How do you use AI-drafted feedback without losing control of marking?

What makes a feedback draft worth editing?

A generated draft is only as good as the marking material behind it. Hand a tool a one-line task description and you get comments that could apply to any student in the country. Give it a rubric with the levels described in full, success criteria written for this task, and a bank of phrases you already use, and the draft has something specific to point at in the response. The preparation you would do anyway before marking a class set is the same preparation that makes a draft usable.

Be honest about what the tool can actually see. Typed work is read cleanly. Handwritten work has to be scanned or photographed first, and faint pencil or cramped margins cost you accuracy. Diagrams, annotated texts and anything where layout carries meaning need a check by eye before you trust a comment about them. Knowing where the input is weak tells you which drafts to read hardest.

One habit is worth building early. Mark two or three scripts yourself before you look at any draft. Those become your anchors, and they tell you within minutes whether the drafts are pitched near your standard or well off it. If they are off, the fix is almost always in the material you loaded rather than in the individual comments.

• The rubric or achievement standard, with the level descriptors written out, not just the grade labels
• Success criteria for this task, in the words you gave students
• Comment banks of feedback phrases you already use and trust
• The task sheet and any stimulus the response depends on
• Clean scans of handwritten work, one file per student, in reading order

How do you check a draft against an NCEA achievement standard?

Read the draft with the standard open beside it and confirm that every judgement is anchored in something you can see in the student’s work. For NCEA internal assessment the teacher is the assessor, and grades are confirmed through national moderation, so a suggested Merit only stands if you can find Merit-level evidence yourself. Treat the draft as a colleague’s first opinion that you verify, not a verdict you accept.

The checks are the ones you already run in a moderation meeting. Has the draft been too generous or too harsh? Does it point at the part of the response it is describing? Do the borderline cases sit on the right side of the line? If a moderator would ask where the evidence for this grade is, the draft should already answer that. If it does not, you fix it before anything is recorded.

Read scripts against each other as well as on their own. Where two students show similar evidence, the drafts should land on similar grades, and where they diverge, that is a prompt to look at both again rather than to split the difference. Cross-checking a handful of scripts in each grade band is the quickest way to catch a run of drafts that has drifted.

How do you edit a draft so it sounds like your own feedback?

Rewrite until the comment reads the way you would have written it for that student. Drafts default to a neutral, even tone, which is safe and forgettable. You know which student needs encouraging and which one is ready to be pushed, and that judgement is not available to software. Keep what is accurate, cut the padding, and change the register.

Then add the part that makes feedback useful: one concrete next step the student can act on before the next piece. Not three, and not a list of everything that went wrong. If the draft has produced five observations, decide which one will move this student furthest and lead with it. Drafting is the mechanical half of marking. Choosing is the half that teaches.

Watch for the small things that give a draft away. Comments that praise without naming what was good, next steps written in rubric language rather than student language, and the same sentence appearing in four different scripts. Those are quick to fix once you know to look, and fixing them is what stops the feedback reading as if nobody was home. Patterns that show up across the whole class are worth pulling into your next lesson or into the study guides students revise from.

What routine keeps a whole class set consistent?

Draft the whole set before you review any of it, then review in one sitting with the standard open. Marking a class in a single pass is what keeps expectations steady from the first script to the last, and it is the reason experienced markers avoid spreading thirty essays across five evenings. When the first draft already speaks the language of your criteria, holding that line takes less effort.

Spend the time you save on the scripts that need it. Borderline grades, students whose results have moved sharply, and anyone whose response does something the rubric did not anticipate all deserve a slower read. The rest can be checked, adjusted and approved quickly, which is the point.

• Load the rubric, success criteria and comment banks once, at the task level
• Draft the full class set before reviewing any of it
• Review in a single pass so the same standard applies throughout
• Flag borderline grades for a second look before you approve
• Record only what you have read and confirmed yourself

Where should the line between the tool and your judgement sit?

The line is easy to state: the tool drafts, you decide. Any setup where generated comments can reach a student without a teacher reading them first has moved the professional decision to the wrong place. Approval should be an action you take, not a default you have to remember to switch off.

This matters most where the grade carries weight. NCEA internal assessment records you as the assessor, and moderation asks you to justify what you awarded. Keep enough of a trail that you could reconstruct a decision months later: the standard you used, the evidence you relied on, and the final wording that went to the student. Most departments are better off agreeing a short position first, covering who may use these tools, what material gets loaded, and how student work is stored and handled.

How does a marking tool built for this fit the routine?

Everything above works with any drafting tool. Here is how it looks in one built for the job. Jeddle is an assessment platform for secondary teachers, and its marking engine drafts against material you load yourself: your rubric or the achievement standard’s descriptors, your success criteria and your comment banks. Student work is either connected from submissions or uploaded, including handwritten responses as PDFs.

A draft comes back as an overall comment plus criterion-level notes, each tied to the part of the response it refers to, with a suggested mark where you have set one. Nothing is sent anywhere. Every comment and every mark is a proposal you accept, edit or discard, and the student sees only what you approve. Teachers who use it for expert essay marking tend to put the reclaimed time into the borderline scripts and the conversations that follow.

The practical test for any tool in this category is whether it makes your criteria easier to apply consistently without taking the final call away from you. If you cannot see what a comment was based on, or you cannot change a mark in one click, the loop is broken. JeddAI is built around that review step, but the discipline described here is the part that transfers, whatever you end up marking with.