The settled ground
Australian law on the standard of proof for student misconduct is not in doubt. Misconduct is an administrative matter decided on the balance of probabilities, and the High Court's decision in Briginshaw v Briginshaw (1938) adds the qualification that matters here: the more serious the allegation and its consequences, the more cogent the evidence must be before a decision-maker can be satisfied. Briginshaw is routinely misread as creating a higher standard of proof; it does not. The standard stays the same, but flimsy evidence that might carry a trivial allegation will not carry one that can end a nursing student's registration pathway. Section 140 of the Evidence Act 1995 (Cth) codifies the same factors for courts and tribunals, and university committees, though generally not bound by the rules of evidence, apply the principle by analogy as part of procedural fairness.
CDU's own instruments sit inside this frame without naming it. The Student Academic Misconduct Procedure requires that facts be established "to the satisfaction of the decision-maker", names no formal standard and does not reference Briginshaw, though the Complaints Policy applies a balance-of-probabilities test at the appeal stage. The Procedure permits text-matching software "as a guide" only, and neither it, the Academic Integrity Policy nor the Generative AI Policy contains any methodology for detecting AI use or any rule about what a detector score can establish. That silence is worth knowing about: when a staff member treats a percentage as a finding, they are not applying CDU policy, because CDU policy says no such thing.
One more piece of settled ground, of the negative kind: as at the research cut-off for this course, no Australian court, tribunal or ombudsman decision could be located that adjudicates a misconduct allegation resting on an AI-detector score. The principles are settled; their application to detector-based allegations has never been judicially tested in the located Australian record. Every institution running such cases is doing so without a precedent to lean on.
What the regulator and the vendor both say
On the central question, the Australian regulator's position is explicit. The TEQSA-hosted Academic Integrity Toolkit states that "the AI score alone is insufficient to bring an allegation of misconduct" and that additional evidence is required. The same guidance carries an instruction that is easy to skip past and does most of the fairness work: look for evidence that disconfirms AI use, not only evidence that confirms it. A flagged score seeds confirmation bias, and everything the reader then notices about the essay tends to become corroboration; deliberately seeking the innocent explanation is the corrective.
The vendor agrees. Turnitin's own guidance states that its AI indicator is not a determination of misconduct and should not be the sole basis for adverse action, and that false positives are a possibility. When the company selling the tool and the regulator overseeing the sector say the same sentence, the sentence is worth taking seriously.
The VET side needs stating separately, because the frame is different. Under the Standards for RTOs, assessment evidence must satisfy the rules of evidence: it must be valid, sufficient, authentic and current, and ASQA's Assessment Practice Guide (2025) defines authenticity as the assessor being assured the evidence is the original and genuine work of the student, expressly extended to work not generated with AI tools. A detector score speaks weakly, and only, to authenticity; it contributes nothing to validity, sufficiency or currency, and it fails most often on exactly the short, formulaic, templated responses that competency assessment produces. Where authenticity is in doubt in VET, the documented methods are positive ones: supervised or observed tasks, oral questioning, re-assessment under supervision, a competency conversation. Doubt is resolved by gathering assurance, not by trusting a probability.
The evidence mosaic
If a score cannot found an allegation, what can? The honest answer from the Australian record is: an accumulation of signals, none of which is sufficient alone and which differ sharply in how much weight they can bear. The TEQSA-hosted guidance ranks them unevenly without assigning numbers: a student's inability to answer questions about their own work is described as a clear signal, a disparity between supervised and unsupervised performance as a possible signal, and the raw score as insufficient on its own. The useful ordering principle underneath is checkability. The further a signal moves from "a model estimated something" toward "anyone can verify this directly", the more probative weight it carries.
Two tiles need footnotes. Fabricated citations earn their high placement because they are falsifiable by anyone with library access: the cited article exists and supports the claim, or it does not, which is a different epistemic category from a probability estimate. Process evidence is genuinely probative, and in documented cases version history and browser records have exonerated students; but it is not self-authenticating in either direction, because current AI tools can generate a convincing revision history, a point made in both Australian peer-reviewed work and University of Sydney guidance. Oral explanation sits at the strong end with its own qualification: Australian systematic-review evidence finds the reliability of oral assessment depends on how well it is designed and scaffolded, not on the format itself, and it carries known equity considerations around anxiety, language background and power dynamics.
CDU's Academic Integrity Policy already reaches for the strong end of the mosaic: where authenticity is questioned, it authorises asking a student to present, explain or defend their ideas verbally, and to provide research notes, drafts or version histories. The policy machinery points at conversation and process, not at a percentage.
Drafts, and the fairness trap
A suspicion forms, and the reflex follows: ask the student for their drafts. Handled carelessly, this reflex converts a weak signal into an unfair process, and the trap has a specific shape. If students were never told to keep drafts, what does their absence prove? A student who wrote in one library sitting, or drafted on paper, or simply never enabled version history, has nothing to produce, and producing nothing looks like guilt precisely when the burden has been quietly flipped onto the student.
The status of the fairness principle here deserves precise reporting. Australian peer-reviewed work (Bassett and colleagues, 2026) states that documentation requirements must be communicated in advance and that it is procedurally unfair to penalise a student for not producing records they were never told to keep, adding that the burden of proof lies with the institution and students should not be required to prove their innocence. University of Sydney guidance says the same: silence or lack of drafts cannot be interpreted as guilt. No TEQSA or ASQA instrument codifies that rule in those terms; what the regulator's material establishes is the components underneath it, the civil standard, the need for corroboration, natural justice, and no single indicator sufficing. So the principle is well grounded in Australian scholarship and institutional guidance, and it is not, or not yet, a regulator's rule. There is also an unresolved tension worth naming plainly: CDU policy permits the University to ask a student for evidence that work is their own, while the scholarly and United Kingdom ombudsman position holds that the burden of proof stays with the institution. Both statements are true of their sources, and this course does not pretend they have been reconciled.
What the Australian record shows
One Australian episode dominates the documented record, and it needs to be reported with its caveat attached. In October 2025, ABC News reported that Australian Catholic University registered approximately 6,000 alleged misconduct cases across nine campuses in 2024, about 90 per cent involving AI-use concerns, with roughly a quarter of referrals dismissed on investigation. ACU's deputy vice-chancellor stated that any case where the Turnitin AI detector was the sole evidence was dismissed immediately, and the university discontinued that detector in March 2025 after finding it ineffective. The caveat: these figures are ACU's own account, reported through journalism; no regulator, ombudsman or court has independently confirmed the episode or its resolution, and this course treats it accordingly, as a documented institutional account rather than an adjudicated outcome.
What the reporting documents beyond the statistics is the harm a process can inflict on its way to the right answer. A nursing student's transcript carried a "results withheld" annotation for about six months while her case was investigated; she was cleared, and believes the annotation cost her a graduate position. A paramedicine student whose essay was flagged 84 per cent AI was asked to produce internet search histories, handwritten notes and typed notes to rebut the score. Both were ultimately cleared. The finding underneath, well supported in procedural-justice research, is that process harm is real and independent of outcome: delay, intrusion and reversed burdens damage students, and damage trust in the institution, even when the eventual decision is correct.
Two further Australian findings complete the picture. Individual academic judgement, unaided, is poor at this task: in blind Australian experiments, experienced markers identified purchased assignments only 48 to 62 per cent of the time. And staff responses to suspected misconduct are documented as varied and inconsistent, shaped by individual beliefs, discipline norms and workload; Australian peer-reviewed studies from 2011 through 2026 have found no consistency in how similar conduct is interpreted or penalised. Neither instinct nor a score, alone or together, is a reliable instrument, which is why everything on this page keeps returning to corroboration.
Everything above converges on a short list. Treat the score as a trigger for inquiry, never as the finding. Give the student the report and the evidence, and a real opportunity to respond; CDU's procedure requires written notice, a copy of the evidence and five working days. Seek disconfirming evidence as deliberately as confirming evidence. Prefer the checkable end of the mosaic: citations, process artefacts offered by the student, and a conversation about the work. Match the cogency of the evidence to the seriousness of the consequences, which is all Briginshaw asks. And remember the base rate: at realistic levels of misuse, a meaningful share of flags are landing on honest work.
A smoke alarm, not a meter
The framing that best survives contact with all of the evidence is this: an AI detector is a smoke alarm, not a meter. A smoke alarm going off is a reason to get up and look; it is not a measurement of how much fire there is, and it goes off for burnt toast. Nobody evacuates a building, or ends a nursing career, on the sound alone. The number on a detector report earns exactly that much authority: attention, followed by human inquiry into evidence that can actually be examined. Anything more is the machine being asked a question, "did this student cheat?", that it was never built to answer.
Watch
ABC News reporting on Australian students wrongly accused by detection tools, and what the appeals process felt like from inside; the Australian companion piece to the cases discussed above.
ABC News · AI-Detector Systems Falsely Accusing Students of Cheating · 2024
Sources
Claims on this page rest on Australian primary and peer-reviewed material wherever it exists; news reporting is identified as such, and the ACU figures are reported with their verification status stated.
- Briginshaw v Briginshaw [1938] HCA 34; (1938) 60 CLR 336, and Evidence Act 1995 (Cth) s 140. The civil standard and the cogency principle.
- CDU Student Academic Misconduct Procedure (id=239), approved 29 October 2025, and CDU Academic Integrity Policy (id=50), effective 31 October 2025. Decision standard wording, notice and response requirements, oral defence and process evidence provisions.
- Detecting plagiarism of AI-generated text in student assessments, Guy Curtis, TEQSA Academic Integrity Toolkit, date unavailable. The "AI score alone is insufficient" statement, the signal hierarchy and the disconfirming-evidence instruction.
- ASQA Assessment Practice Guide, published 17 June 2025. The rules of evidence and the authenticity definition extended to AI.
- Heads we win, tails you lose: AI detectors in education, Bassett et al., Journal of Higher Education Policy and Management, online 29 January 2026. Burden of proof, advance notice of documentation requirements, and the base-rate argument; Australian authors.
- False flags and broken trust, Teaching@Sydney, 30 October 2025. Silence or lack of drafts cannot be read as guilt; AI-generated revision histories.
- ABC News reporting on the ACU episode (via RNZ), 9 October 2025. The ACU figures and named-student accounts; institutional statistics via journalism, not independently confirmed by any regulator or court.
- Dawson and Sutherland-Smith, Assessment & Evaluation in Higher Education, 2018, and Dawson, Sutherland-Smith and Ricksen, 2020. Unaided marker detection at 62 and 48 per cent in Australian blind experiments.
- Birks, Mills, Allen and Tee, International Journal for Educational Integrity, 2020, and De Maio, Dixon and Yeo, Issues in Educational Research, 2019. Documented variation and inconsistency in Australian staff responses to misconduct.
- Nallaya, Gentili, Weeks and Baldock, Issues in Educational Research 34(2), 2024. Oral assessment reliability is design-dependent; Australian systematic review.
- Office of the Independent Adjudicator casework note, 8 July 2025. United Kingdom comparison only: burden on the provider; students receive the detection report.