Where AI should assist, but not own the judgement, in higher education

A university can use AI to help design a course, draft teaching material, organise assessment evidence or review an application. Those contributions can improve speed, consistency and the amount of material staff have time to consider. They do not transfer responsibility for the academic decision that follows.

What belongs in a course, whether a student has demonstrated the required learning, whether they pass, whether they are admitted, and whether misconduct occurred are judgements the provider later has to stand behind. AI can participate in the work that leads there. The institution still has to understand the basis of the outcome, inspect the evidence, be able to disagree with the system, explain what it did and correct it.

The current Higher Education Standards Framework already keeps that responsibility with the provider. It requires specified learning outcomes, assessment capable of confirming those outcomes, and grades that reflect student attainment. It requires course design to establish a coherent educational proposition. It requires academic oversight of teaching, quality and integrity, including critical evaluation of educational innovations. Those provisions do not say a model may never be used. They keep the provider responsible for what the institution then does.

Curriculum and the course that later has to be assured

Curriculum work is easy to confuse with document production, which is why generated outlines get treated as if they were already approved courses. AI can research comparable curricula, surface current literature, map proposed content against outcomes, draft learning activities, suggest assessment ideas and identify gaps. That can be a material improvement on assembling the same material from memory and a handful of familiar sources.

TEQSA's Course Design Guidance Note treats course design as more than the documents that record it. The design specifies what students should learn, how that learning will be assessed, what content and activities will get them there, and whether the sequence is coherent. The Guidance Note itself says it is not definitive or binding. Current Standard 3.1 still puts that specification on the provider.

Someone with academic responsibility still has to decide what belongs in the course, what level of knowledge is appropriate, which outcomes matter, whether the material is academically current, what disciplinary judgement the work requires, and whether the course remains coherent. A generated map can help that person see more of the field. It does not become the accountable curriculum owner.

Course material is a narrower case. Models are often useful for first drafts of examples, exercises, explanations, quizzes, cases and summaries. Generation is not the problem. Treating speed of production as academic assurance is. Generated teaching material can contain factual errors, invented references, outdated information, inappropriate examples, or subtle disciplinary mistakes that sit close enough to the truth to be published. A qualified academic can use that output as working material, in the same way they can use a colleague's draft, a publisher's chapter or an open educational resource. They do not have to type every sentence themselves. They remain responsible for the material presented as part of the course.

Teaching, and what students still have to demonstrate

AI can also support delivery: further explanation, practice activities, question answering, translation or accessibility support, and adaptive resources. TEQSA's June 2026 resource on assuring quality learning treats gen AI as already part of the environment and discusses evaluative judgement, critical thinking and ethical reasoning. An institutional response that simply keeps AI away from teaching and assessment would sit poorly with that direction, and with the workplaces graduates enter.

Some interactions still require an accountable academic or professional response. Interpreting ambiguous student performance, deciding how to handle an unusual learning difficulty, exercising pastoral judgement, or determining whether competence has been demonstrated in a particular context all depend on circumstances a model may not see. Students do not need a person in the room because a person cares. They need someone who can be asked to explain the response, and to change it when the context requires a different one.

Students have a version of this as well. If a model performs every difficult intellectual step, the submitted work may stop demonstrating the student's own learning. The provider still has to be able to say that the specified outcomes were achieved. Used deliberately, AI can sit inside the assessment: students critique generated output, identify errors, compare it with evidence, justify what they keep and improve what they discard. That can assess evaluative judgement rather than hide it. How to design those tasks is a separate piece of academic work.

Assessment the provider can defend

Marking is where institutions most often look for a prohibition that is not there. Software, including AI, can check objective answers, apply stated rules, organise a rubric, identify passages for review, draft feedback and surface inconsistencies across a cohort. Those uses can improve consistency and give staff more time for the work that actually needs them. As assessment becomes interpretive, professionally consequential, or central to confirming course learning outcomes, the institution still needs a process that can defend what the grade means.

Standard 1.4 does not say only a person may mark. It does require assessment capable of confirming outcomes, and grades that reflect attainment. If a model substantially shaped a grade, the provider still has to assure that judgement. A feedback draft that a marker can rewrite is one thing. A recorded result that affects progression or completion is another.

A final grade can change whether a student continues, completes, becomes eligible for an award, or meets a professional pathway. The qualification itself is awarded by the provider. Fully autonomous academic certification is a different proposition from a system that helps an assessor organise evidence or draft comments. Current sources do not say automated grading is unlawful. They do leave the provider responsible for being able to say that the required learning was demonstrated.

Academic work often involves expert evaluation that is not arbitrary just because it is not mechanical. The quality of an argument, clinical reasoning, design quality, originality, synthesis and research quality all require judgement. A model can offer another reading, apply an explicit rubric, compare passages or draft comments. When that evaluation determines attainment, the institution needs an accountable process that can stand behind the result. That process might involve one assessor, several markers, a moderation panel, or a combination of software and people.

A later signature is not that process if the reviewer cannot see the work, does not understand how the result was produced well enough for the task, has no authority to change it, or is processing more cases than anyone can actually inspect. If a student asks why they received a particular mark, "the model gave you 62" is not an institutional explanation. The institution has to be able to point to the task, the criteria, the evidence and the judgement that was made. Every model does not need to be mathematically explainable. The provider's own decision does need to remain intelligible.

Integrity, admissions and other decisions that change a student's position

Academic integrity already shows how a useful signal can be mistaken for a finding. TEQSA's Academic Integrity Toolkit currently says an AI-detector score alone is insufficient evidence to make a misconduct allegation, and that additional evidence is required. A detector can flag work for review. It should not become the adjudicator. Whether misconduct occurred still depends on other evidence, academic judgement, conversation with the student, institutional policy, procedural fairness and, if needed, review. Where AI is useful in Australian higher education discusses how to treat those signals in practice. Contributing evidence is not the same as owning the finding.

Admissions carries a different consequence. A model can extract an institution, qualification, dates, subjects, grades and supporting documents from an application, check completeness, map qualifications for review, highlight missing material and route files. That can remove a large amount of administrative work, including from international applications. Inferring a candidate's quality and ranking applicants on weights nobody has written down is a different design. One produces an inspectable intermediate result. The other starts making the admission decision. The same standard applies to international and domestic applicants. International files often simply contain more documents that need to be read.

TEQSA's Admissions Guidance Note makes clear that admissions frameworks must support students being appropriately prepared for their intended study. The provider owns that framework. A generated score should not quietly become the policy.

This is easy to miss in implementation. If an admissions team has a documented policy, and a system is then told to score each applicant from 1 to 100 from the file, the software is not automating the policy unless the institution has already defined what the score represents, which factors matter, how they are weighted, what exceptions exist and how the decision is reviewed. Otherwise the implementation has invented part of the rule. The same thing happens when a marking model invents a weighting the rubric did not contain, or when a progression model turns an unexplained combination of signals into an academic status.

Progression and intervention sit next to that. Institutions increasingly have data that can identify students who may need support: attendance changes, unusual assessment patterns, changes in engagement. Those signals can be useful for staff review. They describe part of a situation. They do not explain why it happened, and they should not, by themselves, decide exclusion, progression or academic status. A prediction can be a reason to look. It is not the judgement.

Appeals and complaints are a smaller instance of the same pattern. Summarising a large file, organising correspondence, finding the relevant policy and building a chronology can reduce administrative load. Deciding whether an academic appeal succeeds still depends on policy, evidence, context and procedural fairness. The summary can be checked before anyone acts on it. The determination cannot be treated as another generated paragraph.

Where those systems use personal information to make, or do a thing substantially and directly related to making, a decision that could reasonably be expected to significantly affect a person's rights or interests, APP entities that meet the statutory tests will have additional privacy-policy transparency obligations from 10 December 2026 under the Privacy and Other Legislation Amendment Act 2024. Human review does not automatically remove that obligation, and ranking or advisory outputs can still fall within the scope depending on the facts. That is a transparency requirement, not a ban, and this is not legal advice.

Academic governance, and what software must not invent

When AI begins to shape academic outcomes, the adoption is not only an IT purchase. Current Standard 6.3 requires accountable academic oversight of teaching, learning, research, academic quality and integrity. It also requires critical evaluation of educational innovations or proposed innovations. TEQSA did not write that clause about AI. AI in curriculum, teaching, assessment or admissions is still an educational change the institution has to evaluate.

The legislation does not say every AI system must go to Academic Board. Decision-making authority should stay with the governance structure that already owns the affected academic judgement: an assessor, an admissions officer, a review panel, an authorised delegate, or academic governance itself. A technology implementation should not redefine admissions policy, assessment practice or integrity process by the way it is built.

The 2026 Threshold Standards amendments add later-applying governance requirements around accountability, academic standards and independence, ethical and responsible operation, and risk management. They were not introduced as an AI code. They do reinforce that these systems sit inside existing governance and risk structures. Relevant application dates begin in 2027 and differ by provider type.

The Australian Framework for AI in Higher Education, published by ACSES, treats contestability and human review as design principles. It is not a Threshold Standard. It is useful because students may need to question an outcome, correct information, provide context or appeal. If AI contributed, that contribution should not make the institutional decision impossible to interrogate.

There is no particular virtue in having a person perform a task manually because the task is academic. A rules-based check, a well-reviewed draft, or a flag that points staff at the right file can be better than a tired person repeating the same clerical step. The objective is not human labour. It is the capacity of the institution to say what it decided, on what basis, and to change that decision when the evidence requires it.