A conversation at yesterday’s Nova Education Trust INSET stayed with me: AI can offer pupils a useful explanation, but an answer isn’t understanding. Rather than simply “use it” or “ban it”, what routines help pupils try first, check claims and explain their own thinking?

Building Classfolio in public should mean showing the checks, not just the features: a real backup restore, a privacy gap found and fixed, and who sees a pupil message before it appears. Still pre-launch. What would you ask a classroom-software supplier to demonstrate?

We had backups. We hadn’t tested recovery. A real database restore took 13m 58s, but exposed a missing protection on the restored copy. That isn’t a 14-minute promise for the whole service. The test and its limits: classfolio.co.uk/blog/backups...

I checked Classfolio’s AI marking against 350 real pupil attempts. It was consistent — and I still wouldn’t let it decide.

Written answers remain “needs review” until a teacher agrees, changes or overrules them.

Accuracy and authority are different things.

Where would you draw the line?

Classfolio blocks a lesson launch when it finds an activity pupils cannot use. It does not just warn you: it stops and links to the exact slide.

Tempting to soften. But the failure is not a dialog box. It is 30 pupils watching you improvise.

Assessment technology should make the next teaching decision clearer. In week one, that might mean spotting who needed prompting, where explanations broke down, or which misconception appeared across the room. A fuller dashboard is not the same as less workload. What would you want to see?

Safe class spaces need guardrails built in.

Classfolio Class Spaces:
• teacher approval for every pupil post/reply
• no pupil DMs
• staff-only uploads
• recorded moderation

Safety shouldn't rely on enabling a setting. What are your non-negotiables?

classfolio.co.uk/features/class-spaces/

Teachers: where does technology create the most friction in an ordinary lesson?

Moving from slides to a quiz? Collecting work? Marking? Recording evidence? Giving pupils somewhere useful to revisit it?

I’m interested in the awkward hand-offs between tools—the bits that quietly cost five minutes.

I expected AI marking to be inconsistent.

So before trusting it I marked 350 pupil attempts myself and compared. One assessment, one week.

It matched me. 350 times.

I built it, and still made no AI mark final without a teacher:

classfolio.co.uk/blog/what-ai...