic is understandable. They are familiar, they slot into existing workflows, and they produce a percentage score that feels decisive. The problem is that the technology does not reliably deliver on that promise.
Detection tools produce false positives: students writing in a second language, students who write concisely, or students who follow a genre convention closely can all trigger AI flags without using generative tools at all. Those false positives carry real consequences for innocent students, including formal misconduct processes, damaged academic records, and the erosion of trust between lecturers and their cohorts.
There is also a race dynamic that detection tools cannot win. Generative AI models improve continuously. Each new generation produces output that is harder to distinguish from human writing. Any detection tool calibrated to today’s AI will be less effective against next year’s. This is not a software update problem. It is a structural one. The Quality Assurance Agency (QAA) has acknowledged that institutions have a responsibility to design assessments that uphold academic standards in the context of AI. Relying solely on detection does not meet that standard.
What assessment design principles resist AI-generated cheating?
The following four principles represent a coherent framework for integrity-resilient assessment. None of them require the elimination of written work. They require a shift in what written work is asked to do.
1. Personalisation
Tasks that require students to draw on their own placement experience, locally gathered data, or a specific case from their own professional context cannot be fully outsourced to AI. A generative model can write a strong essay about supply chain disruption in general. It cannot write accurately about the disruption a specific student witnessed during a placement at a named logistics firm last autumn. Personalisation does not need to be elaborate. It needs to be genuine.
2. Process visibility
Portfolios, annotated drafts, and reflective logs shift the assessment object from a single polished output to a visible thinking process. This matters because the process is difficult to fabricate. A student who submits a polished essay with no evidence of drafting, no engagement with feedback, and no trace of iteration is easier to query. A student whose portfolio shows progressive revision, engagement with specific comments, and reflective notes on their own development is demonstrating something that generative AI cannot simulate at scale.
3. Oral components
An oral examination, viva, or brief follow-up conversation requires the student to defend their own work in real time. This is not new: doctoral education has always included it. What is new is extending the principle more broadly across undergraduate and taught postgraduate assessment. A student who used generative AI to produce a submission will typically be unable to explain the specific reasoning, answer follow-up questions about their sources, or respond to disciplinary challenges that require genuine understanding. Oral components do not need to be long to be effective.
4. Criterion specificity
Generic rubrics reward surface-level content coverage, which is exactly what generative AI does well. Rubrics that reward disciplinary reasoning, methodological awareness, engagement with counterarguments, or the application of theory to unfamiliar cases require something AI cannot supply: actual understanding. Rebuilding rubrics around these criteria raises the floor for what a passing submission must demonstrate.
How does formative feedback reduce the incentive to cheat?
Research by Nicol and Macfarlane-Dick (2006) established that students who engage in iterative feedback cycles, receiving feedback on drafts and acting on it before final submission, develop deeper understanding and stronger self-regulation. The implications for integrity are direct.
When students receive detailed, personalised formative feedback at the draft stage, two things happen. First, the feedback itself is specific to their work: it references their argument, their evidence, their gaps. That feedback cannot be applied to a generically AI-generated submission without creating obvious inconsistencies. A final submission that ignores or contradicts the feedback it received is a signal worth investigating. Second, students who experience the feedback loop as genuinely useful have a stronger intrinsic reason to engage with the task themselves. Engagement is the single most reliable protection against academic misconduct.
Eduface’s Paper Grader and Feedback Tool delivers per-criterion written feedback on every submission, including at the draft stage. This creates a documented, personalised record of each student’s engagement with the task. In pilot institutions, Eduface feedback has achieved 95% alignment with lecturer assessments, which means the feedback students receive is credible and worth engaging with.
What role do oral examinations play?
The oral examination is the most direct available response to the question that AI-generated submissions raise: does the student actually understand what they submitted? A well-designed oral component does not require a student to memorise their submission. It requires them to reason from it. Contextualised follow-up questions, drawing on the specific arguments and evidence in the student’s own work, reveal whether the submitted text represents genuine understanding or a capable simulation of it.
Eduface’s Oral Examination tool, currently in early access with pilot institutions, generates contextualised follow-up questions derived from each student’s own submission. The AI then conducts an adaptive oral assessment, adjusting the line of questioning based on how the student responds. The result is an examination that cannot be prepared for in the abstract: the questions are specific to what that student wrote. Every assessment is confirmed by a lecturer, preserving human academic judgement throughout the process. This directly addresses Annex III and Article 14 of the EU AI Act (Regulation 2024/1689), which classifies AI used in student assessment as high-risk and mandates human oversight: the tool supports the examiner, not replaces them.
How does Eduface support integrity-resilient assessment?
Eduface is not a detection tool. It does not attempt to determine whether a submission was written by a human or an AI. It provides infrastructure for the kind of assessment that makes that question less critical. The Paper Grader and Feedback Tool creates an iterative, personalised record of each student’s engagement with their assessment. The Oral Examination tool adds a real-time comprehension check that is derived from the student’s own submission, making it impossible to prepare for using generative AI alone. Both tools operate with a human-in-the-loop: lecturers confirm every grade. Eduface is Jisc/CHEST approved and EU AI Act compliant. The goal is not to catch students using AI. It is to design assessment that rewards genuine understanding and makes academic misconduct less rewarding and more visible.
Assessment integrity resilience by type
Assessment type
AI risk
Resilience
Notes
Standard essay (closed topic)
High
Low
Easily outsourced to generative AI
Essay with personal/placement data
Medium
Medium
Requires student-specific content
Portfolio with reflective log
Low
High
Process visibility makes outsourcing visible
Oral examination
Very low
Very high
Requires real-time comprehension and defence
Written + oral follow-up (Eduface)
Very low
Very high
Combines written assessment with contextualised oral defence
Multiple choice
Low
High
Not generatable the same way; limited feedback value
Frequently asked questions
Can AI tools detect if a student used ChatGPT to write their essay?
Current AI detection tools, including Turnitin AI and GPTZero, can flag text that may be AI-generated, but they are not reliable. False positive rates are a documented concern, particularly for non-native English writers and students who write in formal academic register. No detection tool currently available can determine with certainty whether a submission is AI-generated. This is why assessment redesign, not detection, is the more defensible institutional strategy.
What assessment types are hardest to outsource to AI?
Oral examinations are the most difficult to outsource, because they require real-time comprehension and cannot be delegated. Portfolios with reflective logs are also highly resistant, because they make the thinking process visible and difficult to fabricate at scale. Tasks that require genuinely personalised content, such as placement reflections or locally gathered data, raise the bar further. The combination of written work with an oral follow-up, as supported by Eduface, provides the strongest available protection.
How does Eduface’s oral examination tool support integrity?
Eduface’s Oral Examination tool generates follow-up questions from each student’s own submission, then conducts an adaptive oral assessment based on how the student responds. Because the questions are derived from the specific arguments and evidence in that student’s work, the examination cannot be prepared for using generic AI tools. Every session is reviewed and confirmed by a lecturer. The tool is currently in early access with pilot institutions.
Does using AI for marking create its own integrity risks?
This is a legitimate concern. AI marking tools must be transparent about how they work, auditable, and subject to human oversight. Eduface addresses this through its human-in-the-loop model: AI assessment is always reviewed and confirmed by a lecturer before grades are finalised. Eduface is also EU AI Act compliant, which requires that high-risk AI systems in education operate with meaningful human oversight. Using AI to support assessment is not itself an integrity risk, provided the system is designed with appropriate accountability.
What does the QAA say about AI and academic integrity?
The Quality Assurance Agency has published guidance on AI and academic integrity, recognising that generative AI presents a significant challenge to traditional assessment models. The QAA’s position is that institutions have a responsibility to design assessments that genuinely uphold academic standards, not merely to update their policies or rely on technology-based detection. Assessment redesign, including the approaches outlined in this article, is consistent with what the QAA expects of quality-conscious institutions.
Conclusion
Detection tools are not a strategy. They are a stopgap that creates risk for innocent students while failing to keep pace with the technology they are designed to catch. The durable response to AI-generated academic misconduct is assessment that rewards what AI cannot replicate: genuine understanding, disciplinary reasoning, and the ability to defend one’s own thinking in real time. Eduface supports that shift. From formative feedback that creates a personalised engagement record, to oral examinations that test comprehension directly from the student’s own submission, Eduface gives institutions the tools to make integrity-resilient assessment practical at scale.
Build assessment that holds up
Request a demo or create a free lecturer account, and see how per-criterion feedback and oral examination work in practice.