Can universities still prove students are learning when AI can do the assignment?

As generative artificial intelligence (GenAI) becomes increasingly capable of producing essays, solving problems and completing university-style assessments, educators face a fundamental challenge in determining how they can be confident that a student has actually learned. 

New research from Southern Cross University provides a real-world example of what can happen when assessment is redesigned for a world in which AI is already available to students. 

The study, published in the International Journal of STEM Education, examined how 36 engineering and pre-service teaching students used GenAI across three assessments in a second-year physics unit after changes to the University’s approach to AI use. 

Rather than predominantly using AI to complete assessment tasks for them, students reported increasingly using it to check their work, ask questions, clarify concepts and receive feedback. 

Southern Cross University Senior Lecturer in Teaching and Learning Dr Zachery Quince said the findings point to a bigger issue facing universities. 

“The old practice of the final artefact being the demonstration of knowledge is no longer enough,” Dr Quince says. 

“If AI can reproduce the artefact, we need other ways of seeing the student’s thinking, judgement and engagement along the way.” 

What happens when university students are allowed to use generative AI? 

The study examined three assessments in a physics unit taken by engineering and education students. 

Students initially worked under conditions allowing GenAI for specific purposes, before later assessments permitted more open use with students required to declare the tools they used and how they used them. 

The researchers found students initially used AI largely for learning and formatting tasks. Across later assessments, its use increasingly shifted towards feedback and verification. 

Students used it to sense-check answers, question their reasoning and seek explanations – behaving more like someone consulting a tutor or peer than simply asking AI to complete an assessment. 

Reported use of AI to simply “do the task” remained comparatively low. 

Dr Quince said clear expectations were an important part of the result. 

Across close to 120 assessment submissions, he says there were only one or two instances in which students appeared to have over-relied on AI. Because expectations had been clearly scaffolded, those cases could be addressed through a conversation with the student rather than immediately escalating to an academic misconduct process. 

The research found no statistically significant difference in outcomes between students who reported using GenAI and those who did not, indicating that the study does not show that AI improves student marks. 

Instead, its significance lies in what it reveals about how assessment design can influence students' engagement with the technology. 

How can universities design assessments for the age of AI? 

Southern Cross University has progressively changed its approach to GenAI as the technology has evolved. 

Assessments now use clear descriptors to establish how AI may be used, while the University's Assessment Adaptation Model – GenAI helps academics examine whether an existing task still provides meaningful evidence of learning when generative AI is available. 

Both the University's GenAI policy evolution and its assessment adaptation model have been published by the Tertiary Education Quality and Standards Agency as case studies for the Australian higher education sector. 

For Dr Quince, the goal is not to design assessments to “beat” AI, but to ensure students still demonstrate the knowledge, reasoning and professional judgement their qualification requires. 

That may mean gathering evidence of learning through drafts, reflection, peer engagement and questioning, rather than relying on a final submission alone. 

“The question is becoming less about whether a student has used AI and more about whether we have enough evidence that they understand what they have submitted and can exercise judgement over it,” Dr Quince says. 

Do students want universities to teach them how to use AI? 

A separate study led by Dr Quince, involving 48 engineering students across three Australian universities, suggests that students themselves recognise the tension. 

Students described GenAI as an on-demand tutor and productivity tool, but also warned that over-reliance could weaken critical thinking, independent problem-solving and creativity. 

They expected routine elements of engineering work to become easier to automate, while human skills, including judgement, communication, collaboration, accountability and decision-making under constraints, would become increasingly important. 

Rather than advocating blanket bans, students called for clearer guidance on appropriate AI use and assessment that prioritised reasoning, verification and quality assurance. 

Dr Quince says the challenge is preparing students to use AI in ways that reflect the very different expectations they will encounter across industries and workplaces. 

“There isn’t one model of AI literacy that will work for every graduate because every profession and employer is using AI differently,” he says. 

“What matters is that students develop the critical thinking, evaluative judgement and agency to know when to use AI, how to assess what it gives them and when to reject it. The human still needs to remain in control of the thinking.” 

Can generative AI also improve teaching? 

Southern Cross University researchers are also exploring how GenAI can support teaching without replacing academic judgement. 

In a separate practice report, Dr Quince trialled a multimodal feedback approach combining personalised video commentary, assessment rubrics and AI-generated written summaries. 

The academic still determined the grade and provided the original feedback. GenAI was used only to help convert that feedback into an additional concise format, with every AI-generated summary reviewed before it was returned to the student. 

The work provides another example of what Dr Quince describes as using “GenAI for good”, applying AI to clearly defined tasks such as improving feedback, accessibility or efficiency, while keeping academic judgment and decision-making with the human. 

What comes next for AI and university assessment? 

The challenge is becoming increasingly urgent as generative AI systems become more capable. 

Universities are now trying to do two things at once – prepare students to use AI effectively in the workplace, while still being confident they have genuinely learned the knowledge and skills their qualification represents. 

That means moving beyond a focus on detecting inappropriate AI use and thinking more carefully about how assessment can build capability, judgement and independence. 

The physics study was small and based on students' self-reported AI use, and the researchers say larger studies across different disciplines are needed. 

But Dr Quince says universities cannot wait for the technology to stop changing before adapting. 

The next stage is to continue developing ways to assure learning across whole courses and degrees, rather than expecting a single piece of assessment to carry all the evidence. 

That approach is already shaping course design at Southern Cross University, including the Bachelor of Information Technology, where AI is woven through the curriculum, so graduates develop the skills industry is calling for, including using AI tools critically and exercising professional judgement over what they produce. 

“We’re trying to find the right balance between policy, practice and pedagogy,” says Dr Quince. 

“AI is going to keep changing. Our job is to make sure students remain the ones doing the thinking.”