Overview

GenAI has unsettled an assumption that well-formed academic writing signals a learner’s understanding. When a chatbot can write a plausible answer to many types and topics of questions, and when detection tools struggle to identify AI-generated text with any consistency¹, the link between a ‘good submission’ and ‘a learner’s own grasp of the work’ becomes less secure. We need to ask what assessment is for and what kinds of learner capability universities ought to recognise when the production of fluent writing is no longer a meaningful indicator on its own.

We propose a philosophical case for ‘AI-agnostic’ assessment design, where assessment remains valid and defensible, whether learners use genAI or not. In this framing, marks are associated with judgement in context: taking a stance, weighing explanations, justifying choices and making the limits of evidence explicit. Our core claim is that these are precisely the moves that differentiate human understanding from the pattern-matching of LLMs, especially where moral, methodological or situational justification is required. The point is to ensure that it is the learner’s accountable decision-making that is rewarded.

AI-agnostic assessment treats learning as something that develops over time and in dialogue, not something captured in a single artefact. Where learners must carry arguments through stages, respond to critique and explain why they changed their mind or refined a claim, their agency becomes easier to see and easier to assess. This matters across disciplines, because the content of a good justification differs by field, but the need to demonstrate how conclusions were reached is widely shared. Designing assessment that includes ambiguity and even contradictions, pushes learners into interpretation rather than reproduction, making it harder for any submission, AI-generated or not, to slide by on generic coherence alone.

This way of thinking changes what we mean by academic integrity and fairness - it becomes less about proving who typed words and more about making responsibility clear. One answer is to treat openness about AI use as normal academic practice rather than as a confession. Learners can be asked to state what tools they used, how they used them and what they did to change, check or reflect the outputs. This matters ethically because strict bans and unreliable detection can create uneven risk, with some learners more likely than others to be wrongly suspected². A focus on transparency, by contrast, gives everyone the same clear expectations and allows markers to assess what actually matters – whether the learner can explain the decisions they made, the evidence they relied on and the limits of their argument. In short, it shifts the important questions from ‘who wrote this’ to ‘who is taking responsibility for the judgements in it, and can they justify them?’

1. Weber-Wulff D, Anohina-Naumeca A, Bjelobaba S, et al. Testing of detection tools for AI-generated text. Int J Educ Integr. 2023;19(1):26.

2. Liang W, Yuksekgonul M, Mao Y, Wu E, Zou J. GPT detectors are biased against non-native English writers. Patterns (N Y). 2023;4(7):100779.

Speaker bio

Dr Donna Johnson is the Course Director for Postgraduate Biomedical Sciences at Leeds Beckett University, where she’s overseen MSc provision since 2015. She’s a Senior Fellow of the Higher Education Academy and a Chartered Science Teacher. Much of her work centres on curriculum design that promotes student agency and supports learners from a range of academic and professional backgrounds. She’s especially interested in equipping postgraduate students to think critically and work independently, while still offering structured, inclusive support. Her teaching draws heavily on authentic assessment and skills development, particularly in communication, applied research, and data interpretation. Alongside teaching, she’s led a series of research projects exploring learning innovation in biomedical science.