Artificial intelligence has moved faster than the rules written to govern it, and in Australian VET the gap has landed on the assessor. Standard 1.4 requires assessment evidence to be authentic, the student's own original and genuine work, and that obligation has not changed; what has changed is the scale of the threat and the inadequacy of the AI detectors many RTOs are reaching for. No detector on the market satisfies the rule, because the Standard does not require a tool. It requires a judgement. This article sets out what the authenticity rule actually demands, why detection cannot discharge it, and what it means for RTOs, assessors and the students whose competence the system must genuinely prove.
A Gap That Has Landed on the Assessor
Artificial intelligence has moved faster than the regulation designed to govern it, and in Australian vocational education and training that gap has landed squarely on the assessor. Standard 1.4(2)(b)(iii) of the Outcome Standards for NVR Registered Training Organisations Instrument 2025 requires that assessment evidence be authentic: the student's own original and genuine work. That requirement has always existed. What has changed is the scale of the threat to it, and the inadequacy of the tools many RTOs are reaching for to meet it. No AI detector currently on the market passes the test Standard 1.4 sets, because the Standard does not require a tool. It requires a judgement.
1. The Assessment System and Its Legal Architecture
The Instrument, which came into full regulatory effect on 1 July 2025, replaced the Standards for RTOs 2015 and is built around outcomes rather than prescriptive processes, but Standard 1.4 is among its most detailed provisions. Standard 1.3 sets the overarching requirement: the assessment system must be fit for purpose and consistent with the training product, and assessment tools must be reviewed prior to use to ensure assessment can be conducted in line with the principles of assessment and rules of evidence. Standard 1.4 then defines what those principles and rules demand, and is the operational heart of assessment compliance.
Standard 1.4(2)(a) sets out four principles governing how assessment is conducted: fairness, flexibility, validity and reliability. Standard 1.4(2)(b) sets out four rules of evidence that must inform every individual assessment judgement: validity, sufficiency, authenticity and currency. These are not aspirational values; they are the standard against which every judgement is tested. An assessor who reaches a competency determination without satisfying all four rules has made a judgement that cannot withstand audit. The AI authenticity problem sits inside this framework. The question is not whether students use AI; they do, widely and increasingly, but whether the assessment system is designed to produce evidence that satisfies the authenticity rule, and whether assessors are equipped to make defensible judgements where authenticity is in doubt.
2. What the Authenticity Rule Actually Says
The authenticity rule requires the assessor to be assured that the student's assessment evidence is the student's own original and genuine work. Three words carry the weight of the obligation, and each matters.
Assured does not mean suspicious, uncertain, or reliant on a platform score. It means the assessor has reached a positive state of confidence, and that the assessor, not an algorithm, is the one who is assured. This is a judgement standard, not a detection standard, and the obligation sits with the credentialled person who makes the determination, not a software vendor. Original means the work has not been derived or substantially reproduced from an external source without acknowledgement; a student who has a language model produce a written response and submits it as their own has submitted work that is not original, because the content was generated by the model, not by the student's application of their own skills and knowledge. Genuine means the work authentically represents what the student knows and can do, which is subtler: a student could write original prose that still does not represent genuine competency, by reproducing memorised material they do not understand. When a language model writes the response, it is neither original nor a genuine representation of competency. Taken together, the rule requires the assessor to be positively and personally satisfied, on the evidence before them, that the work was produced by the student and reflects what the student actually knows and can do. That satisfaction cannot be outsourced to a detector or delegated to a percentage score. It must be grounded in assessment design and professional judgement.
|
The Standard Requires a Judgement, Not a Tool |
|
Standard 1.4(2)(b)(iii) places the obligation to be assured on the assessor. It does not say a tool must clear the work; it says the assessor must be assured the work is the student's own original and genuine work. That is a positive, personal, professional standard, not a number generated by a third-party platform. The Standard has always demanded a judgement. AI has not changed the standard. It has changed how hard the judgement is to make well. |
3. Why AI Detectors Do Not Satisfy the Standard
At the ASQA workshop that prompted this article, participants asked whether ASQA recommends a system for detecting AI-generated responses. ASQA's stated position, reflected in its AI Transparency Statement and the draft AI Principles it has been sharing with the sector during its 2026 workshops, is that AI detection tools are not foolproof, and that regulatory and assessment decisions must remain human-made. That is an understatement of the problem these tools create when relied on as the primary authenticity mechanism.
Detection tools produce both false positives and false negatives at rates well documented in published research, and both create compliance risk. A false positive triggers an integrity process against a student who submitted genuine work; research has found that detectors mislabel the writing of capable non-native English speakers at several times the rate of native speakers, and pursuing a student on a tool's incorrect flag is not fair assessment under the Standard 1.4(2)(a)(i) fairness principle, exposing the RTO to complaints under Standard 2.7 and appeals under Standard 2.8. A false negative is the more common problem: modern models produce content that detectors increasingly fail to identify, and studies show that editing only a small fraction of AI output collapses detector recall. An assessor who relies on a clean detection report has not been assured of anything; they have been handed a number that may not reflect reality. There is also a deeper legal problem. Standard 1.4(2)(b) places the obligation on the assessor to make individual judgements justified by the rules of evidence, and the word individual means a credentialled professional must reach a personal, reasoned judgement. Marking work authentic because a tool said so, without independent evaluation, is not making that judgement. Detection tools may form one element of an investigation when authenticity is specifically in question, but they cannot substitute for assessment design that builds authenticity in, and they cannot replace the assessor's evaluation of the evidence.
|
Why Detectors Fail the Standard |
|
A false positive accuses a genuine student and breaches the fairness principle. A false negative clears AI-generated work and assures the assessor of nothing. And even an accurate detector cannot make the individual judgement Standard 1.4(2)(b) reserves to the credentialled assessor. A tool can inform an investigation. It cannot be the authenticity mechanism, because the Standard puts the obligation on a person, not a platform. |
4. The Assessor's Actual Obligation
The rule does not require the assessor to prove a student used AI. It requires the assessor to be assured the work is the student's own, which is a positive obligation, not a reactive one. In practice, the assessor must be able to answer three questions about each piece of evidence. Does the evidence demonstrate the student's skills and knowledge consistently with the training product and the student's performance across other tasks? Is there anything about the quality, structure, vocabulary, coherence or consistency of the submission that raises a reasonable question about whether the student produced it? And if a question has arisen, has the assessor verified the student's understanding through supplementary assessment or questioning? The first is mainly a design question, answered before assessment begins; the second is the assessor's responsibility during marking; the third is the response when doubt arises.
What the assessor cannot do is proceed to a competency determination while holding unresolved doubt about authenticity. A finding of competence made in the face of unresolved doubt is not a valid judgement under the Standards; it is a risk event for the RTO. This is not a counsel of perfection. The Standard does not require mathematical certainty, but a state of professional assurance, which is a reasonable and defensible standard for a credentialled practitioner, and it requires genuine engagement with the evidence rather than a mechanical pass through a tool.
5. Assessment Design Strategies That Build Authenticity In
The most effective response is not reactive detection after submission. It is design that makes AI-generated responses visible, implausible or unhelpful to a student who has not genuinely engaged with the training. Several approaches have direct grounding in Standards 1.3 and 1.4, set out in the following table.
|
Design strategy |
How it builds authenticity in |
Standards anchor |
|
Contextualisation and specificity |
Tasks requiring application to scenarios the student observed, supervised practice they performed, or cases discussed in class produce responses a language model cannot generate without information it does not have |
Standard 1.4(2)(a)(iii), practical application in a practical setting |
|
Progressive evidence collection |
Evidence accumulated across training, through participation, drafts, verbal check-ins and formative tasks, builds a picture of genuine understanding that a single polished submission unrelated to prior performance cannot satisfy |
Standard 1.4(2)(b)(ii), sufficiency |
|
Oral assessment and supplementary questioning |
A student who used AI often cannot explain, extend or apply the content under structured questioning; this remains the most reliable authenticity mechanism, and can be a routine element rather than a formal viva |
Standard 1.4(2)(a)(i), fairness, which provides for supplementary assessment |
|
Third party and workplace evidence |
A supervisor confirming a specific task performed on a specific day in a specific workplace provides evidence that is authentic and independently verified, which AI cannot replicate |
Standard 1.4(2)(b)(i), validity and adequacy |
|
Design Authenticity In, Do Not Detect It After |
|
The practical application requirement is not only a validity and fairness tool. It is a structural authenticity mechanism: an assessment built around demonstrated, contextual, observable performance does not need a detector, because a language model cannot produce the evidence. An assessment built entirely on generic written submissions has a structural authenticity vulnerability no detector can cure. |
6. What Assessors Must Document Before Making the Call
When an assessor forms a view about authenticity, that view must be documented; this flows from Standard 1.4(2)(b) and the assessment system obligations in Standard 1.3. Where the assessor is satisfied the work is the student's own, the evidence on which that satisfaction rests- observation notes, verbal confirmation, progressive tasks or third-party reports- should be identifiable from the record, though not every submission needs a lengthy narrative. Where the assessor has formed a doubt and acted on it, the doubt, its nature, the action taken such as supplementary oral questioning, and the outcome must all be on the file. An assessor who formed a concern, did nothing, and marked the student competent has created an indefensible record; an assessor who formed a concern, conducted a structured verbal check, satisfied themselves of genuine understanding, and documented it has done exactly what the rule requires.
Where a student is found to have submitted AI-generated work and a formal integrity process begins, the RTO's feedback and complaints process under Standard 2.7 and appeals process under Standard 2.8 must be followed; the student must be afforded procedural fairness, and the decision, evidence and outcome must be documented and communicated. A student declaration of AI non-use is a sensible step that establishes an evidentiary baseline and strengthens the RTO's position where a declaration was false, but it does not by itself satisfy the authenticity rule. The assessor must still be independently assured. The declaration supports the process; it does not substitute for it.
7. Building a Standard 4.3 AI Risk Policy
Standard 4.3 requires the RTO to identify, manage and review risks to students, staff and the organisation, to manage conflicts of interest, and, where it trains students under 18, to manage risks to their safety and wellbeing in accordance with the National Principles for Child Safe Organisations. AI use in assessment falls squarely within the risk identification and management obligation, and an RTO that has not connected its AI response to this risk system has a governance gap. A compliant policy is not a long document; it is a proportionate, structured response, set out in the following table.
|
Policy element |
What it covers |
|
Risk identification |
AI use in assessment as a risk category, with sub-categories: AI-generated written submissions; AI-assisted verbal responses; AI-generated workplace documentation; and AI-mediated paraphrasing of third-party material, each with a different authenticity profile |
|
Risk assessment |
Calibrated to the cohort, delivery mode and qualification mix; an online, written-assessment-heavy, digitally capable cohort carries higher risk than face-to-face practical delivery in a trade, and the response should be proportionate |
|
Controls |
Preventive controls in design (contextualised tasks, progressive evidence, oral questioning, workplace evidence) and detective controls in response (assessor development in recognising AI characteristics, defined escalation, and a structured supplementary assessment process) |
|
Monitoring and continuous improvement |
Periodic review under Standard 4.4, informed by assessor reports, integrity cases, student feedback and ASQA guidance, because a policy adequate when written will not stay adequate as AI advances |
8. The Assessment Tool Review Checklist
Because Standard 1.3 requires assessment tools to be reviewed prior to use, the AI authenticity risk should be a standard item in that pre-use review. The following questions frame it.
|
Review question |
What to confirm |
|
Could a generative AI system produce an acceptable response that a student who had not engaged with the training could submit as their own? |
If yes, the task must be redesigned or supplemented; a question any model can answer needs a contextual or applied component a model cannot |
|
Does the task include at least one component generating inherently personal, contextual or observable evidence? |
A workplace observation, practical demonstration, verbal confirmation, reflective journal tied to specific training, or case analysis from the student's placement; not every task needs this, but a wholly generic written assessment is structurally vulnerable |
|
Is there a documented process for what the assessor does when not satisfied about authenticity? |
The triggers, the steps, the documentation required, and the outcome options: supplementary assessment, integrity referral, or proceed where supplementary assessment resolves the doubt |
|
Are assessors receiving AI-literacy professional development under Standard 3.2(c)? |
AI literacy is now part of the current skills and knowledge required to assess effectively, and must be in the development program |
9. What This Means for RTOs
For an RTO building a defensible position on AI, the work falls in a clear order.
First, treat it as a design problem before a detection problem. Build at least one contextual, applied or observable component into every assessment, so authentic evidence is the natural product of genuine engagement and AI substitution is visible or useless. This is the single most effective control, and it is grounded in the practical application and sufficiency rules, not bolted on.
Second, equip and require the assessor's judgement, and document it. Make clear that the authenticity obligation is the assessor's, not a tool's; give assessors a defined process for resolving doubt through supplementary questioning; record the basis for assurance and the steps taken where doubt arose; and provide AI-literacy development under Standard 3.2(c). A detector may assist an investigation, but it never makes the decision.
Third, govern AI as a managed risk. Document an AI risk framework under Standard 4.3 with preventive design controls and detective response processes, follow Standards 2.7 and 2.8 where integrity matters proceed to a finding, and review the framework under Standard 4.4 as the technology changes. Passive documentation will not hold; an actively managed system will.
10. Conclusion: An Old Rule Facing a New Challenge
The questions practitioners are asking reveal a sector looking for regulatory certainty in an uncertain technological environment, wanting ASQA to name a tool, set a threshold, or prescribe a standard. That certainty is not coming, and for good reason: prescribing a detection tool as the compliance mechanism would define the authenticity rule by reference to a technology obsolete within months. The Outcome Standards deliberately define what must be achieved, not the mechanism, because the how must respond to the training product, the cohort, the delivery mode and the technological environment. That flexibility is the point, and it means RTOs cannot adopt a passive posture. The question is not whether ASQA will audit AI policies. It is whether the assessment system produces authentic evidence of competence, a question that existed long before AI did.
The authenticity rule is not a new rule. It is an old rule facing a new challenge. The assessors who satisfy it will not be the ones with the most sophisticated detection software. They will be the ones who understand what the rule requires, design tasks that make authentic evidence the natural product of genuine engagement, and make professional judgements grounded in that evidence. That is what the Standard has always demanded. AI has simply made doing it well more urgent.
|
Summary: Meeting the Authenticity Rule in the Age of AI |
|
1. Standard 1.4(2)(b)(iii) requires the assessor to be assured that the evidence is the student's own original and genuine work; the obligation is the assessor's, not a tool's. 2. No AI detector satisfies the rule; the Standard requires a judgement, not a detection score. 3. Detectors produce false positives (accusing genuine students, breaching the fairness principle) and false negatives (clearing AI work), and neither makes the individual judgement Standard 1.4(2)(b) reserves to the assessor. 4. The assessor cannot reach a competent determination while holding unresolved authenticity doubt; doing so is not a valid judgement. 5. The primary response is design: contextualised and applied tasks under the practical application principle make AI substitution visible or unhelpful. 6. Progressive evidence collection satisfies the sufficiency rule and resists single-submission substitution. 7. Oral and supplementary questioning is the most reliable mechanism and is grounded in the fairness principle's provision for supplementary assessment. 8. Third-party and workplace evidence provides independently verified authenticity that AI cannot replicate. 9. The assessor's assurance, and any doubt and its resolution, must be documented; a student declaration supports the process but does not satisfy the rule. 10. AI must be governed as a managed risk under Standard 4.3, with assessor AI literacy under Standard 3.2(c) and review under Standard 4.4. |
References and Further Reading
Australian Skills Quality Authority (2026). AI Transparency Statement; and draft AI Principles shared through the 2026 sector workshops.
Australian Skills Quality Authority (2025). Practice Guides on Assessment and on the Integrity of Nationally Recognised Training Products.
Federal Register of Legislation (2025). National Vocational Education and Training Regulator (Outcome Standards for NVR Registered Training Organisations) Instrument 2025, Outcome Standards 1.3, 1.4, 2.7, 2.8, 3.2, 4.3 and 4.4.
National Principles for Child Safe Organisations. Australian Human Rights Commission.
