What the AI assumes
The feedback shown is written to show the pattern the study found. It is not real AI output.
1 of 4
It learned from us
Tools like Copilot learn from an enormous amount of human writing. That writing includes our habits: who gets praised, who gets corrected. The tool has no opinions of its own. It repeats our patterns back to us, fluently, and this is what that looks like on one report.
For exampleIt is not neutral because a computer wrote it.
Pick an activity. Each takes two or three minutes.
Three rules for asking Copilot about learners' work, and a check to paste after any answer.
Why this works in your room
Describe the work, not the learner, and the AI has less to assume.
For your learners: Feedback drafted with AI is about their work, held to the same criteria as everyone's.
For you: You get a more useful draft, and you stay the one who decides.
Describe the work
Never the learner.
The task and the criteria.
No names, no labels, no history.
Give it the criteria
So it has a standard to use.
Paste them in.
Ask for one strength, one fix, one next step.
The check
Paste this after any feedback.
Does any of that depend on who the writer is and not on the work?
Rewrite it so it would be the same for anyone who wrote this.
Read it as a draft
You are still the tutor.
Would I say this to any learner who wrote this?
If not, change it.
Watch out for
Telling it more, to be kind.
A kinder label is still a label.
She lacks confidence is a label too.
The next step
One piece of real work this week.
Ask with the criteria only.
Paste the check. See what changes.
Elsewhere in Bradford ThreadWorks
Big 4Writing detailed prompts in Copilot, so the request says exactly what you want.Open Big 4 (opens in a new tab)
Lesson WeaverEDI, built in: how its AI adapts resources for inclusion, and the research behind it.Open Lesson Weaver (opens in a new tab)
Prompts to paste into Copilot
Two prompts that keep Copilot on the work. Change what is in brackets, and put nothing in about who the learner is.
Feedback from the criteria only
Here is a piece of [level] [subject] work: [paste the work, with the name removed]. Here are the criteria: [paste]. Give one strength, one thing to fix and one next step, each tied to a criterion and to a place in the work. Assume nothing about the writer.
Examples with a real range
Write [number] short case studies for [level] [subject] learners about [topic]. Vary the names, ages and backgrounds across all of them, and do not match jobs to the usual gender. Keep each under 80 words. Then list the names and jobs you used, so I can check the range.
The finding
A Stanford team generated AI feedback on 600 essays, changing only the description of the writer. The feedback shifted in line with stereotypes: more praise and less real critique for some writers, with nothing in the essay changed. It has not yet been peer reviewed.
Tan, Phalen and Demszky, 2026 (a preprint) (opens in a new tab)What the guidance says
Government guidance for schools and colleges in England says AI tools can be wrong and can reproduce bias, and that a person must always check what they produce before it is used with learners.
Department for Education, 2025
How sure is this?
Not very. One preprint is not a rule, and the feedback on this page is written by us to show its pattern, not copied from it. It does not mean do not use it. Feedback built from the work and the criteria can be quick and useful.
No source claimed