How do you start using AI marking as a new teacher?

What do you need in place before your first AI-assisted marking run?

Three things decide how useful a first run is: the task your students responded to, the marking materials you want applied, and the student work itself. Nearly all of the quality is settled before anything is generated. A model cannot infer the standard you hold in your head. It can only apply the standard you have written down, which means your rubric and criteria are doing most of the work.

None of it has to be polished to begin. One assignment for one class is enough. If you already mark against a rubric or a set of success criteria that lives in a document somewhere, have it open so you can paste it in rather than retype it. If you have never written your criteria down, this is the point at which you find out how much of your marking standard is tacit.

• The task or question your students actually responded to, in full.
• Your rubric or marking criteria, aligned to the outcomes you report against.
• Success criteria describing what a strong response does, not how good it is.
• A bank of feedback comments you already reuse, if you have one.
• One class set of responses, named clearly and matched to the right student.

How small should the first run be?

Pick one real class and one task you genuinely need to mark this week. Resist the urge to set up your whole timetable. The point of a first run is to find out where the drafts are strong and where they miss, and you can only judge that on work you know well.

The best first choice is a task you have marked before, ideally one where you still have your own marked set in a folder. That gives you a reference point. Without it you are comparing generated feedback against a memory of how you mark, which is not a fair test of either.

Treat the whole assignment as a trial rather than a migration. Once you have seen how the drafts read across thirty responses you understand well, you will have a much clearer sense of which of your other tasks are worth running this way and which are better left alone.

How do you write criteria a machine can actually apply?

Vague descriptors produce vague feedback. “Excellent analysis” gives nothing to check a response against, so anything generated from it will be equally unfalsifiable. Name the observable thing instead: explains how at least two techniques shape meaning, with quoted evidence for each. That version is checkable by a colleague, by a student and by a model.

Your bands also have to be distinguishable from one another. If the only difference between a B and a C in your rubric is the word “thorough” against “very thorough”, you have not described a difference in the writing. You have described a difference in your impression of it. Rewrite band boundaries as changes in what the response does: from naming a technique to explaining its effect, from including evidence to selecting evidence that fits the claim.

Order your success criteria the way a student writes, introduction through body to conclusion, so the feedback arrives in the order they will act on it. Keep a comment bank of the phrases you reuse, so a class hears one voice rather than thirty variations of the same point. It is worth giving students the same criteria before they write, alongside any study guides you already hand out, because criteria that only appear at marking time cannot change the work.

How do you check the drafts before you trust them?

Calibrate before you scale. Mark three or four responses yourself first: one you consider strong, one middling, one weak. Then generate drafts for those same responses and compare. You are not looking for agreement on every mark. You are looking for the direction and the size of the gap, and whether it is consistent.

Most drift shows up at band boundaries. A first pass is often a little generous in the middle of the range, where a response has the right shape but thin evidence, and a little harsh at the top, where a strong answer breaks the expected structure deliberately. Once you know which way it leans, you can correct for it in seconds on every remaining response instead of re-reading the rubric each time.

Read for four specific failures: praise that is not tied to a piece of the student’s own writing, comments that restate the rubric back at the student, a mark that does not match the comment sitting beside it, and a real weakness the draft did not notice at all. Keep a short list of what you edited. That list is your rubric revision for next term, and it is the most valuable thing a first run produces.

Where does the teacher's judgement have to stay?

A generated mark is a proposal, not a result. Nothing should reach a student unread. Assessing student learning and reporting on it sits with the teacher under the professional standards, and changing your workflow does not move that responsibility anywhere else. In practice this means the review step is the job, and the drafting is the part that was taking your evenings.

Check your school or system policy before you upload anything. Sectors differ on which tools may hold student work and where that data may be processed, and the answer is not the same across every state and system. If you are not certain yet, start with de-identified responses or with a task that carries nothing sensitive while you get a proper answer.

Be straightforward with students about the process as well. Telling a class that a first pass was drafted against the criteria they were given, and that you read and edited every comment, tends to increase their trust in the feedback rather than reduce it. It also puts the criteria at the centre of the conversation, which is where you want them.

How does a marking tool built for schools fit into this?

Here is the part where a purpose-built tool differs from a general chatbot, and it is worth being concrete rather than vague about it. Jeddle applies the rubric, success criteria and comment bank you supply, across a whole class in one pass, rather than you pasting one response at a time into a chat window and hoping the standard held from the first script to the thirtieth. If most of your load is extended writing, that is where the case is strongest: expert essay marking is where the drafting hours actually go, and where an evenly applied rubric matters most.

The control points described above are part of the flow rather than left to your discipline. JeddAI drafts a mark and comments against your criteria, you review and edit every one before anything is returned, and nothing is finalised without you. It was built in Australia and is used across Australian and New Zealand secondary classrooms, so the subject and year level framing matches what you report against.

None of that removes the work in the earlier sections. Any tool applies the standard you hand it, so a thin rubric produces thin feedback faster. The setup is the same whichever one you use: one class, one task, clear criteria, a calibration check, then the rest of the set.

What should you change after the first assignment?

The strongest signal from a first run is your own edit history. If you rewrote the same comment on nine scripts, the criterion behind it is underspecified. If you moved marks in the same direction repeatedly, your band descriptors are not separating what you assumed they were separating.

Fix the criteria rather than the individual comments. Consistency comes from the criteria, not from the automation, and a sharpened rubric pays off on every future task that reuses it, including the ones you go on to mark entirely by hand.

• Rewrite any band descriptor you had to explain to yourself more than twice.
• Add the comments you typed repeatedly into your comment bank.
• Reuse the same criteria on the next task of the same type so students see continuity.
• Expand to a second class only once you can predict where the drafts will need editing.