How do you stay in control of your marking when using AI?

What does staying in control of marking with AI actually mean?

Staying in control means the software drafts and you decide. An AI marker reads a student response against a set of criteria, then proposes where it sits and what to say about it. Control is the step after that: you read the draft beside the student’s work, correct what it has misread, reshape the wording, and confirm or change the mark. Nothing reaches a student until a teacher has approved it.

This sits closer to normal department practice than it first sounds. Most faculties already mark from a shared guide, and many run a second read on scripts near a boundary. The drafting is the slow, mechanical part. The deciding is the professional part. Under the Australian Professional Standards for Teachers, assessing student learning and providing feedback on it sits with the teacher, and that responsibility does not move to a piece of software because the software wrote the first sentence.

The useful test is a single line: at every point where a judgement is made, a teacher makes it. Write that down before you use any tool at scale, and be specific about what it means for your faculty. Teachers who lose control rarely decide to. They start with a clear boundary, then let a deadline move it one script at a time.

How do you anchor an AI marker to your own standard?

Give the tool the same materials you would use marking by hand. That means your rubric with its band descriptors, the success criteria you gave students, and the comment phrasing you actually use. A general chatbot with no rubric will assemble a standard out of whatever it has read, and that standard will not be yours. It will usually reward polish and length, because that is what most writing it has seen rewards.

Anchoring also changes the review job from something hard into something fast. Instead of asking whether this is a good mark, you ask whether it matches your descriptor. The second question has an answer you can point to on the page. And because the inputs are yours, you can change them whenever your judgement says so: tighten a criterion, rewrite a phrase, draft again.

There is a side effect worth having. Vague descriptors that human markers quietly paper over with shared assumptions become obvious once a machine applies them literally. If drafts keep over-crediting structure and under-crediting analysis, the wording of that descriptor is the problem, not the tool. Fixing it improves your hand marking too.

• Your rubric or marking guide, with the band descriptors written out in full
• The success criteria students were given, so feedback targets the task you actually set
• Comment phrasing in your own voice, so drafts do not read like a stranger wrote them
• Two or three annotated exemplars showing what each band looks like in practice
• Task context: the question, the conditions, and anything you told the class on the day

What should you check before you approve AI-drafted feedback?

Check that every comment and every mark is supported by evidence in the work in front of you. Read the response and the draft side by side. Models misread subtle arguments, credit quotations a student never used, and praise structure that is not there. They do it fluently and confidently, which is exactly why skimming a draft is not a review. Ten seconds with the actual script catches most of it.

Then sense-check tone and fairness. Feedback should be specific enough that only this student could have received it, and the written comments should agree with the band you have awarded. Watch for length bias in particular: long, tidy answers tend to read as stronger than they are, and a short precise one can be underrated. A deliberate review of each draft is still far quicker than marking from a blank page.

• The mark is justified by the descriptor and by evidence in the response
• Comments refer to what the student actually wrote, not a plausible assumption
• The written feedback and the awarded band tell the same story
• At least one comment gives the student a concrete next step
• Comparable responses across the class have been treated the same way

Which marking decisions should always stay human?

High-stakes, borderline and genuinely unusual work. Marks that feed a formal record or contribute to certification, results sitting on a grade boundary, and responses that take an approach you did not anticipate all need interpretive judgement that someone is accountable for. These are also the cases where a quick draft is least reliable, because they are the least like anything the model has seen before.

That does not rule AI out of those tasks. A draft can still give you something to argue with, and a tool that flags where it is uncertain is useful. But the closer a decision sits to a student’s formal record, or to their sense of their own ability, the more firmly the call belongs to the teacher who knows the class, the context and what the student was reaching for.

Two other categories are not marking decisions at all. Anything that raises a wellbeing concern, and anything that looks like an academic integrity issue, goes to a person and follows your school’s process. A tool’s confidence is not a substitute for either conversation.

How do you keep AI-assisted marking consistent and defensible?

Consistency comes from applying one anchored standard to every student, then moderating a sample. A tool running the same rubric across a whole set does not drift the way a tired human does at eleven at night on the last script. Your job is to check that the consistency it produces is the consistency you wanted. Pull one script from each band plus every borderline case and re-read those properly. If your judgement disagrees with the draft on those, fix the setup before you trust the rest.

Defensibility is being able to explain a mark to a student, a parent or a moderation meeting. Keep enough of a record to do that: which rubric you used, and where you overrode a draft and why. Achievement standards in the Australian Curriculum give you the shared reference point. Read your school and sector policy as well, alongside the Australian Framework for Generative AI in Schools, which sets the national expectations schools are asked to work within. The same discipline applies to anything else you produce with AI for students, from practice questions to study guides.

How does a purpose-built marking tool fit into this?

Everything above works with any tool, including a general chatbot and a careful prompt. What a purpose-built marking tool changes is where the effort goes. Anchoring and review stop being habits you have to remember and become steps in the workflow. Jeddle takes that approach: JeddAI marks against the rubric, success criteria and comment banks you supply, produces a provisional mark with drafted comments, and holds them until you edit and approve. Nothing is released to a student on the tool’s own authority.

Ask the same four questions of anything in this category before you commit a faculty to it. Can I supply my own rubric, in my own words? Can I edit every comment before release? Is there a step where nothing reaches a student without my approval? Can I see which criterion a mark was based on? A tool that cannot answer all four is offering nominal control rather than real control. If you are weighing up Jeddle for expert essay marking, run one class set through it and compare your overrides against the drafts. Two hours of that tells you more than any feature list.