In the scramble to AI-proof assessment, fluency isn’t competence

Higher education is caught in a high-stakes scramble to protect academic integrity.

As generative AI makes traditional written assignments ever harder to verify, universities across the UK are orchestrating a rapid return to human, real-time evaluation.

Viva voce examinations, live classroom presentations, and spontaneous oral questioning are fast becoming the new gold standard for authentic assessment.

The logic behind the shift is understandable. Testing a student’s ability to defend an argument in real time offers a compelling defence against automated text generation. But in our haste to outsmart the algorithm, we risk stumbling into a severe accessibility blind spot.

Replacing written coursework with live verbal performance, without fundamentally reforming how oral tasks are designed and marked, risks systematically penalising students who stammer, as well as neurodivergent learners and those with other speech disfluencies.

In a 2026 paper, Oliver Irvine and I explored how support for stammering in higher education remains poorly communicated and inconsistently implemented.

If universities don’t address this now, our response to generative AI will inadvertently entrench structural inequality at the heart of university assessment.

The fluency fallacy

The problem lies in an assumption embedded deep within university assessment cultures – that spoken fluency is a neutral proxy for academic competence.

Oral assessment rubrics frequently conflate smooth delivery, rapid pacing, and unhesitating answers with subject mastery. Phrases like “confident delivery”, “articulate presentation”, and “prompt and concise response” are common in traditional marking schemes and in the feedback assessors give.

For the roughly one per cent of adults who stammer, a silent block or a long pause reflects a physical, neurological difference in speech production rather than cognitive hesitation, lack of preparation, or a hidden AI script.

When assessors equate conversational speed with domain knowledge, dysfluent students are marked down for how their vocal cords function under pressure rather than for what they know.

The problem is compounded by the sector’s reliance on superficial adjustments. Granting a student an extra 25 per cent of time in a live presentation slot does nothing to resolve a silent speech block or ease the pressure of real-time performance.

But waiving oral assessment altogether is just as flawed – it deprives students of the chance to build communication skills and signals lower expectations.

As universities expand oral assessment to safeguard academic standards, continuing to measure competence through the lens of normative speech delivery simply isn’t tenable.

Fixes for sector leaders

Preserving academic integrity in the age of AI doesn’t require high-anxiety, single-format oral assessments. By translating recommendations from national advocacy groups like STAMMA alongside emerging pedagogical research, we can build assessment frameworks that are both resilient against AI and equitable for all.

As our research highlights, support for speech dysfluency often exists in isolation or is buried deep within general disability services. Universities must formally adopt recognised guidelines such as STAMMA’s higher education resources, and actively signpost them in student-facing information and staff guidance. Visibility is the first step toward systemic change.

Universities must audit their marking rubrics to ensure core learning outcomes are decoupled from speech delivery. Unless a programme explicitly assesses public speaking as an accredited professional standard – as in the performing arts – marking criteria must focus strictly on the quality of reasoning, evidence, and critical analysis. Removing terms like “polished delivery” or “unhesitating response” costs nothing, yet immediately removes a systemic bias.

Our paper argues that as instructional practice evolves, technology-enhanced learning offers powerful routes to inclusion. Rather than requiring every student to deliver a single live presentation, universities can normalise asynchronous and hybrid formats.

Pre-recorded video presentations paired with short written or asynchronous Q&A exchanges let dysfluent students communicate without the mounting pressure of a live audience, while keeping the AI-proofing benefits of authentic questioning.

Accessibility can’t be treated solely as a student-facing issue. A genuinely inclusive academic culture recognises that speech dysfluencies affect staff as well as students, and designing inclusive oral practices and professional development frameworks means academics, researchers, and professional services staff who stammer are equally supported to thrive at work.

A job for executive teams

This is an institutional governance issue as much as a pedagogical one.

As executive teams update assessment strategies to address AI, pro-vice-chancellors for education and heads of quality assurance must make sure accessibility is built into those frameworks from the outset rather than patched on later.

The challenge posed by generative AI gives higher education a rare opportunity to rethink how it evaluates learning. Academic rigour lies in assessing the depth and authenticity of a student’s understanding rather than in watching them perform fluently under pressure. By modernising oral assessment now, universities can protect academic integrity while building an assessment culture where all students and staff feel supported to fulfil their potential.