Australian Universities Are Turning AI Access Into an Assessment Problem

University students are already experimenting with AI. The harder question is how institutions turn that informal use into a learning experience that is fair, useful and assessable. OpenAI’s November 5 account of collaborations with UNSW Sydney, the University of Sydney, La Trobe University and Adelaide University offers several approaches: staged tool access, applied projects, staff training and explicit rules about when AI is allowed or required. ^[1]

Access is only the first institutional decision

OpenAI reports that UNSW plans to bring ChatGPT Edu to more than 80,000 students and staff in stages and integrate AI foundations and Codex into some educational programs. The University of Sydney plans a staged rollout, with access for all students scheduled for Semester 1, 2027 after a trial. La Trobe says staff have access already and plans broader student access in 2027; its curriculum plans include an AI Engineering degree and practical industry work. Adelaide University describes researchers and students using its API and authorized defensive-security tools. These are institution-specific plans and company-reported examples, not a claim that every student already has access. ^[1]

The same OpenAI post cites a survey conducted by Pureprofile on its behalf. It surveyed 1,018 Australian university students aged 18 or older from August 3 to 7, 2026. OpenAI reports that 81% used AI several times a week and 79% wanted more opportunities to use advanced tools on real-world projects. Because the study was commissioned by a provider selling the tools discussed, the sponsor and questionnaire context should stay visible when interpreting the results. The survey records stated behavior and preferences; it does not test learning outcomes or establish that a specific deployment improves them. ^[1]

Assessment rules make the boundary legible

UNSW’s “Can, Can’t, Must” framework gives instructors three clear choices: permit AI when it supports the goal, prohibit it when independent demonstration is essential, or require it when learning to use the tool is itself part of the task. This is more actionable than a single institution-wide rule because the relevant skill changes by assignment. A student may need to calculate a result unaided in one course, then use a model and audit its work in another. ^[1]

UNESCO’s guidance for generative AI in education similarly emphasizes human agency, data protection, age-appropriate use and pedagogical validation. A university therefore needs more than a license or an acceptable-use memo. It needs to define what competence looks like, train staff to design assessments around that competence, and explain how student data is handled. ^[2]

Practical learning can make the policy visible. A paramedicine student might use generated scenarios to practice triage, but still explain the evidence behind a decision. A trainee teacher might rehearse a difficult conversation with a simulation, then receive feedback on judgment and communication. In each case, the AI helps create practice opportunities; it does not replace the student’s responsibility to understand the subject.

A well-designed assignment should tell students what evidence of learning looks like. If the objective is to interpret a clinical result, AI can generate a practice case, but the student should still explain the basis for the diagnosis. If the objective is to learn coding, students might submit both the result and a short account of how they checked model-suggested code. If the objective is unaided recall, the instructor can explicitly prohibit AI and use a supervised task. The rule should track the skill being assessed, rather than treating every tool interaction as either acceptable or cheating.

Measure learning, not logins

Sydney’s student guidance describes a phased approach: a pilot before the planned Semester 1, 2027 student release, with training and academic-integrity guidance alongside access. That matters because institutions can change the rollout after a pilot, based on support demand, educator feedback and student experience. A timetable is a plan, not evidence that a rollout has already reached every student. ^[3]

A well-designed assignment should tell students what evidence of learning looks like. If the objective is to interpret a clinical result, AI can generate a practice case, but the student should still explain the basis for the diagnosis. If the objective is to learn coding, students might submit both the result and a short account of how they checked model-suggested code. If the objective is unaided recall, the instructor can explicitly prohibit AI and use a supervised task. The rule should track the skill being assessed, rather than treating every tool interaction as either acceptable or cheating.

For universities, useful indicators include whether students can identify errors, explain when they chose not to use AI, transfer a skill to a supervised setting, and demonstrate baseline knowledge independently. Access rates and usage frequency can show adoption, but they do not prove competence. Programs should also review who has reliable access, whether staff have time to redesign courses, and whether AI-generated scenarios are accurate enough for the field.

The Australian examples point to a sensible sequence: provide supported access, give students real tasks, train educators, set assignment-level rules, and evaluate learning rather than activity. The survey helps explain student interest; institutional evidence will determine whether the approach prepares graduates well. Universities should publish that evidence as the programs mature, including where tools are ineffective and where independent work remains essential.