Introduction
The public release of ChatGPT in November 2022 confronted Australian universities with a problem that moved faster than any policy cycle. A free, conversational tool could suddenly produce fluent essays, reports, reflections and code within seconds, tailored to almost any brief a marker might set. The immediate institutional responses, a mixture of bans, hurried policy revisions and renewed faith in detection software, treated generative artificial intelligence (AI) as a compliance threat to be contained. The deeper difficulty is structural. Unsupervised written assessment has long served as the sector’s principal proxy for learning, and generative AI has severed the assumed connection between fluent text and demonstrated capability (Lodge et al., 2023). A take-home essay no longer certifies what it once appeared to certify.
This essay argues that Australian universities should resist two tempting settlements: policing the problem away through detection, and retreating wholesale into the examination hall. Assessment should instead land on the program-level architecture now encouraged by the national regulator, the Tertiary Education Quality and Standards Agency (TEQSA): a deliberately limited secured lane of supervised, interactive assessment that verifies learning outcomes at critical gateways, and a redesigned open lane in which working with AI is expected, scaffolded and assessed, with AI literacy recognised as a graduate outcome in its own right. The argument proceeds by situating the integrity shock within Australia’s regulatory framework, examining why detection cannot bear the weight placed upon it, evaluating three redesign directions, and defending this position against the strongest counterargument, a general return to invigilated examinations.
The Integrity Shock in Its Regulatory Context
Australia entered the generative AI era with an unusually mature academic integrity apparatus, built largely in response to contract cheating. Bretag et al. (2019) surveyed more than 14,000 students across eight Australian universities and found that a small but persistent minority outsourced assessable work, often through commercial services. The Commonwealth responded with the Tertiary Education Quality and Standards Agency Amendment (Prohibiting Academic Cheating Services) Act 2020 (Cth), which criminalised the provision and advertising of cheating services and empowered the regulator to seek the blocking of offending websites. That regime presumed an identifiable supplier, a transaction, and a boundary between student and ghostwriter that an investigator could, at least in principle, expose.
Generative AI dissolves each of those assumptions. There is no supplier to prosecute, no payment trail to follow and no stable boundary between legitimate and illegitimate use, because the same tool that can ghostwrite an essay can also, quite defensibly, explain a difficult concept, summarise readings or tidy expression. The regulatory centre of gravity therefore shifts from misconduct enforcement to assessment validity. The Higher Education Standards Framework (Threshold Standards) 2021 (Cth) requires providers to use assessment methods capable of confirming that learning outcomes have been attained; where AI can complete a task undetectably, that confirmation quietly fails even when no misconduct occurs. TEQSA (2023) has accordingly framed generative AI as an assessment design problem rather than a policing problem, and has since asked every registered provider to demonstrate a credible institutional response, publishing examples of emerging good practice from across the sector (TEQSA, 2024).
The stakes extend well beyond pedagogy. International education generated close to A$50 billion in export income in 2023-24 (Australian Bureau of Statistics, 2024), and the value of that export rests on confidence in qualifications issued under the Australian Qualifications Framework. If employers, professional accreditation bodies or overseas governments come to doubt that an Australian degree certifies genuine capability, the damage would be economic as well as educational. Assessment integrity is, in this sense, a piece of national infrastructure.
TEQSA and the Two-Lane Settlement
TEQSA’s (2023) discussion paper, developed with leading Australian assessment researchers, advances two propositions: that universities must equip students to participate ethically in a society and workforce where AI is pervasive, and that judgements about attainment require multiple, inclusive and contextualised points of trustworthy evidence across a program of study. Neither proposition can be satisfied by treating every AI-assisted submission as suspect. Read together, they imply that some assessment must be secured, some must be opened, and the two must be deliberately orchestrated rather than left to the discretion of individual unit coordinators.
The clearest operational translation is the two-lane approach developed at the University of Sydney (Liu & Bridgeman, 2023). Lane one comprises secured assessment, undertaken in supervised or otherwise authenticated conditions such as invigilated examinations, in-class tasks, practical demonstrations and interactive orals, and exists to assure that graduates actually hold the outcomes their testamur claims. Lane two comprises open assessment, in which AI use is permitted and frequently required, and exists to develop and evaluate the capabilities students will exercise in real workplaces. The lanes are not interchangeable options on a menu; they perform different epistemic work, assurance in one case and education in the other.
The settlement succeeds or fails at program level. If every unit independently bolts on a viva or an invigilated test, staff workload and student assessment load both balloon while the program as a whole remains incoherent. Mapped across a whole degree, however, secured points can be few, deep and strategically placed at gateways such as capstones and professional milestones, with open tasks carrying the formative load between them. Sector practice is already converging on this architecture (TEQSA, 2024). Its underlying premise, that the open lane cannot be policed into integrity, is precisely what the evidence on detection confirms.
The Limits of Detection and the Equity of False Positives
Detection software promises to restore the old settlement, and it fails on its own terms. In the most comprehensive independent evaluation to date, Weber-Wulff et al. (2023) tested 14 detection tools and concluded that none was sufficiently accurate or reliable to ground misconduct decisions: overall accuracy fell below 80 per cent, results were inconsistent across text types, and performance degraded sharply once machine-generated text was lightly paraphrased or translated. Detection is also an arms race in which the defender’s methods are public, the attacker’s iterations are free and every new model generation resets the contest (Dawson, 2021).
The deeper objection is distributive, because false positives do not fall randomly. Liang et al. (2023) found that widely used detectors misclassified more than half of authentic essays written by non-native English speakers as machine-generated, while judging native-speaker prose almost flawlessly, since the lower lexical variability typical of writing in an additional language reads to a detector as a machine signature. In a sector educating hundreds of thousands of international students, this is not a marginal defect. It concentrates accusatorial risk on the cohort with the most to lose, for whom an adverse misconduct finding can jeopardise not only enrolment but also visa status. An integrity system that structurally over-suspects one group of students is itself an equity failure.
There is also a problem of procedural fairness. Australian universities must give students a fair opportunity to answer allegations against them, yet a detector outputs an opaque probability score that neither party can interrogate: the student cannot prove a negative, and the decision-maker cannot inspect the reasoning behind the number. Evidence of that kind cannot responsibly satisfy the balance of probabilities on its own. At most, detection can operate as a triage signal that prompts a conversation about drafts, notes and process. Integrity must therefore be designed into assessment before submission rather than adjudicated after it.
Redesign Directions
If neither detection nor denial will hold, the productive question becomes what assessment worth securing, and worth opening, actually looks like. Three directions dominate current Australian practice and scholarship: authentic assessment, interactive oral assessment and programmatic assessment. Each carries real value, and each is conditional.
Authentic assessment
Villarroel et al. (2018) define authentic assessment through realism, cognitive challenge and evaluative judgement: tasks mirror the problems, audiences and standards of professional practice rather than rehearsing academic genres for their own sake. The educational case is strong, since such tasks align assessment with the outcomes degrees claim to develop and are consistently associated with deeper engagement and employability. The security case, however, is routinely overstated. Workplace genres such as briefing papers, client communications and project plans are precisely the text types generative AI produces most convincingly, and authenticity by itself offers no assurance of authorship (Dawson, 2021). Authentic design therefore belongs primarily in the open lane, as a validity and employability strategy, not as an AI-proofing device.
Interactive oral assessment
Where authorship must be verified, dialogue outperforms surveillance. Sotiriadou et al. (2020) report Australian evidence that interactive orals, unscripted professional conversations anchored in work the student has already submitted, reliably reveal whether understanding sits behind a text while simultaneously developing the communication skills employers value. Orals scale more readily than folklore suggests when questions probe process and reasoning rather than recall, when rubrics are shared in advance and when sessions are recorded for moderation. Equity requires deliberate design, including practice opportunities and reasonable adjustments consistent with the Disability Standards for Education 2005 (Cth), because oral formats redistribute anxiety rather than remove it. Designed well, the interactive oral is the workhorse of the secured lane.
Programmatic assessment
Programmatic assessment, developed in the health professions, reframes certification as a program-level judgement built from many low-stakes information points and a small number of high-stakes decision points (van der Vleuten et al., 2012). The fit with the two-lane architecture is close to exact: open-lane tasks generate rich, frequent, feedback-oriented evidence of learning, while secured gateways aggregate and verify that evidence before progression or graduation. This is also the level at which the Threshold Standards actually bite, since providers must assure outcomes for qualifications rather than for individual tasks. Programmatic thinking converts the two lanes from a slogan into an auditable design.
AI Literacy as a Graduate Outcome
An assessment settlement built only on securing and permitting would still be incomplete, because it would treat AI as a condition to be managed rather than a capability to be taught. Bearman et al. (2023) show how university discourse tends to cast AI as either saviour or corruptor, obscuring the pedagogical question of what graduates need to be able to do with it. The answer increasingly resembles a literacy: decomposing problems, prompting effectively, verifying outputs against authoritative sources, recognising bias and fabrication, disclosing use honestly, and exercising judgement about when AI use is inappropriate on privacy, confidentiality or professional grounds.
These capabilities can be assessed rigorously in the open lane. Process portfolios that capture prompts, drafts and an annotated critique of AI output, accompanied by reflective commentary, make the student’s evaluative judgement the object of marking rather than the surface polish a model supplies for free. Emerging institutional practice is already moving in this direction (TEQSA, 2024), and the research agenda on learning in hybrid human-AI settings is expanding rapidly (Lodge et al., 2023). Framing AI literacy as a graduate outcome also honours the Australian Qualifications Framework’s insistence that qualifications certify the application of knowledge and skills with judgement, in the contexts graduates will actually inhabit. Australian employers, including the public sector, are normalising AI-assisted work faster than curricula are changing.
The Counterargument: Securing Everything
The strongest objection to the position defended here holds that if unsupervised assessment can no longer be trusted, universities should simply stop relying on it: expand invigilated examinations, verify everything and let integrity rest on supervision rather than design. The argument has genuine force. Examinations are procedurally defensible, familiar to students and markers, and uniform in their conditions. They also neutralise unequal access to premium AI subscriptions, an equity concern that the open lane must otherwise manage deliberately, for instance by providing institutional access to approved tools. For high-stakes professional gateways, secured examination remains exactly the right instrument.
As a general settlement, however, wholesale reversion fails on validity, education and feasibility. Examinations sample a narrow band of time-pressured individual performance and align weakly with the sustained, collaborative, resource-rich work that degrees claim to prepare students for; the Threshold Standards require methods consistent with outcomes, not merely secure ones (Villarroel et al., 2018). An examination-only regime would also graduate students into AI-saturated workplaces without ever having taught or assessed disciplined AI use, abandoning the forward-looking duty that TEQSA (2023) places at the centre of reform. Feasibility tells the same story: examining everything at scale strains venues, budgets and academic workload, and the pandemic experiment with remote proctoring demonstrated both its intrusiveness and its fallibility (Dawson, 2021). The counterargument correctly identifies that verification is indispensable. Its error is to generalise verification into the whole of assessment, when verification belongs where it is designed to matter, at the gateways, and nowhere else.
Conclusion
Generative AI did not break Australian university assessment so much as expose how heavily it leaned on an assumption of unassisted authorship that was already fraying during the contract cheating era. This essay has argued that the sector should land where TEQSA’s guidance points: a program-level architecture in which a limited secured lane of interactive orals, supervised performances and well-aligned examinations warrants outcomes at critical gateways, an open lane teaches and assesses AI-integrated work honestly, and programmatic design binds the two into an auditable whole. Detection cannot anchor this settlement, because the evidence shows it is neither accurate enough to be safe nor neutral enough to be fair, bearing hardest on students who write in English as an additional language. Retreating entirely to the examination hall would be safer only in appearance, trading validity and educational duty for the comfort of supervision. The credibility of Australian qualifications, and of the export sector built upon them, will belong to the institutions that redesign assessment deliberately rather than defend, or merely surveil, what no longer works.
References
Australian Bureau of Statistics. (2024). International trade: Supplementary information, financial year 2023-24.
Bearman, M., Ryan, J., & Ajjawi, R. (2023). Discourses of artificial intelligence in higher education: A critical literature review. Higher Education, 86(2), 369-385.
Bretag, T., Harper, R., Burton, M., Ellis, C., Saddiqui, S., van Haeringen, K., Rozenberg, P., & Newton, P. (2019). Contract cheating: A survey of Australian university students. Studies in Higher Education, 44(11), 1837-1856.
Dawson, P. (2021). Defending assessment security in a digital world: Preventing e-cheating and supporting academic integrity in higher education. Routledge.
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7).
Liu, D., & Bridgeman, A. (2023). What to do about assessments if we can’t out-design or out-run generative AI? Teaching@Sydney, University of Sydney.
Lodge, J. M., Thompson, K., & Corrin, L. (2023). Mapping out a research agenda for generative artificial intelligence in tertiary education. Australasian Journal of Educational Technology, 39(1), 1-8.
Sotiriadou, P., Logan, D., Daly, A., & Guest, R. (2020). The role of authentic assessment to preserve academic integrity and promote skill development and employability. Studies in Higher Education, 45(11), 2132-2148.
Tertiary Education Quality and Standards Agency. (2023). Assessment reform for the age of artificial intelligence.
Tertiary Education Quality and Standards Agency. (2024). Gen AI strategies for Australian higher education: Emerging practice.
van der Vleuten, C. P. M., Schuwirth, L. W. T., Driessen, E. W., Dijkstra, J., Tigelaar, D., Baartman, L. K. J., & van Tartwijk, J. (2012). A model for programmatic assessment fit for purpose. Medical Teacher, 34(3), 205-214.
Villarroel, V., Bloxham, S., Bruna, D., Bruna, C., & Herrera-Seda, C. (2018). Authentic assessment: Creating a blueprint for course design. Assessment & Evaluation in Higher Education, 43(5), 840-854.
Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P., & Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19, Article 26.