- Evidence Snacks
- Posts
- Unbundling AI
Unbundling AI
Why blunt verdicts backfire
Hey
It’s been an interesting week! Character building. Anyways, onwards, with a new series on AI and education…
Big idea 🍉

Is AI good or bad for learning? Well, this is not a helpful way to think about it... it’s a bit like asking if intelligence is good or bad: it depends how it’s used.
Ban it and we lose out... teachers who plan lessons with AI save almost 50 minutes a week (with no drop in quality). Evangelise it and we lose out... when students got hold of AI, homework scores rose by 18% while exam scores fell by 20%.
Blunt verdicts on AI tend to backfire.
A better way to think of AI is as having a team of smart, tireless but remote postgrad students on tap. They can do a LOAD of different things, such as:
Talking with learners: from open chatbots that hand over answers to tutors built to make students do the thinking.
Making materials: generating explanations, images, videos, questions, practice games etc.
Checking & responding: marking, giving live feedback, and remediating misconceptions.
Orchestrating learning: tracking what each student knows, and what they should meet next, including sequencing retrieval.
Running a school: triaging emails, generating reports, aggregating data, aligning policies etc.
AND, AI is (currently) jagged in its performance. It’s getting rapidly good at things where answers are easy to check (coding or maths). But likely less so for things which are harder to check (like teaching).
So, if we reframe our original question through this lens... is a team of remote postgrads good or bad for learning?... the answer starts to get clearer. It depends which job we give them, how well we brief them, and how closely we check their work.
In short: to think better about AI (and eventually make better decisions), we need to unbundle it. That’s the first step to making sense of this messy era, and giving our students the best possible deal.
🎓 For more, check out the OECD Digital Education Outlook 2026.
Summary
“Is AI good or bad for learning?” isn’t a helpful question: AI is a load of different things.
Blunt verdicts backfire: bans lose real gains, and evangelism hands students real harms.
Think of AI as a team of remote postgrads: what matters is the job, the brief, and how closely we check.
Little updates 🥕
Research on sleep & academic achievement → finds that changes in sleep duration do not predict subsequent academic performance.
Article on vocabulary learning → suggests retrieval practice & sentence generation improve vocabulary retention more than copying words.
Article on self-testing → argues that questions & feedback may improve factual recall without developing the ability to connect ideas.
Paper on student habits → finds that repeating behaviours increases automaticity, making learners more likely to follow through on intended actions.
Upgrade your evidence edge → Get Snacks PRO
Laters
Jen & Peps 👊