Beyond "AI-proof": Generative AI and the end of traditional assessment practices

For decades, unsupervised coursework has let us assess students often and cheaply. Generative AI has ended that approach, and making every task "AI-proof" risks more over-assessment.

Online postgraduate education takes a common approach: we set many unsupervised tasks, trust that submitted work is the student's own, and treat the volume of submissions as evidence that learning outcomes have been met. Online programmes rely on this more heavily than any other mode of delivery, because almost all of their assessment takes place outside supervised campus-based conditions.

Generative Artificial Intelligence (GenAI) has ended the traditional approach. In this article, I review some of the sector responses and argue that searching for "AI-proof" assessment is the wrong goal. For time-poor online postgraduate taught (PGT) students in particular, it risks compounding the over-assessment problem I explored in an earlier article.

The scale of change

Adoption of GenAI has been rapid, particularly amongst students. The Higher Education Policy Institute (HEPI) and Kortext survey found that the proportion of undergraduates using GenAI for assessments rose from 53% to 88% in a single year (Freeman, 2025, reference). By the 2026 survey, 94% reported using GenAI to help with assessed work, and 12% reported including AI-generated text directly in submissions, up from 3% in 2024 (Stephenson & Armstrong, 2026, reference). Nearly two-thirds said assessment had changed significantly in response, while some described growing anxiety about being falsely accused of misconduct.

These surveys sample full-time undergraduate students, and evidence on online and PGT students is still lacking. However, many postgraduate professionals now use these tools daily at work and can be expected to be just as familiar with them.

Why detection is not a strategy

The instinctive response from institutions to GenAI has been to a) detect and b) deter. This follows the traditional academic approach, though it is not supported by the evidence. For instance, Weber-Wulff et al. (2023), reference tested a range of AI-text detection tools and found them neither accurate nor reliable, and the Quality Assurance Agency for Higher Education (QAA, 2023, reference) advised that the problem would only intensify as GenAI becomes embedded in everyday software. For technical reasons, detection tools are both unreliable and pedagogically counterproductive, leading to a false sense of assessment security and 'false positives', accusing innocent students of academic misconduct.

The problem with "AI-proof"

With the recognition that detection fails, attention turns to assessment design: make each task somewhat resistant to student use of AI. Liu and Bridgeman (2023), reference put the difficulty plainly, concluding that take-home assessments simply cannot be made AI-proof as the technology continues to advance. The expert group behind the guidance from Australia's Tertiary Education Quality and Standards Agency (TEQSA) reached a similar view, framing assessment reform around trustworthy judgements of learning rather than attempts to exclude AI (Lodge et al., 2023, reference).

There is also a structural problem. Corbin et al. (2025), reference distinguish between discursive changes, such as instructions telling students which AI uses are permitted, and structural changes to the assessment itself. Discursive approaches, including popular AI-use scales (Perkins et al., 2024, reference), communicate expectations but cannot enforce them in unsupervised settings.

From a governance perspective, the most overlooked consequences of GenAI are cumulative. In a modular programme, AI-proofing happens task by task, assessment by assessment, with practical application left to individual staff. Academic teams add process logs, reflective statements on AI use, declarations, and mini-vivas to each assessment. Each addition is reasonable in isolation. Summed across a programme of dozens of summative points, they substantially increase the burden on students who have the least time to absorb it.

Reduce volume first

The most useful recommendation in QAA's (2023), reference advice has been the least acted upon across the sector. The first desirable outcome it sets for any assessment review is to reduce assessment volume, removing items vulnerable to misuse and repurposing the time for other learning. Its second outcome is a shift towards synoptic assessment of programme-level outcomes, and its third is authentic assessment in real, often workplace, contexts.

The order matters. Redesigning every task to resist AI is challenging for programme teams and, on current evidence, unachievable. Removing tasks that never needed to be summative is faster, cheaper, and provides a better student experience.

The two-lane logic

The clearest framework for the redesign task comes from the University of Sydney's two-lane approach (Liu & Bridgeman, 2023, reference; University of Sydney, 2024, reference), which aligns with TEQSA's principles for assessment reform:

This provides a crucial insight for online PGT programmes: assessment does not need to be everywhere. It needs to be somewhere credible. A small number of well-designed and assured points, such as short vivas, an observed task, or a defended portfolio, can carry more evidential weight than coursework submitted only at the end of a task.

A framework for online programme teams

Drawing on these principles, I suggest online PGT teams sort their existing summative tasks into three categories:

1. Assure: Tasks that evidence a programme learning outcome at a meaningful point. Secure these properly, using remote-appropriate formats such as recorded vivas.
2. Open: Tasks whose real purpose is learning. Make them formative or low-stakes, permit AI, and teach students to use it critically.
3. Remove: Tasks that duplicate evidence already gathered elsewhere and are easily reproduced using GenAI tools.

Looking ahead

Online programmes cannot fall back on proctored, hall-style examinations, and should not desire to. QAA (2023), reference itself describes the on-campus retreat to handwritten examinations as regressive and inauthentic. The online answer is fewer secure moments, designed specifically for remote students, alongside an open lane where learning to work with AI is practice-based and authentic. This is not a lowering of standards but a redesign of the student experience to better meet future graduate needs. That raises a further question, which I will address in a forthcoming article: if we have fewer, more secure assessment points, what is it we want future graduates to be able to demonstrate?

References