Universities in Australia are transforming generative AI from personal accounts set up by students themselves into public tools that the schools can manage, train on, and set boundaries for use. Public cooperation information with the University of Sydney indicates that ChatGPT Edu is being integrated into teaching, research, and campus operations. On the surface, this appears to be an expansion of campus software; however, a more profound change is that universities are beginning to rewrite what it is expected that students should know upon graduation.
From personal trials to public infrastructure
Students have been using AI in dormitories, libraries, and part-time jobs for a long time. What universities now need to accomplish is to transform a tool that is highly personalized, has uneven payment capabilities, and varies in usage methods into a public facility that comes with training, boundaries, and is auditable. Moving from individual subscriptions to unified purchases by the school not only means issuing additional accounts but also implies that the school must take responsibility for identities, data, courses, and usage records.
UNSW plans to expand ChatGPT Edu to include over 80,000 students and staff in phases. The University of Sydney, on the other hand, plans to open it to all students in the first semester of 2027, while establishing a AI sustainability research and innovation fund of 1 million Australian dollars. Once the scale shifts from a pilot account to a university-wide infrastructure, the issues are no longer just about whether the "model is useful" or not, but rather how permissions, data, computing power budgets, and teaching responsibilities can be managed together.
A change in the assessment logic was made with tables Can, Can, and Must.
UNSW adopts the frameworks of Can, Can, and Must. Some tasks can utilize AI, while others must be completed independently, and still others require students to proactively use AI. This tripartite approach may seem simple, but it is closer to real-world work than the general prohibition against cheating. In the future, doctors, lawyers, engineers, and teachers may all need to use AI, yet they still have to demonstrate that they have a solid grasp of basic knowledge and are able to identify errors in tools at critical points.
Therefore, the assessment has shifted from checking whether "the answers were written by one person" to examining whether "students can explain their reasoning process." An assignment that allows AI to participate may require the submission of original materials, verification steps, and revised suggestions; whereas an exam that prohibits AI is used to determine whether students truly possess the underlying skills necessary. The most noteworthy change is not that universities have adopted a certain chat tool, but rather that they have begun to include the choice of tool as part of the assessment of professional competence.
What is truly scarce are not the cue words.
When AI is able to provide a complete answer in a short time, the focus of teaching should not remain on whether it is “useful” or not, but rather shift to how students can verify, how they can make corrections, and how they can take responsibility for their final decisions. Models can serve as repeatable practice partners, but they cannot bear the consequences of professional judgments on behalf of students. This boundary is more important than the ability to “write prompts.”
This also explains why “being able to write prompts” is not the ultimate goal. True abilities include defining problems, selecting materials, discerning illusions, protecting privacy, understanding disciplinary norms, and having the courage to say “no” even when tools provide seemingly perfect answers. When everyone can generate decent written text, the ability to judge which statements cannot be trusted outright becomes an even rarer skill.
Large-scale opening up can also lead to new inequalities.
Centralized procurement can alleviate the disparities caused by individual subscriptions, but it does not automatically eliminate them. Whether different specialties can obtain cases suitable for them, whether teachers have the time to redo their courses, whether students with disabilities can use the materials smoothly, and whether schools record sensitive research data are all factors that can affect the actual benefits. A centralized platform may also lead to dependency on suppliers: once courses are designed around a particular set of tools, the cost of switching to different models, prices, or policies in the future will rise rapidly.
Therefore, what universities need is not just the issuance of accounts, but also multiple model options, exit pathways, data segmentation, and clear learning objectives. The public statement from the University of Sydney clearly states that accounts are managed through single sign-on by the school, and the input content is not used for training public models by default. It is also accompanied by identity management, encryption, audit logs, and configurable data retention. These are not merely accessories to product selling points; they are prerequisites for the success of any university-wide promotion.
The most interesting aspect of the experiment conducted by Australian universities is that it shifted the debate from “whether students can use AI” to “what students must prove when using AI.” Of course, the form of assignments will change, but a more profound transformation occurs in the standards of competence: future graduates will not be able to rely solely on models, nor can they pretend that models do not exist. What universities truly aim to produce are not people who are proficient at using certain products, but individuals who know when to invoke machines, when to stop, and who are willing to take responsibility for the final judgment.
