A teacher has thirty essays to mark, a class to prepare and a familiar dilemma: give each student thoughtful feedback, or finish the pile before tomorrow. The problem is not that teachers lack judgment. It is that good judgment takes time, and time is unevenly distributed across a school day.
AI grading products promise to change that equation. Edexia, an Australian assessment platform, presents a particularly interesting model: use AI to prepare rubric-based marks and comments, then require a teacher to review and approve what students receive. It is a different proposition from asking a chatbot to grade an essay and treating the output as final.
Can Edexia reinvent assessment? It might help make some feedback faster and more consistent, but that is not the same as proving that students learn more. Its most promising contribution is a workflow in which teachers can use AI as a second reader, preserve professional judgment and return their attention to the learner. The case for adoption depends on the quality of feedback, the validity of grading, privacy safeguards, local rules and what teachers actually do with the time they recover.
What is Edexia, and who is it for?
Edexia is a teacher-facing AI marking and feedback platform with a strong emphasis on senior-secondary English assessment in Australia. Its published materials describe alignment with VCE English rubrics and support for other Australian senior-school pathways and IB English. Rather than presenting itself as a universal answer engine, it organises assessment around specified criteria, examples and curriculum context.
A teacher can submit typed work or scanned handwriting. The platform produces criterion-level marking suggestions and comments attached to passages of student writing. The teacher can then revise grades, delete comments, add their own observations or voice notes, and decide when feedback becomes visible to the student. According to Edexia's workflow description, release follows teacher review and approval.
That distinction is critical. It means the intended end user of the first AI output is not the learner. It is the educator who must interpret the work in context.
For a school considering the product, the first question should therefore be neither “How fast does it mark?” nor “How impressive is the demo?” It should be: Does this process help our teachers make better educational decisions?
The four stages of AI-assisted marking
Edexia describes a straightforward sequence.
Collect the evidence. A teacher uploads essays or scans handwritten work. Transcription creates one more point where mistakes can enter, so comparing the text with the scan matters.
Generate a first assessment. The system applies the relevant rubric and suggests marks and passage-specific feedback.
Exercise professional judgment. The teacher checks whether the evidence supports the grade and whether the feedback is accurate, fair, comprehensible and helpful.
Release and teach from the result. Students receive approved comments and next steps; teachers can identify patterns that deserve further instruction.
The fourth stage matters most. A class does not become better at writing merely because it receives more annotations. Feedback has educational value when learners can understand it, act on it and transfer the resulting improvement to another task.
What does “learns from teachers” really mean?
The phrase sounds like a model that continuously trains itself on teachers' marking decisions. That is not an accurate description of every form of AI adaptation, and it should not be assumed here.
Edexia's current FAQ says that it does not use school submissions to train or fine-tune its underlying AI models. Instead, the platform uses assessment instructions, rubrics, relevant curriculum context, approved exemplars and teacher configuration to construct feedback. The FAQ also describes a transparent, reversible calibration option for eligible assessments, applied by an authorised lead marker to unreleased AI marks.
These are three different operations:
| Operation | What it means | Why it matters |
|---|---|---|
| Rubric alignment | Producing judgments in relation to explicit criteria and performance descriptors | Reduces the risk of judging the wrong construct |
| Local calibration | Adjusting how unreleased marks are interpreted for an eligible activity under authorised teacher oversight | Can support moderation without silently changing another school's practice |
| Model training | Updating the underlying model's learned parameters using data | Raises distinct questions about data use, consent and governance |
The practical lesson is to avoid treating the word learning as a technical guarantee. Ask the provider exactly which settings change, who can change them, whether changes are reversible, whether decisions are logged and whether submissions ever contribute to model training. A responsible school should obtain those answers in writing.
This is also a useful lesson for teachers evaluating any AI product: the most reassuring explanation is rarely a vague promise that “the AI improves over time.” It is a precise account of what changes, who controls it and how the effect is measured.
How accurate is Edexia's grading?
Edexia reports a 2025 trial involving 579 VCE English essays at St Bernard's College. On the company's accuracy page, it reports 81.2% exact agreement or acceptance of the grade and 98.3% agreement within one grade band. Those figures are worth examining, but they need the right interpretation.
A percentage is not self-explanatory. In the method Edexia describes, teachers reviewed proposed grades and accepted or changed them. That measures agreement with teacher decisions in a particular setting; it does not demonstrate that the grade was objectively correct in every case, nor that student attainment improved after feedback. It also does not establish equivalent performance across subjects, age groups, countries or assessment types.
The company describes the trial as independent, but the public summary and figures are published by the vendor. They should be treated as vendor-reported evidence until a school has reviewed the full design, comparison method, sample characteristics and any independent analysis. In particular, a school should ask whether teachers formed their own judgments before seeing the AI's proposed mark. Seeing an AI suggestion first can influence the subsequent human decision, a risk known as anchoring.
It is also useful to separate three outcomes that marketing discussions often merge:
Agreement: How often do the AI and teacher give the same mark?
Efficiency: Does total marking time actually decrease, including correction and checking?
Learning: Do students make stronger revisions and perform better on later, independent work?
Even excellent agreement would not automatically establish efficiency or learning impact. A system might match grades but produce generic comments. Another might generate helpful questions yet require so much editing that teachers save little time. Schools need evidence on the outcome they actually care about.
The biggest opportunity is not marking essays faster
Imagine two teachers receiving AI-generated observations on an essay. The first reads the proposed grade, corrects a few sentences and releases it. The second studies the errors, asks why several pupils struggled with the same paragraph structure, and plans a short lesson in which learners practise building an argument from evidence.
Both used the technology. Only the second clearly turned assessment into a teaching decision.
This is the difference between feedback production and feedback use. Technology can increase the quantity of feedback more easily than it can guarantee that students understand or apply it.
Consider the comment: “Your analysis lacks development.” It may describe a real weakness, but it gives a pupil little to work with. A stronger response might identify a specific paragraph, show which claim is unsupported, and ask the learner to choose one quotation and explain how it changes the reader's interpretation. The teacher can then check the revision and eventually remove the scaffold.
For a writing class, a useful feedback cycle might be:
The learner identifies one personal goal before writing.
The teacher defines two or three criteria that are genuinely teachable.
AI drafts feedback, which the teacher checks against the original work.
The learner selects one comment and makes a visible revision.
The learner explains why the revision is stronger.
On a new task, the learner attempts the same improvement without being prompted.
This sequence gives the assessment a future-facing purpose. It measures not only the quality of a submitted essay but the learner's growing capacity to improve independently.
Why human review is essential, not ceremonial
The value of teacher approval depends on whether review is substantive. A teacher who automatically clicks “accept” has not meaningfully exercised oversight.
Five errors are especially worth watching.
First, inaccurate textual evidence. AI may misread a quotation, misattribute a technique or overlook contradictory context. Before endorsing a comment, the teacher should be able to point to the passage that supports it.
Second, rubric drift. A comment may sound intelligent while assessing fluency or vocabulary more heavily than the actual criterion requires. Compare the proposed grade to every descriptor, not to the polish of the explanation.
Third, overconfident language. Student writing can support more than one defensible interpretation. A narrow algorithmic judgment may discourage original analysis if teachers do not protect legitimate differences of interpretation.
Fourth, inconsistent treatment of students. Dialect, multilingual expression, spelling differences, disability-related access needs and the quality of handwriting transcription can affect what the system sees. A school must investigate whether certain groups experience systematically different errors, while respecting privacy.
Fifth, automation bias. People can defer to a numerical recommendation because it appears neutral. One practical safeguard is for the teacher to mark a sample of scripts independently before viewing AI suggestions, then compare the two decisions and discuss disagreements.
Human oversight is therefore an active practice: checking evidence, asking questions, correcting errors and retaining responsibility for consequential decisions.
Can AI grading be fair across different kinds of writing?
A rubric gives a shared language for judgment, but it does not eliminate ambiguity. In literary analysis, for example, the quality of an interpretation depends on what the student argues and how the argument is supported. An inventive, defensible reading may differ from a model example.
This is where teachers should protect the difference between meeting a criterion and matching an expected answer. An AI grader should not reward conformity at the expense of genuine reasoning.
Handwritten assessment adds another layer. A pupil who writes small, makes extensive deletions or uses a nonstandard layout may be harder to transcribe correctly. If an AI system evaluates an inaccurate transcription, the resulting mark can appear precise while being built on faulty input. Verification needs to begin with the text itself.
A school trial should deliberately include varied scripts: concise and extended responses, high- and low-scoring examples, original interpretations, multilingual writing and handwriting samples. The aim is not to search for a flattering average, but to identify where the system fails and who bears the cost.
Formative assessment and high-stakes grading are different jobs
For a practice essay, the most valuable output may be a question that prompts a better draft. The student can challenge the comment, revise and learn. A mistaken suggestion, if identified quickly, can become a useful discussion.
For a final grade, the stakes are different. Scores may affect qualifications, opportunities and family expectations. The required standard of documentation, moderation and accountability rises accordingly.
Edexia itself distinguishes teacher-reviewed formative work from consequential assessment in its guidance for schools. Schools should not assume that because a product technically supports a qualification's rubric, its use is automatically permitted by the jurisdiction or awarding body.
Australia's Framework for Generative AI in Schools provides national guidance, but local system, sector and assessment-body requirements still need checking. International schools should apply their own laws, policies and examination rules rather than copying a product's Australian defaults.
A proportionate starting point is low-stakes, teacher-reviewed feedback on practice work. Before considering summative uses, leaders should document which decisions may be AI-assisted, which must remain independent, what records are retained and how families can query a disputed result.
What about student privacy and data protection?
Essays contain more than answers. They can include names, classroom context, personal experiences, identifiable incidents and information about learning needs. Some students reveal sensitive experiences in creative or reflective writing. Uploading that material to a platform is a governance decision, not a casual productivity choice.
Edexia's privacy materials state that school data is segregated, that uploaded work is not used to train its models, and that it uses Australian data storage and de-identification before external AI processing. Its transparency policy also makes clear that AI outputs can be incomplete or wrong and require educator review.
These are useful claims to investigate, not substitutes for a school's due diligence. An institution should confirm:
Which personal information is collected, and for what purpose?
Which subprocessors receive assessment content, and in which countries?
Can the school control retention periods, deletion and access?
How are handwritten scans and information written inside them handled?
What contracts, audits and incident-response arrangements can the provider document?
What notice or consent is appropriate for students and families under local rules?
UNESCO's guidance on generative AI in education argues for human-centred safeguards, data protection and age-appropriate use. That principle applies equally to impressive classroom software and ordinary administrative tools.
No teacher should test a new system using identifiable student work until school approval and data-protection requirements have been addressed.
A practical school pilot: measure learning, not just speed
A small, carefully designed trial can be more informative than a large deployment driven by enthusiasm. The following pilot is a suggested evaluation approach, not a report of Edexia use by Learning Coach.
Week 1: establish a human baseline
Choose one department and a low-stakes writing task. Agree on a fixed rubric, select representative samples, and have teachers mark a subset without AI suggestions. Record the amount of time spent, examples of disagreement and the comments teachers consider useful.
Do not use student names in an experimental dataset. Obtain institutional approval before uploading any work.
Week 2: test teacher-reviewed AI support
Where permitted, process comparable pieces of work through the platform. Teachers inspect each suggestion, log changes, check handwriting transcription where relevant and record total review time. Compare the resulting comments with the original human-only feedback.
Ask not simply whether the grades match, but whether comments are specific, accurate, appropriately challenging and written in language learners can use. Disagreements should become material for departmental moderation rather than being hidden as inconvenient exceptions.
Week 3: test what students can actually do
Give pupils time to revise one passage after receiving teacher-approved feedback. Collect their explanations of the changes they made. Later, use a short independent writing task to check whether the targeted skill transfers.
You can now examine several different measures.
| Measure | Useful question |
|---|---|
| Grade agreement | Where do teacher and AI judgments diverge, and why? |
| Full workload | Does AI save time after review, correction and administration? |
| Feedback quality | Can students understand and act on the comments? |
| Revision behaviour | Do learners make evidence-based improvements? |
| Transfer | Can students apply the skill on a new task without the same prompts? |
| Equity and safeguards | Are errors or access burdens concentrated in particular groups? |
Set thresholds before the trial begins. If the system saves time but repeatedly undermines original interpretations, the saving is not sufficient. If it improves the quality of dialogue but does not reduce workload, leaders can still consider whether that outcome justifies the cost. Neither decision should be disguised as a universal verdict on AI.
Evidence about AI-assisted marking more broadly is still developing. A Jisc National Centre for AI report published on November 3, 2026 describes a pilot in postsecondary settings where discussion moved beyond workload reduction toward sustainably improving feedback quality. Its context differs from Australian school English marking, but the evaluation question is relevant: What kind of educational improvement is the tool helping staff achieve?
From automated comments to a teacher-as-coach approach
Learning Coach advocates moving beyond the simple transfer of information toward building the learner's awareness, resources and independence. That requires teachers to notice patterns in thinking and ask questions that students can eventually ask themselves.
Suppose an essay is rated weak in evidence selection. A traditional response might be a red mark beside a quotation. An AI-generated comment might say the quotation is insufficiently analysed. A teacher-as-coach can go one step further:
“What claim are you making here?”
“Which part of the quotation actually supports that claim?”
“What would a reader need explained before accepting your interpretation?”
“How will you check this for yourself in the next essay?”
These are not just friendlier comments. They transfer responsibility for the thinking process back to the learner.
Teachers can also use a quick conversation to separate a knowledge gap from a strategy gap. Perhaps the student knows the text well but does not know how to structure an argument. Perhaps the structure is fine, but the evidence does not support the interpretation. A machine-generated score alone cannot determine what kind of help will be most productive for that individual without context and dialogue.
This approach connects naturally to the teacher as learning coach: clear goals, nonjudgmental inquiry, awareness of the learner's resources and progressive independence. It does not imply that Edexia itself provides coaching, nor that its use automatically develops autonomy. Those outcomes depend on what educators design around the tool.
Educators who want to strengthen those human skills can explore the Learning Coach training pathway, which introduces a broader coaching approach rather than promising that any software can replace professional relationships.
What should a school leader ask before buying?
An attractive interface is not a procurement case. Ask for a demonstration using a task and rubric from your own setting, subject to privacy approval. Require the vendor to show where teachers can override marks, correct comments, view change history and control release.
Ask which curricula are truly supported today, rather than which appear on a roadmap. Request documentation for the trial outcomes most relevant to your learners. Verify pricing and contracts directly rather than relying on a blog article or an old screenshot.
Most importantly, decide where the product fits into professional learning. If staff do not have time to moderate disagreements, discuss feedback quality and observe what pupils do next, the technology may simply accelerate an assessment routine that needed redesign.
A strong pilot includes teachers, leaders, students and privacy or IT representatives. Their questions are different but complementary. Teachers want trustworthy workflow; leaders want fairness and accountability; students want understandable feedback and the opportunity to improve; families want reassurance that a person's learning is not reduced to an opaque automated judgment.
Is Edexia right for teachers outside Australia?
Edexia's strongest published case concerns its Australian curriculum-aligned assessment workflows. That focus may be a feature for the intended audience and a limitation elsewhere.
A US middle-school teacher, a UK primary school or a French international school should not assume that VCE-specific accuracy figures apply to its own writing standards. Before using the product, confirm actual curriculum availability, eligible account types, local compliance requirements, languages, feedback formats and data-transfer arrangements.
This does not make the underlying idea irrelevant internationally. The design principle is portable: define valid assessment criteria, let technology generate provisional feedback, preserve teacher control, and evaluate whether students learn from the result. The product's suitability must still be assessed locally.
Frequently asked questions
Does Edexia replace teachers who mark essays?
No. In its documented workflow, Edexia produces proposed grades and feedback that teachers review, modify and approve. The teacher remains responsible for what reaches the learner.
Does Edexia train its AI on student essays?
Edexia says that it does not use school submissions to train or fine-tune its underlying models. It distinguishes contextual rubric use and eligible, reversible activity-level calibration from training on student data. Institutions should verify contract and privacy details for their own account.
Are the reported accuracy figures proof that it improves student achievement?
No. The 579-essay VCE trial concerns agreement or acceptance of marks. It does not on its own establish gains in learning, writing quality or later independent performance.
Can AI mark creative writing fairly?
It can help organise evidence against explicit criteria, but creative and interpretive work requires careful human reading. Teachers should check whether unusual but defensible responses are penalised and whether rubric interpretation remains sound.
What is the safest first use case?
Where policy permits it, teacher-reviewed feedback on low-stakes practice work is a more proportionate starting point than automated consequential grading. Begin with data protection, a baseline and a plan to evaluate learning.
The real test of assessment innovation
Assessment is not simply the act of producing a mark. It is a process for understanding what someone knows, identifying what they need and helping them move forward.
Edexia makes a plausible case for using AI to reduce repetitive elements of marking while keeping teachers in the decision loop. Its rubric-aligned workflow and published agreement figures deserve attention. They also deserve scrutiny, particularly when a school's decisions affect young people's opportunities.
The decisive question is not whether AI can imitate a teacher's mark. It is whether a teacher, supported by a carefully governed tool, can give a student more accurate, timely and usable guidance—and then help that student become less dependent on guidance altogether.
If technology gives teachers more time, the educational gain will come from how that time is reinvested: in conversation, diagnosis, practice, revision and independent thinking. That is the point at which smarter marking could become genuinely better learning.
Resources and references
Edexia — How It Works: current product workflow and teacher approval.
Edexia — Accuracy: company-reported 2025 VCE English agreement figures and methodology summary.
Edexia — FAQ: calibration, model-training claims and data-handling explanations.
Edexia — Privacy and Security: provider disclosures and governance documents.
Australian Government Department of Education — Framework for Generative AI in Schools: policy principles and context.
UNESCO — Guidance for Generative AI in Education and Research: human-centred AI and data safeguards.
Jisc National Centre for AI — When Does AI Improve Marking and Feedback?: postsecondary pilot report, November 3, 2026.
