Can ChatGPT write your report comments? An honest audit
Plenty of teachers are already doing it. It sort of works, it fails in predictable ways, and there is one line you must not cross. Here is the full picture, including the workflow that makes it actually useful.
Every staffroom now has someone quietly running their report comments through ChatGPT or Claude. The question is not really whether you can (you can), it is whether the output is good enough to send home, how much work it actually saves once you do it properly, and whether you are allowed to do it at all. We have used these chatbots on our own classes, so this is an audit from practice, not a hit piece: what they are genuinely good at, the workflow that makes them work, and where they fall down.
What a raw chatbot is genuinely good at
- Phrasing. Given the substance, a chatbot turns it into fluent, professional sentences instantly. If your bottleneck is wording rather than knowing what to say, it removes that bottleneck completely.
- Tone conversion. "Make this firmer without being harsh", "say this to a parent rather than a colleague": genuinely excellent, and a legitimate everyday use.
- Unblocking. The blank box at student fourteen is real. A mediocre draft to react against is often faster than a blank page, because editing is easier than composing.
- The awkward ones. It can help you find defensible words for the honest-but-difficult message, the low result, the effort concern, without drifting into either brutality or euphemism.
The workflow that makes it work
Here is the thing the "just use ChatGPT" advice skips: a chatbot knows nothing about your students. Every true, specific sentence in the output has to come from evidence you typed in. So the teachers doing this well have all converged on the same workflow, which is essentially manual data assembly:
- Go through the markbook and write an evidence note per student: strongest topic, weakest topic, anything notable. Some teachers literally do a post-it note per student, or a photo of a handwritten grid.
- Write one careful mega-prompt: the comment structure you want, length limits, tone, banned phrases, your school's conventions.
- Feed it the class in batches, then edit every single comment against your own knowledge of the student.
This works. The comments come out specific, because you supplied the specifics. But notice what happened to the time saving: you are still doing the evidence gathering (the slowest part of report writing), you have added a prompt-engineering step, and you must still edit everything. What the chatbot removed is the typing, which was never the expensive part. The honest description is: better comments for similar total effort, not the same comments for less.
Where it falls down
- It only knows what you typed. Leave the evidence thin and the model does not leave gaps; it fills them, fluently and plausibly. A comment praising "consistent effort in class" about a student the model has never met is fabrication wearing a nice shirt. The thinner your input, the more of the output is invented.
- Praise inflation. Left unedited, every student trends toward "a pleasure to teach". Thirty warm, interchangeable comments are exactly the AI tell parents are learning to spot, and they devalue the comments that earned their warmth.
- Consistency drift. Chat sessions have no memory of your standard. The comment quality and tone you get for student three and student twenty-six can differ noticeably, and the model will happily hand four students the same distinctive phrase, which is fatal if two of them are siblings' families comparing reports.
- No evidence trail. When a parent asks "what is this based on?", the answer needs to be in your markbook, not in a chat log you deleted. A comment you cannot trace to marks is a comment you cannot defend.
- The privacy line. This is the serious one. Pasting full names, school details and academic results into a consumer chatbot is sending student data to a third party without consent, very likely against your school's policy and your obligations under privacy law, and depending on your settings that data may be used to train future models. Schools have disciplined teachers for less. If you do any of this: first names or initials only, never the school, never anything identifying, and check what your employer's policy actually says before you start, not after.
The minimum safety rules if you use a chatbot anyway: initials or first names only; no school name, no year-and-full-name combinations, no health or behavioural detail; turn off chat history/training where the setting exists; edit every comment against your own knowledge; and never send a line you could not defend to the family with your markbook open.
The honest verdict
A raw chatbot is a very good phrasing engine with no evidence and no memory of your standard. The workflow that fixes this (assemble evidence per student, prompt carefully, edit everything, keep names out of it) genuinely produces good comments, and if reports come around twice a year and you enjoy tinkering, it may be all you need. What it does not do is save the time you wanted saved, because the expensive parts of report writing were always the evidence gathering and the judgment, and those stayed on your desk.
We built the post-it workflow so you don't have to
Markpilot is that exact workflow, productised: your marked tests or rubric judgments become the per-student evidence automatically, every comment is drafted from the actual marks (with nothing invented to fill gaps), tone and standard hold steady across the whole class, first names are all it ever needs, and you edit and approve every line. The phrasing engine, with the evidence layer attached.
Start free 15 credits a month free · no card required · built by practising teachers