The Machine in the Room: Artificial Intelligence, Recognition and the Integrity of VET Assessment

Artificial intelligence has arrived in a vocational education system that never fully fixed its assessment practices, and it is now industrialising them, good and bad alike, faster than oversight can keep pace. Recognition of Prior Learning is where the stress concentrates, because it is the most compressed, document-dependent, remote and authenticity-reliant form of assessment in the system. The Standards never mention AI, yet they already govern it, and the one defence no generative tool can breach is the oldest one: make the candidate demonstrate the competence by doing the work. What that means for RTOs, assessors and the integrity of every qualification issued through recognition is the subject of this analysis.

A Second Battle, on Top of the First

For three decades the great battle in Australian vocational education has been a battle over people: over whether trainers and assessors had the skills, whether providers had the will, and whether the system as a whole could be persuaded to assess competence honestly. That battle is not over, and on any fair reading it has not been won. Now a second battle has opened on top of the first, one that is faster, larger and more consequential than anything the sector has faced before. Artificial intelligence (AI) has arrived in the middle of a system that never fully fixed its assessment practices, and it is now industrialising those practices, the good and the bad alike, at a speed and scale the sector is not prepared for.

Recognition of Prior Learning sits at the sharpest edge of this. If AI is a stress test for assessment integrity, recognition is where the stress concentrates, because recognition is already the most compressed, the most document-dependent, the most remotely delivered and the most authenticity-reliant form of assessment in the system. Whatever AI is going to do to assessment integrity, it will do to recognition first and worst. This article is about that collision: how AI threatens the integrity of recognition and assessment, what the Standards require even though they never mention the technology, how recognition can be designed to withstand it, where AI can legitimately help, and why the deepest answer turns out to be the same answer this entire field of practice keeps arriving at.

A short orientation for readers outside the system may help. Vocational Education and Training (VET) is Australia's practical, occupation-focused education sector. Its qualifications are issued by Registered Training Organisations (RTOs), and the sector is regulated by the Australian Skills Quality Authority (ASQA). Recognition of Prior Learning (RPL) is the process that lets a person have skills and knowledge they already hold assessed and recognised toward a qualification rather than being trained again in what they can already do. A training product, in the language of the Standards, is the qualification, skill set or unit of competency being assessed. Generative AI refers to the now widely available tools that can produce fluent text, images, documents and other content on demand in response to a prompt, in a way that can be difficult to distinguish from human work.

1. Why Recognition Is the Softest Target

To understand the threat, start with why recognition is uniquely exposed. A full training course leaves a long trail. There is delivery over weeks or months, repeated interactions between trainer and student, work produced and observed across time, and many opportunities to notice when something does not add up. Recognition has almost none of this. It compresses the entire judgement into a short engagement. It frequently involves no training at all and a candidate the provider has known only briefly. It leans heavily on two kinds of evidence that AI is now extraordinarily good at producing or assisting with: written documentation and verbal answers to questions. And it is increasingly conducted at a distance, through screens, where the assessor cannot fully see what is in the room with the candidate.

Every one of these features is a vulnerability in an age of generative AI. A candidate seeking recognition can use these tools to generate written statements, responses, portfolios and supporting documents that read as competent and personal but reflect no competence the candidate actually holds. In a competency conversation conducted by video, a candidate can have an AI tool generate polished answers in real time and read them out as though the knowledge were their own. The compression that makes recognition efficient is exactly what makes this hard to detect, because there is no long relationship in which the truth would eventually surface. Recognition gives a dishonest candidate a short, document-and-conversation-based, often remote process in which to present AI output as evidence of competence, and it issues, at the end, a full qualification identical to one earned the hard way. That is the softest target in the system, and it is where the integrity of recognition will be won or lost.

2. The Standards Do Not Mention AI, and They Still Govern It

It would be easy to assume that because the Standards for RTOs, remade in 2025 and in full effect from 1 July 2025, do not mention AI, they have nothing to say about it. The opposite is true. The Standards are written in technology-neutral terms, which means they bind regardless of what technology is involved. The obligations were not designed with generative AI in mind, but they apply to it completely, and read in the light of the technology they turn out to be remarkably well aimed.

The most directly relevant obligation is the authenticity rule of evidence in Standard 1.4. Authenticity means the assessor must be assured that the evidence is the genuine work of the candidate being assessed. This is the rule AI attacks most directly, because the entire problem with AI-generated evidence is that it is not the genuine work of the candidate at all. Every form of AI-assisted recognition fraud is, at its core, an authenticity failure, and the authenticity rule already prohibits it. The Standards do not need to name the technology to forbid the deception. ASQA's own assessment guidance already treats work generated with AI tools as an authenticity concern, which confirms the point.

The validity principle in Standard 1.4 is the next anchor, and it is decisive. Validity now expressly requires that assessment includes practical application components that let the candidate demonstrate skills and knowledge in a practical setting. A practical demonstration of a physical or applied skill, observed by an assessor, is something a candidate must do themselves, in real time, with their own hands and judgement. Generative AI cannot perform a practical task on a candidate's behalf in front of an assessor. It can write about the task fluently, but it cannot do the task. The validity principle, by demanding practical demonstration, points the way to the single most effective defence against AI-assisted fraud, a point this article will return to as its central argument.

The currency rule matters too, because currency requires evidence of the candidate's present competence, and AI-generated text describes nothing real about the candidate at all, present or past. The assessment system definition matters, because it requires recognition to produce consistent and valid judgements, which AI-contaminated evidence cannot support. The workforce Standards 3.2 and 3.3 matter, because they require assessors to hold the relevant credentials and current industry skills and knowledge, and the Credential Policy reserves the assessment judgement to a credentialled assessor, which together means the judgement cannot be delegated to a machine. And the whole of Quality Area 4 matters, because governing AI risk is exactly the kind of integrity and risk obligation that governance Standards exist to impose. The Standards, in short, already contain the tools to govern AI. What they require is that providers apply them with the technology in mind.

3. The Threat in Detail: Where Candidate-Side AI Attacks

It is worth being specific about where AI attacks recognition, not to provide any kind of instruction, but because a provider cannot defend against a threat it has not clearly identified. The regulator itself has signalled that the misuse of AI by candidates is an emerging integrity risk, and naming the attack surfaces is the first step in protecting them.

Across all four surfaces, the common thread is authenticity. AI does not create a new category of compliance failure. It massively amplifies an old one. The recognition processes most exposed are exactly those this series has criticised throughout: the document-heavy, conversation-dependent, observation-light processes that were already the weakest on rigour. AI has simply made their weakness catastrophic rather than merely serious.

4. The Antidote Is Structural, Not Technological

Faced with this, the sector's instinct is to reach for detection: AI-detection software, plagiarism tools, forensic analysis of submitted text. This instinct is understandable and almost entirely wrong as a primary strategy, for a simple reason. Detection is an arms race the assessor cannot reliably win. The tools that generate content improve faster than the tools that detect it, detection tools produce both false positives and false negatives, and a compliance strategy built on unreliable detection is a compliance strategy built on sand. A provider that relies on catching AI-generated evidence after the fact has already lost, because it has designed a process that invites the fraud and then hopes to detect it.

The real antidote is structural, and it is the argument this entire field keeps arriving at from every direction. The defence against AI-assisted recognition fraud is to anchor recognition in the direct observation and demonstration of current practical performance, because that is the one form of evidence AI cannot fake on the candidate's behalf.

Consider what direct observation requires. The candidate must perform a real task, in real time, in a practical setting, while a credentialled assessor watches. A generative tool cannot lay the bricks, cannot dress the wound, cannot wire the circuit, cannot care for the patient, cannot operate the machine, cannot prepare the meal, cannot conduct the interview, cannot do the physical and applied work of the occupation in front of the assessor. It can describe all of these things in fluent prose, but description is not the evidence the validity principle requires. The moment recognition demands that the candidate demonstrate the skill rather than write about it, AI is locked out of the evidence, because the evidence is the doing, and the doing must be the candidate's own.

This is why the evidence hierarchy this series has insisted on throughout is not merely good practice but is now the sector's principal defence against AI. A recognition process whose spine is observed practical performance, supported by questioning grounded in what was just observed, is structurally resistant to AI fraud. A recognition process whose spine is submitted documents and a remote conversation is structurally defenceless against it. The technology has, in effect, raised the stakes on a design choice the Standards already pushed providers toward. The providers who built observation-based recognition for reasons of rigour will find they also built the strongest available defence against AI. The providers who clung to document-based recognition will find that AI has rendered their model not just weak but indefensible.

5. Designing Recognition to Withstand AI

Anchoring in observation is the foundation, but a provider can do more to harden recognition against AI, and several practical design choices follow.

The first is to weight the evidence decisively toward observed performance and away from written submission and unstructured conversation. The less the judgement depends on text the candidate supplies or answers they could read from a screen, the less AI can corrupt it. Where written or verbal evidence is used, it should corroborate observed performance, never form the judgement.

The second is to ground questioning in what the candidate has just done. A competency conversation that asks general questions about the field invites generic answers a machine can supply. A conversation that asks the candidate to explain the specific decisions they just made during an observed task, why they did this rather than that, what they would have done if a particular variable had changed, what went differently from how they expected, cannot be answered by a generative tool, because the tool was not present for the task and does not know what the candidate did. Specific, contingent, performance-anchored questioning is far more resistant to real-time AI assistance than general knowledge questions, and it also happens to be a better assessment, because it probes genuine understanding rather than recall.

The third is to require live reasoning rather than polished answers. A candidate asked to think aloud, to work through a problem in real time, to respond to unpredictable follow-up questions that build on their own previous answers, is in a very different position from a candidate who can pause, prompt a machine and read back a result. Unpredictability is the enemy of AI-assisted fraud, because the tool cannot anticipate where a skilled assessor will take the conversation next.

The fourth is to take identity and presence seriously in remote settings. Where any part of recognition must be conducted at a distance, the provider should verify the candidate's identity, be alert to the signs of off-screen assistance, and treat the remote competency conversation as the weakest and most vulnerable element of the process, to be minimised in favour of observed performance wherever possible. Remote-only recognition of practical competence should be regarded with deep suspicion, because it maximises every AI vulnerability at once.

The fifth is to treat fluent generic answers as a warning, not a reassurance. There is a natural tendency to be reassured by a candidate who answers smoothly and comprehensively. In an age of generative AI, an answer that is fluent, comprehensive and curiously generic, that sounds like a textbook rather than a practitioner, should raise rather than lower the assessor's vigilance. Real practitioners answer with the specific, sometimes messy texture of actual experience. Machines answer with polished generality, and training assessors to notice the difference is now part of the craft.

The last is to use detection tools, if at all, only as a minor supplement and never as the strategy. Where AI-detection or similar tools are used, they should inform an assessor's judgement at the margins, not determine outcomes, given their unreliability. The strategy is design, not detection.

6. The Other Threat: AI Industrialising Poor Practice

So far this article has discussed AI as a tool that dishonest candidates use against assessment. But there is a second threat, less obvious and arguably more dangerous in the long run, which comes from providers themselves using AI to industrialise their own practice, including their poor practice.

Generative tools can now produce assessment materials, including recognition tools, mapping documents and assessment instruments, at enormous speed and volume. Used carelessly, this is a recipe for embedding flaws at scale. A provider that uses a generative tool to mass-produce recognition kits without genuine expert review will produce kits that look professional and are subtly or grossly non-compliant: kits that map incorrectly, that miss requirements of the training product, that import errors and invented content, that reproduce the very document-based model AI has just rendered indefensible, and that do so across dozens of qualifications at once. The speed of the tool means that a single flawed approach can be replicated across an entire scope of registration before anyone reviews a word of it.

Mapping is a particular danger. Generative tools can produce mapping documents that align assessment content to units of competency in a way that reads plausibly but is wrong, because the tool does not truly understand the training product; it produces text that resembles correct mapping. A provider that trusts machine-generated mapping without expert verification is building its assessment judgements on a foundation that merely looks sound.

The control for this is already in the Standards. Standard 1.3 requires that assessment tools be reviewed before use to ensure assessment can be conducted consistent with the principles of assessment and the rules of evidence. That pre-use review is precisely the checkpoint at which machine-generated materials must be caught and corrected. A provider may use AI to draft a recognition tool, but it cannot use AI to discharge the obligation to review that tool, which requires credentialled human expertise. The faster AI allows tools to be produced, the more important, not less, the human review before use becomes, and the more disciplined a provider must be about not letting volume outrun verification.

A related and absolute limit concerns the assessment judgement itself. The Workforce Standards require assessors to hold the relevant credentials and current industry currency, and the Credential Policy reserves the judgement of competence to a credentialled assessor, with anyone working under direction not permitted to make assessment judgements. An AI tool is not a credentialled assessor and cannot make an assessment judgement. A provider may use AI to help organise evidence or draft observations, but the determination of whether a candidate is competent against the training product must be made by a credentialled human assessor exercising professional judgement. Delegating the judgement itself to a machine is not a grey area. It is a direct breach, and it strips the qualification of the human accountability the system depends on.

7. A Worked Illustration: The Same Unit, Two Ways

The abstract argument becomes concrete when the same recognition outcome is approached two ways. Take a single unit of competency from a hands-on occupation, and two providers offering recognition for it.

The first runs a document-based process. The candidate is sent a workbook of written questions, asked to write detailed responses describing how they perform the work, asked to supply examples of past documentation, and asked to provide a reference from an employer. The assessor reviews the written responses, reads the documents, holds a thirty-minute video conversation to clarify a few points, and makes a judgement. The second runs an observation-based process for the same unit. It uses the candidate's background only to confirm suitability and plan the assessment, then arranges to observe the candidate performing the actual tasks of the unit, in their workplace or in a realistic simulated environment, against a clear plan of what competent performance looks like. As the candidate works, the assessor records what is done, to what standard, under what conditions, then asks the candidate to explain the specific decisions they just made, to reason aloud about a complication introduced on the spot, to account for why they did this and not that. The table below sets out what AI does to each.

The two providers are assessing the same unit, to the same standard, for the same qualification. The difference between them is entirely in design, and that design difference is now the difference between a process AI destroys and a process AI cannot touch. This is the whole argument in miniature. The defence against AI is not a tool a provider buys. It is a design a provider chooses.

8. The Cross-Border Dimension

AI sharpens a set of risks that were already present wherever recognition crosses borders, and these deserve specific attention because they combine several vulnerabilities at once. Recognition is sometimes sought by candidates located overseas, or delivered through offshore arrangements, or conducted entirely at a distance across jurisdictions. Each of these already strains the authenticity and observation requirements, because the assessor's ability to verify identity, to observe genuine performance, and to control the conditions of assessment is reduced. Add generative AI to a remote, cross-border, document-and-conversation-based recognition process, and very nearly every safeguard is simultaneously compromised. The candidate's identity is harder to verify, off-screen assistance is impossible to rule out, supplied documents are unattributable, and the practical demonstration that would defeat AI is the hardest element to arrange at a distance.

There is a further layer specific to overseas competence. The regulator has warned that evidence of overseas qualifications and experience must be mapped to Australian legislative and regulatory requirements, giving the example of work health and safety obligations, because competence demonstrated under another country's rules is not automatically competence under Australia's. AI can now generate fluent accounts of how a candidate supposedly meets Australian requirements that the candidate has never actually met, making this mapping harder to trust and more important to verify through genuine demonstration. The combination of cross-border delivery, remote conversation, document-based evidence and AI is, in integrity terms, close to a worst case, and providers operating in this space need the strongest observation-based safeguards precisely because every other safeguard is weakest there. Where genuine observation of practical performance cannot be reliably arranged, the honest conclusion may be that defensible recognition cannot be conducted in that configuration at all.

9. Preparing the Assessor Workforce

None of the defences described here is self-executing. They depend on assessors who understand the threat and have the skills to counter it, and that is a workforce capability the sector has barely begun to build. Standards 3.2 and 3.3 require assessors to hold the relevant credentials and current skills and knowledge, and in an age of generative AI, currency of assessment skill now necessarily includes an understanding of how AI affects evidence and how to assess in ways that resist it. An assessor whose practice has not adapted to the existence of generative tools is, in a meaningful sense, no longer current.

The capabilities assessors now need are specific. They must be able to design and conduct observation that anchors the judgement in genuine performance. They must be skilled in performance-anchored, contingent questioning that cannot be answered from a script or a screen. They must be able to recognise the texture of fluent generic answers and treat it as a prompt for deeper probing rather than reassurance. They must understand how to verify identity and detect the signs of off-screen assistance in remote settings. And they must understand the difference between a candidate legitimately using AI as assistive technology and a candidate using it to fake competence, so that they neither wave through fraud nor wrongly penalise a candidate making a reasonable adjustment.

These are teachable skills, but they are not yet widely taught, and the providers that invest in building them will have a defensible recognition workforce while those that do not will have assessors conducting processes they do not realise have already been compromised. This investment connects directly to the validation and continuous improvement obligations discussed elsewhere in this series, because validation that reveals assessors are not equipped for the AI environment has identified a workforce development need the provider must then meet. Building assessor capability for the AI era is not an optional enhancement. It is now part of maintaining the currency the workforce Standards require.

10. The Problem of Speed

There is one more dimension of the AI challenge that is easy to overlook because it is about tempo rather than content. Generative tools do not just change what can be produced. They change how fast it can be produced, and speed itself is a risk the sector's assurance mechanisms are not designed to absorb.

The Standards assume a certain natural pace. Tools are developed, reviewed before use, deployed, and then validated over a five-year cycle under Standard 1.5. Human oversight is feasible because human production was always the bottleneck. AI removes the bottleneck. A provider can now generate a complete suite of recognition materials across its entire scope in the time it once took to draft a single tool, and the temptation to deploy at that speed, ahead of the human review the Standards require, is enormous and largely unguarded. The danger is not that AI produces a flawed tool. It is that AI produces fifty flawed tools faster than anyone can review one, and that they enter use before the pre-use review under Standard 1.3 has caught up. Speed converts a single point of failure into a systemic one almost instantly.

The discipline this demands is counterintuitive in a culture that prizes efficiency. A provider must deliberately refuse to deploy at the speed the technology allows, holding production back to the pace at which genuine human review and verification can keep up. The pre-use review and validation obligations are not merely steps to be completed, but a tempo to be maintained, and a provider that lets AI set the tempo has, in effect, let the machine override its assurance system. Faster is not better when what is being accelerated is the entry of unreviewed materials into live assessment.

11. Where AI Legitimately Helps

It would be a serious mistake to read all of this as a case against AI in recognition and assessment. The technology is not the enemy, and a blanket prohibition would be both unenforceable and foolish. There are genuine, compliant and valuable uses of AI in this work, and distinguishing them from the harmful uses is essential. The distinction that runs through all of it is the same: AI that supports a human professional, improves access for a genuinely competent candidate, or handles administration is legitimate and often beneficial; AI that fakes a candidate's competence, replaces the credentialled judgement, or mass-produces unreviewed materials is a threat. The technology is the same. The use determines whether it strengthens recognition or destroys it.

The line on assistive technology deserves emphasis, because it is where access and integrity meet. A candidate using a tool to help them express themselves, in an occupation that does not require independent written expression, is not committing fraud. They are using assistive technology to demonstrate competence they genuinely hold, which is exactly what reasonable adjustment under the fairness principle envisages. The line, here as everywhere, is whether the tool is helping a competent candidate demonstrate real competence, which is legitimate, or substituting for competence the candidate does not have, which is fraud. Drawing that line requires judgement, but the line itself is clear in principle.

12. Governing AI Under Quality Area 4

Because AI presents both threats and opportunities, it must be governed, and the governance obligations of Quality Area 4 provide the framework. Standard 4.3, the risk management Standard, requires the provider to identify, manage and review risks to students and to the organisation. The integrity risks AI poses to recognition are exactly the kind of risk this Standard requires a provider to have identified and to be actively managing, with controls in place, rather than discovering for the first time at audit. A provider that has never assessed its exposure to AI-assisted recognition fraud has a gap in its risk management.

Standard 4.2, on roles and responsibilities, requires clarity about who does what, which in an AI context means clarity about who is accountable for materials AI helped produce, who reviews them, and who retains the assessment judgement. The accountability cannot evaporate into the machine. A human must own every output. Standard 4.4, on monitoring and continuous improvement, provides the means to detect AI-driven anomalies in recognition data, such as sudden changes in the character of submitted evidence, patterns suggesting fabricated documentation, or clusters of suspiciously uniform responses, and to improve the system in response.

Standard 4.1, on integrity and the culture set by governing persons, sits beneath all of it. A provider's response to AI will ultimately reflect its culture. A provider whose leadership treats recognition as a revenue stream to be maximised will be tempted to use AI to produce more recognition faster and cheaper, embedding the worst risks. A provider whose leadership genuinely values integrity will use AI to support good practice while defending hard against its misuse. The technology amplifies whatever culture it is dropped into, which is why the integrity culture obligation in Standard 4.1 is, in the end, the most important governance control of all.

A practical consequence is that providers now need explicit policies on AI use, covering both candidates and staff. Candidates should be told clearly what use of AI is and is not acceptable in their recognition assessment, and recognition should be designed so that prohibited use is difficult and detectable through its structure. Staff should have clear guidance on how AI may be used in producing materials and supporting assessment, with the boundaries around the assessment judgement and pre-use review made explicit. A provider with no AI policy is not technology-neutral. It is simply ungoverned in an area of serious and growing risk.

13. Validating Recognition in an AI World

The assurance mechanisms discussed elsewhere in this series acquire a new dimension under AI, and validation in particular must adapt. When a validator examines a sample of recognition decisions, they must now ask a question that would not have arisen a few years ago: can the provider be assured that the evidence underpinning these decisions was the genuine work of the candidates, in a world where so much evidence can be machine-generated? A validation that examines whether the tools were well designed and the judgements consistent, but never considers whether the underlying evidence might have been artificially generated, is no longer a complete validation. Authenticity under AI is now part of what validation must scrutinise, and it reinforces, yet again, the importance of recognition evidence being anchored in observed performance that a validator can see was genuinely the candidate's own.

14. The Deeper Risk and the Closing Window

Step back from the detail, and a larger picture emerges that should concern everyone who cares about the sector. AI is being introduced into a system that never resolved its underlying assessment problems. For decades, too much of the sector has relied on written evidence over demonstration, has assessed talk rather than performance, and has treated assessment as a documentation exercise rather than a judgement of competence. AI does not fix any of this. It industrialises it. It takes whatever practices the sector has and embeds and scales them, faster than human oversight can keep pace.

If the sector industrialises good assessment practice with AI, anchored in observation, demonstration and genuine professional judgement, the technology could be a real benefit, freeing human effort for the work that matters and extending access to candidates who were previously excluded. If instead the sector industrialises its existing poor practice, mass-producing document-based recognition kits, automating away human judgement, and chasing volume over rigour, then AI will lock in the sector's weaknesses at a scale and speed that may be impossible to unwind. The choice is being made right now, in thousands of providers, mostly without anyone consciously deciding it.

This is why the unfinished battle over assessment quality and the new battle over AI are really the same battle. The sector cannot defend recognition against AI without finally embracing the assessment design it should have embraced all along: assessment that requires candidates to demonstrate competence by doing the work, observed by a credentialled professional who exercises genuine judgement. AI has, paradoxically, made the case for good assessment more urgent and more obvious than decades of argument ever could, because it has made the alternative not just weak but unworkable. There is a window, and it is closing, in which the sector can get its assessment design right before AI sets the current practice in concrete. The providers that use this moment to rebuild recognition around observed performance will emerge stronger and more defensible. The providers that use AI to do more of what they were already doing will discover, too late, that they have automated their own undoing.

Conclusion: The Machine Cannot Do the Job

The temptation, confronting AI, is to believe that the answer must itself be high-technology: better detection, smarter software, an arms race of tools. It is not. The answer is older and simpler than the technology that prompts it, and the Standards have been pointing at it the whole time. Make the candidate demonstrate the competence by doing the work, in a practical setting, observed by a credentialled assessor, with questioning anchored in what was actually done. A machine can write about welding, but it cannot weld in front of an assessor. It can describe a clinical procedure, but it cannot perform one on the assessor's table. It can compose an essay about the work, but it cannot be the candidate doing the job. The authenticity rule, the validity principle that now demands practical demonstration, and the requirement that a credentialled human make the judgement, together form a defence that no generative tool can breach, for the simple reason that competence, in the end, is something a person either has or does not have, and the only reliable way to know is to watch them use it.

AI has changed almost everything about how text and documents are produced. It has changed nothing about what competence is. A qualification certifies that a person can do a job, and the only future-proof way to assess that, in an age when everything that can be written can be written by a machine, is to require the person to do the job and watch them do it. The machine is now in the room, and it will not leave. The task for the sector is to design recognition so that, however clever the machine becomes, the competence being assessed remains unmistakably, demonstrably, the candidate's own.

References and Further Reading

National Vocational Education and Training Regulator (Outcome Standards for NVR Registered Training Organisations) Instrument 2025 (F2025L00354), Standards 1.3, 1.4, 1.5, 3.2, 3.3 and 4.1 to 4.4. https://www.legislation.gov.au

Australian Skills Quality Authority (2025). Credential Policy. https://www.asqa.gov.au

Australian Skills Quality Authority (2025). Practice Guide: Assessment. https://www.asqa.gov.au

Australian Skills Quality Authority (2025). Practice Guide: Recognition of Prior Learning and Credit Transfer. https://www.asqa.gov.au

Australian Skills Quality Authority (2026). Responsible Use of AI in VET: Sector Workshops and Draft AI Principles. https://www.asqa.gov.au

Disability Standards for Education 2005 (Cth). https://www.legislation.gov.au