Can AI Mark Assessments? What Standard 1.4 and the Credential Policy Actually Permit

Modern language models can plainly evaluate a written response in seconds, and the marking workload across online and high-volume delivery makes that capability tempting. But the real question is not whether AI can read an answer. It is whether using AI in the marking process satisfies the requirement in Standard 1.4(2)(b) that assessors make individual assessment judgements justified by the rules of evidence, and the Credential Policy requirement that only a credentialled person may determine competency. This article draws that line precisely, so RTOs can design AI-assisted marking workflows that are compliant rather than guess where the boundary lies. It is general guidance on the Standards, not legal advice, and the controlling references are the current Outcome Standards and Credential Policy.

The Right Question, Asked at the Wrong Level

The question is being asked in every RTO that has trialled AI tools in the past two years: can AI mark assessments? It is the right question asked at the wrong level. Whether AI is technically capable of evaluating a written response is not in doubt; modern language models plainly are. The real question is whether using AI in the marking process is consistent with the assessment system obligations in Standard 1.4 of the Outcome Standards, and specifically whether it satisfies the requirement for individual assessment judgements made by credentialled persons. The answer turns on understanding precisely what Standard 1.4 and the Credential Policy require, where in the marking process human professional judgement is not negotiable, and what an audit-defensible AI-assisted workflow actually looks like.

1. Why the Question Is Becoming Urgent

RTOs are under real and sustained pressure on assessment workload. Growth in online delivery, rising student numbers, and the spread of written assessment components across training packages have created a marking burden that strains assessor capacity in any organisation conducting genuine assessment rather than superficial confirmation of completion. AI marking tools answer that pressure attractively: they can process a written response in seconds, return a structured evaluation against criteria, flag gaps, suggest feedback language, and produce a preliminary judgement an assessor can review and finalise. The speed gain is real, and so is a certain consistency gain, in that the same criteria are applied to every response without the fatigue and drift that affect a human working through a large batch.

At the ASQA webinar behind this series, two participants asked directly whether AI may perform assessment so long as the output is supervised and checked, and whether the use of AI for significant judgements by large companies sets a precedent VET should follow. ASQA did not give a flat yes or no, which is the correct regulatory position: the Outcome Standards do not categorically prohibit AI involvement in assessment. What they do is set specific requirements any assessment system must satisfy, and those requirements draw a line that AI tools, in their current form and typical deployment, cannot cross unassisted.

2. The Assessment System Architecture: Standards 1.3 and 1.4

Standard 1.3 sets the overarching obligation: the assessment system must be fit for purpose and consistent with the training product, and assessment tools must be reviewed before use to ensure assessment can be conducted in a way consistent with the principles of assessment and the rules of evidence. The consequence for AI is immediate. An AI-assisted marking workflow must be reviewed against the Standard 1.4 framework before it is deployed, not after.

Standard 1.4 then operates in layers. Standard 1.4(1) is the outcome: the assessment system ensures assessment is conducted in a way that is fair and appropriate and enables accurate assessment judgement of student competency. Standard 1.4(2)(a) sets the four principles of assessment: fairness, flexibility, validity and reliability. Standard 1.4(2)(b) sets the rules of evidence, and it is the single most important provision for the AI question, because it opens with the words that assessors make individual assessment judgements that are justified based on the rules of evidence: validity, sufficiency, authenticity and currency. Three elements of that phrase carry the whole analysis. First, assessors, meaning persons, specifically credentialled persons as defined by the Credential Policy, not systems, tools or algorithms. Second, individual, meaning personal and specific to this student, this evidence, this context, not batch-processed or pattern-matched across a dataset. Third, justified, meaning the assessor can explain and defend the judgement by reference to the rules of evidence. A judgement that amounts to "the AI said so" is not a justified judgement. It is a deferred one, and the deferral is not available.

3. The Credential Policy: Who May Make a Competency Decision

The Credential Policy, which forms part of the Standards and must be read with the Outcome Standards, specifies the credentials required for particular training and assessment roles, and its bearing on AI marking is direct. Section 1A specifies the credentials required to deliver training and assessment without direction, including making assessment judgements: the Certificate IV in Training and Assessment or its successor, earlier versions of the Certificate IV, or a diploma or higher-level qualification in adult education or VET, among others. The decisive words are "including making assessment judgements". The Policy treats the making of an assessment judgement, the determination of whether competency has been achieved, as a distinct function requiring a specific credential. It is not a clerical task that any staff member, any software platform, or any entity without the credential may complete.

The Instrument frames an assessment judgement as a determination of whether a student has achieved competency consistent with the training product. A competency determination produced by an AI platform is not made by a person, and so is not an assessment judgement within the meaning of the Credential Policy. It is a processing output that may inform a judgement; it cannot be one. The point is reinforced by how the Policy treats people working under direction: such a person may deliver training and collect evidence but cannot make assessment judgements. If a credentialled human working under direction is not permitted to make the determination, an uncredentialled AI plainly cannot. The rationale is not legalistic. The credential exists because assessment is a professional function requiring knowledge of the training product, understanding of the cohort and context, the exercise of professional discretion, and accountability for the outcome. An AI platform holds no credential, cannot be held accountable for a wrong competency decision, and cannot exercise that discretion. Standalone AI marking bypasses every one of those things.

4. The Four Principles Applied to AI Marking

Evaluating an AI-assisted workflow against each of the four principles in Standard 1.4(2)(a) is part of the pre-deployment review Standard 1.3 requires.

On fairness, which under Standard 1.4(2)(a)(i) requires assessment to accommodate student needs, including reasonable adjustments and reassessment, AI raises real concerns. Language models reflect the linguistic and cultural patterns of their training data, so a student whose first language is not English, who writes in a non-standard register, or whose response style differs from the dominant patterns may be evaluated on communicative difference rather than competency. Reasonable adjustment compounds this: where a student has a disability affecting the form of their written expression, a system that treats writing quality as a proxy for knowledge may systematically disadvantage them, and recognising when a non-standard response nonetheless demonstrates competency is exactly the professional act an independent AI decision cannot perform. The fairness principle, and the broader reasonable-adjustment obligation that also flows from the Disability Standards for Education 2005, requires controls an unsupervised AI workflow does not have.

On flexibility, which under Standard 1.4(2)(a)(ii) requires assessment appropriate to the context, training product and student regardless of how the skills were acquired, AI is by design pattern-matching a response against an expected model. A student who acquired competency through ten years of industry experience may express it differently from one trained in a classroom, and a flexible system values both; a model prompted to score against a standard template may not.

On validity, which under Standard 1.4(2)(a)(iii) requires practical application components allowing demonstration in a practical setting, AI marking applies most naturally to written knowledge tasks, where, used to support an assessor's review of written knowledge evidence, it need not offend the principle. The danger is where written responses are standing in for practical demonstration. That is a validity problem that predates the AI question, and automating the marking of an already-invalid assessment simply produces invalid judgements faster and at greater scale.

On reliability, which under Standard 1.4(2)(a)(iv) requires evidence to be interpreted consistently and outcomes comparable irrespective of assessor, lies the tool's most superficial appeal. The consistency is real but conditional, because an AI system is consistently wrong in the same ways: poorly specified criteria are applied poorly and uniformly, systematic biases are applied uniformly, and a task the model cannot evaluate against the training product's real requirements fails uniformly. Reliability in the Standard's sense means outcomes are comparable because assessors are genuinely evaluating competency. Identical output for identical input is not the same thing.

5. Where the Line Falls: Permitted, Restricted and Prohibited

The analysis above is the legal test. The table maps it onto the scenarios an RTO is most likely to encounter.

6. Documenting the Supervised Workflow

An RTO deploying AI marking under the permitted or restricted scenarios must document the workflow so it demonstrates Standard 1.4(2)(b) compliance. The documentation provides the audit trail showing how the individual-judgement requirement was met, and tells assessors exactly what they must do to discharge their obligations.

The assessor's active review obligation must be explicit: when AI provides a preliminary evaluation, the assessor must read the student's full response independently, apply professional judgement to the criteria, form their own view of competency, and make a determination that reflects their own assessment, which may or may not align with the AI. The policy must state that accepting an AI output without independent review breaches both Standard 1.4(2)(b) and the Credential Policy. Better still, the system should build engagement into its architecture, for instance by requiring the assessor to record an independent preliminary view before the AI output is displayed, rather than relying on policy alone.

The divergence record is the workflow's most valuable quality output. When an assessor reaches a different conclusion from the AI, the file should record what the AI indicated, what the assessor determined, and the reasoning for the difference. Consistent divergence on a particular qualification, cohort or task type is data about the tool's limits, and it must feed the Standard 4.4 continuous improvement cycle. The opposite pattern deserves equal scrutiny: an assessor who never diverges across a large sample is either using a uniformly accurate tool, which should be documented and reviewed, or not exercising genuine independent judgement, which is a Standard 1.4(2)(b) failure. Monitoring the divergence rate and investigating both extremes is a supervisory responsibility under the Standard 3.1 workforce obligations.

The competency determination record must identify the credentialled person who made the determination, not the tool that provided input: the assessor's name and credentials, the date, a summary of the evidence reviewed, the outcome, and, where AI was used, a note that its output was reviewed and that the determination reflects the assessor's independent judgement. That keeps accountability where the Standard places it.

7. Does AI Produce Reliable Outcomes?

Because Standard 1.4(2)(a)(iv) requires comparable outcomes, AI in marking raises an empirical question that must be answered before deployment: does the tool produce outcomes comparable to those a credentialled assessor would reach independently, and does its use improve or degrade the system's reliability? The Standard 1.3 pre-use review includes verifying that any AI component can be used consistently with the reliability principle, so an RTO that deploys a tool without testing this has not done the review the Standard requires.

The pre-deployment check should include a pilot in which the tool and a credentialled assessor independently evaluate the same sample of responses, a comparison that surfaces systematic divergences, particularly any correlated with cohort characteristics such as first language, age or industry background, and a concordance threshold below which the tool will not be deployed. Post-deployment, the Standard 4.4 continuous improvement obligation requires the RTO to keep monitoring: divergence rates, appeal rates and assessor feedback about the tool are exactly the kind of data the improvement cycle should collect and act on. A tool not subject to this monitoring is not being governed as Standard 4.4 requires.

8. The Multinational Precedent Argument

The webinar question about large companies using AI for significant judgements deserves a direct answer in two parts. The first is legal. VET assessment in Australia is governed by a specific instrument, the Outcome Standards made under the National VET Regulator Act, which sets the standard against which assessment must be conducted. A company using AI for recruitment screening or risk modelling is not subject to that instrument; it operates under different frameworks, accountability structures and risk profiles, and its choices do not change what the Outcome Standards require of RTOs. The second part concerns the nature of the decision. A VET competency determination is not a prediction about who will perform best or which application carries least risk. It is a determination that a person has demonstrated the skills and knowledge in a nationally endorsed training product to a standard the AQF recognises, and it founds a qualification that employers, regulators and other institutions rely on as evidence of competency. The integrity of that determination matters in ways many corporate AI decisions do not. This does not mean VET cannot incorporate AI; it means the evolution must be designed around the legal and ethical requirements that govern VET assessment, not around what other industries do in legally distinct contexts.

Conclusion: Keep the Judgement Human, Let the Tool Assist

The position the Outcome Standards make available is generous enough to capture most of the efficiency AI marking offers. AI may be used freely before assessment, to design tasks, draft rubrics, build feedback banks and surface design weaknesses. It may be used in a supervised role during marking, to flag gaps, suggest feedback the assessor adapts, and check consistency across already-marked work for supervisor review. What it may not do is make the competency determination the assessor then adopts unread, mark practical observation records as the primary evidence, judge whether the authenticity rule in Standard 1.4(2)(b)(iii) is met, or stand in for the credentialled assessor's judgement wherever the Credential Policy requires that judgement to be applied. The test to apply at every step of the workflow is a single question: at this point, is a credentialled person exercising individual professional judgement, or is an uncredentialled system making or substantially determining a competency decision? An RTO that stays on the right side of that question, documents it, trains its assessors in it, and monitors it through continuous improvement is using AI in assessment in a way that is both useful and defensible. The efficiency is worth having. The judgement stays human.

References and Further Reading

National Vocational Education and Training Regulator (Outcome Standards for Registered Training Organisations) Instrument 2025: Standard 1.3 (review of assessment tools before use); Standard 1.4, comprising 1.4(1) (the assessment outcome), 1.4(2)(a) (the principles of assessment: fairness, flexibility, validity, reliability) and 1.4(2)(b) (individual assessment judgements and the rules of evidence: validity, sufficiency, authenticity, currency); Standard 1.5 (validation); Standard 3.1 (workforce); and Standard 4.4 (continuous improvement). Federal Register of Legislation.

Standards for RTOs 2025 Credential Policy, Section 1A (credentials to make assessment judgements) and the provisions on delivery under direction. National Training Register (training.gov.au).

Disability Standards for Education 2005 (Cth), reasonable adjustment in assessment. Federal Register of Legislation.

Australian Skills Quality Authority. Practice Guides on assessment, validation and the Credential Policy. https://www.asqa.gov.au/rtos/2025-standards-rtos/practice-guides

This article is general guidance on the Standards and is not legal advice.