How to read this page
A gap is not the same as an inconvenience. When the research behind this workshop was assembled, each item was sorted into one of two piles: things nobody has established yet, and things that are known but awkward. Only the first pile belongs here. A second line was drawn too, between a question the field has not answered and a source we simply could not reach in the time available. The items below are open questions, not failed searches.
Two habits of mind follow from that, and both are worth passing to students. Most of the strong research on AI in education is recent, short, and conducted overseas, so a figure can be real and still not describe an Australian cohort. And a study can be widely shared long before it has been checked, so a striking result is a lead to verify, not a fact to repeat. The dates below matter for the same reason; the evidence was gathered across the middle of 2026, and parts of it will already have moved.
Detector evidence and a fair process
The detection sections explain why an AI-writing score is weak evidence on its own. What is genuinely unsettled is what a fair process should do around that weakness, and how much of the reassuring guidance rests on Australian ground.
No Australian ruling
No Australian court, tribunal or ombudsman has tested the evidentiary weight of a detector score
Every case where a detector-accused student was cleared on the public record is from the United States. Australian student-misconduct case law is thin and predates generative AI. The National Student Ombudsman only began operating in February 2025, and had published no decision on AI integrity as of mid 2026. Whether it will treat these as process complaints is not yet known.
Defensible inference, not settled rule
Whether the absence of drafts counts against a student is a principle, not a codified rule
The sensible position is that missing drafts are not evidence of misconduct, where students were never told to keep them. That position appears in peer-reviewed and institutional sources, but in no TEQSA or ASQA instrument. It is a defensible reading of procedural fairness and the burden of proof. It should be argued as that, rather than quoted as an established rule.
Overseas data only
The bias against writers of English as an additional language is measured only overseas
A 2023 Stanford study found detectors flagged most essays by non-native English writers as AI, and at least one detector flagged nearly all of them, while barely touching native-writer samples. No equivalent Australian false-positive rate exists. The direction of the risk is clear; its size in a CDU cohort is not.
Not stated
What weight CDU itself gives a Turnitin AI percentage is not published
CDU names the tool and the relevant policies. Its public materials do not state a threshold, a requirement to corroborate a score, or a rule that a score alone cannot found an allegation. The prudent practice is well understood; the local rule is not written down where a student or an assessor can point to it.
The cautionary tale is itself unverified
The sector's most-cited warning is the wave of AI-related integrity cases reported at the Australian Catholic University in 2024, put at around six thousand referrals. Every figure traces back to the university's own account, carried in journalism; no regulator, court or ombudsman has confirmed it. The university's own Deputy Vice-Chancellor later called the numbers substantially overstated, and staff and student accounts of how many were dismissed, and how quickly, do not agree. It is a real episode and a fair warning about over-reliance on detectors. It is also a reminder that a number can travel a long way before anyone has checked it.
Whether redesigned assessment actually works
The assessment section sets out defensible designs: secured checkpoints, oral tasks, program-level judgement. What no located study yet shows is that any of these restores the thing they are meant to protect: a valid inference from the work to the student's capability.
Not measured
No evidence that a secured or oral format restores validity under generative AI
The two-lane model of secured and open tasks is endorsed by several institutions and challenged by a peer-reviewed Australian critique on coherence and equity grounds. Neither side rests on a study that measures whether a secured task actually recovers a valid judgement. The design is reasonable; the proof is not in yet.
Perception, not incidence
Interactive orals are harder to cheat in perception, which is not the same as measured
The most-cited Australian study found students perceived interactive oral formats as harder to cheat in. It did not measure a fall in cheating. The sharper claim that orals are not necessarily more authentic belongs to later critics, not to the original authors, and the two are often run together.
Commentary, not evaluation
The claim that interactive orals scale to large cohorts is practitioner commentary
The figures cited for running orals across many students and institutions come from an advocate's blog, with no independent evaluation attached. Scalability may well be real, but it should not be reported as an established outcome. Whether orals defeat a real-time AI avatar is simply unresolved.
No local outcome data
No Australian study measures learning outcomes, or staff consistency, for permitted-use scales
Graduated permitted-use scales are defined and operationalised, but not validated at scale. The most-cited pilot is a single non-Australian site with self-reported results and no measure of whether two markers apply the categories the same way. In VET, there is no current Australian data on how consistently practical and workplace judgements are made.
What AI does to learning
The cognitive-outsourcing section is deliberately careful, and this is why. The headline studies are recent, small, short, mostly overseas, and several are contested or corrected. They are worth knowing and worth doubting in equal measure.
No long horizon
The durable effects of AI use on learning are unmeasured
No located study runs longer than about sixteen weeks. Claims about lasting harm to thinking, or to the developing brain, run ahead of evidence that could only come from far longer work. Short-term findings are just that.
Non-Australian, and not VET
There is no Australian or VET study of AI offloading and skill transfer
The one Australian anchor is a Monash study of a higher-education essay task. The group using ChatGPT gained the most in essay score, yet showed no advantage in what they actually knew or could transfer. It is a careful result. It is also not about VET, and not about long-term skill.
Correlational or unreviewed
Whether AI erodes critical thinking is unresolved, and the alarming studies cannot carry the weight put on them
A much-shared study of "cognitive debt" is a small preprint that has not been peer reviewed, and its own authors call the phrase a metaphor. A published critique lists five methodological problems with it. A widely quoted erosion study is correlational and has been corrected. A 2026 meta-analysis, often miscited under the wrong author's name, finds short-term gains that fade unless the intervention runs for months. None of these establishes that AI use causes a loss of thinking.
Equity, access and who gets watched
Australia has good data on the digital divide in general. What it does not yet have is much data on the AI-specific version of it, or on how AI intersects with disability adjustment and with international study.
Searched, nothing found
No Australian study links paid AI access to assessment outcomes
Whether the students who can pay for the better model do measurably better on assessment is an open and important question, and no Australian study answers it. The nearest evidence compares model capability, not student results. That paying for a premium tool deepens inequality is a reasonable worry, not a measured finding.
No Australian framing
No Australian source draws the line between AI as access support and AI as the thing being assessed
Where a tool is a reasonable adjustment for a disability, and where it starts doing the part of the task under assessment, is a live and unsettled question. The clearest treatments are from the United Kingdom and United States. No Australian regulator statement, and no Australian analysis of the privacy of disability data fed to these tools, was located.
Aggregated only
No Australian survey breaks AI use down by country of origin, and none measures real-world flagging of international students
Surveys treat the international cohort as one block. No Australian data shows whether students who write in English as an additional language are actually flagged more often in live assessment, as opposed to in the overseas laboratory studies. CDU has no published usage data by cohort at all.
The regulator concedes the point without measuring it
TEQSA's own emerging-practice guidance, from November 2024, states that an institutional licence for an AI tool "will not completely mitigate the risk of inequitable access". That is a candid admission that giving everyone the same login does not level the ground. It comes without metrics, and the same guidance offers no reliability or fairness figures for the oral and secured formats it recommends. The honesty is useful; the numbers to act on are not there yet.
Law and governance still in motion
Several of the rules that shape all of this are announced but not yet operating, or were proposed and then dropped. Anything on this page that touches regulation should be read with a date beside it.
Direction reversed
Australia's approach to regulating high-risk AI changed course in late 2025
The National AI Plan of December 2025 withdrew the earlier proposal for mandatory guardrails on high-risk AI in favour of guidance and existing law. The transparency rules for automated decision-making under the amended Privacy Act are enacted but do not commence until December 2026. A copyright exception for AI training was floated and then ruled out in October 2025 in favour of licensing.
Not in the regulator's own words
ASQA's own position on AI in VET is known mostly at second hand
ASQA's transparency statement and its draft AI principles were reached through commentary rather than the regulator's primary pages, which could not be read directly. The Standards for Registered Training Organisations 2025 are confirmed and in force from July 2025. The exact wording of the AI authenticity clause is corroborated here, rather than quoted from source.
Thinly evidenced
How mature university AI governance actually is remains largely unmeasured
The best single figure found is that operational AI-governance structures were documented in only seven of thirty-nine Australian universities. There is no CDU-specific or Northern Australian figure. And no Australian source yet treats locally hosted AI, or browser extensions, as their own data-handling questions.
What we are not unsure about
A page of open questions can read as if nothing is known. That would be the wrong lesson. Several things are settled, and they are the firm ground the workshop stands on.
The points that hold
Student misconduct in Australia is decided on the balance of probabilities. The Briginshaw principle raises the quality of proof needed for a serious allegation, but it is not a separate or higher standard. VET already forbids AI-generated work through the authenticity rule in the rules of evidence, in plain terms. Detectors are unreliable as sole evidence; that part is well established, and only the search for a validated positive alternative is still open. Current Australian copyright law contains no AI-training exception, whatever the future reform brings. And CDU's own Generative AI Policy has confirmed dates and a named owner, so the local framework, whatever its gaps, does exist and can be pointed to.
The honest summary is narrow but usable. We know how to run a fair process and what standard it must meet. We do not yet know, from Australian evidence, how well the new assessment designs work, how AI use shapes learning over time, or how unevenly access falls. Holding both at once is the position the evidence actually supports.
Sources and method
This page draws on the workshop's own evidence-gap registers. Those registers logged each open question during the Phase 1 research, and separated genuine gaps from sources that could not be reached in the time available. The evidence was gathered across the middle of 2026; regulatory items in particular may have moved since. Named studies are described with their limits, because those limits are the point of the page. Some figures are reported at second hand: the Australian Catholic University case numbers, and the exact percentages in the overseas detector and learning studies. These are presented as reported, and are on the workshop's list for primary verification. They are not asserted here as settled.
- Detector evidence and due process draws on the workshop's clusters on academic misconduct, the evidence mosaic and staff consistency. These cover Australian misconduct case law and the United States clearing cases.
- The National Student Ombudsman commenced in February 2025, with no published AI-integrity decision located as of mid 2026.
- Liang et al., "GPT detectors are biased against non-native English writers", Patterns, 2023 (Stanford). The overseas detector-bias finding.
- The reported ACU case numbers: ABC News investigation, 9 October 2025, relayed by RNZ and others.
- The two-lane critique from the University of Western Australia, 2025.
- Sotiriadou, Logan, Daly and Guest, Studies in Higher Education, 2020. The perception of interactive orals as harder to cheat in.
- The AI Assessment Scale pilot: Furze, Perkins, Roe and MacVaugh, AJET, 2024, a single non-Australian site.
- Fan, Gasevic and colleagues, "Beware of metacognitive laziness", British Journal of Educational Technology, 2025 (Monash).
- Bastani et al., PNAS, 2025. A formal correction exists and is on the verification list.
- Kosmyna et al., "Your Brain on ChatGPT", preprint, 2025, with a published critique.
- Gerlich, Societies, 2025 (a corrected version).
- Wu et al., a 2026 meta-analysis in Humanities and Social Sciences Communications.
- TEQSA emerging-practice guidance, November 2024. The inequitable-access quotation.
- Australian digital-inclusion data from the ADII and Mapping the Digital Gap, 2023 to 2025. The confirmed backdrop to the AI-specific gaps.
- The National AI Plan, December 2025, and the Privacy and Other Legislation Amendment Act 2024, with automated-decision transparency from December 2026.
- The Standards for Registered Training Organisations 2025, in force July 2025.
- Atif, Jha, Richards and Cibin, AJET, 2026. Governance structures in seven of thirty-nine universities.
- Briginshaw v Briginshaw (1938) and the balance of probabilities; ASQA's rules of evidence and the authenticity requirement; the Charles Darwin University Generative AI Policy, effective February 2026.
