The Staff-Room AI Test
One job. One boundary. One receipt. Fifteen minutes, then compare notes.
I put three AI systems into the same workstream this weekend. They were fast, capable, and ready to help. They were also perfectly willing to repeat each other's work until I gave them separate jobs.
Teachers have seen this movie before. “Everybody help” sounds collaborative right up until four people edit the same slide and nobody calls the parent.
The classroom version of this problem is the phrase “use AI.” Use it as what? A tutor? An editor? A skeptical colleague? A machine that confidently turns a five-minute task into a new policy meeting?
Give the AI one job
Choose a lesson-planning artifact that you created and that contains no student information. Then assign the AI one narrow role.
“Act as a skeptical teaching colleague. Review these discussion questions for places where students could answer without understanding the reading. Do not rewrite them yet. Identify weak spots, explain why they matter, and tell me what you are least certain about.”
- Keep the source material free of student names, grades, accommodations, and private records.
- Set a timer for 15 minutes.
- Save the first answer and your final version.
- Ask the AI to list its changes and uncertainties.
What you are actually testing
This 15-minute test has a narrow purpose: see whether one clearly named role improves one piece of work without hiding the judgment that belongs to you. The larger argument about whether AI is good or bad for school is too big to settle before first period.
The receipt matters. A polished answer can make weak reasoning look finished. The list of changes, doubts, and rejected suggestions tells you whether the tool helped you think or merely gave you something attractive to paste.
Take five minutes with another teacher
Point to one suggestion that changed the work for the better.
Point to one suggestion that ignored your students, subject, or actual classroom.
Name the decision you would not hand over.
Keep it, narrow it, or fire the intern.
A field note from Jason
The useful moment in my multi-agent experiment came when the work became legible. Named roles exposed duplication, receipts made the handoffs inspectable, and approval boundaries kept the important decisions with me.
That feels like a better starting point for classroom AI than a giant list of approved tools. Tools change. A clear job, a clear boundary, and visible work travel pretty well.
Reply with your result
What role did you test, and what did the AI get wrong?
Two sentences is plenty. If enough people reply, I will share the useful patterns in the next Field Notes.
Until next time: eat your veggies, limit screen time, and never stop learning.
Ciao,
Jason