Tell us about this trial, what you're finding, and what you feel like it might imply for the future of tutoring.

Thanks, Alex.

So I'll start with a kind of a broad introduction so that people understand what we try to do in this trial.

What was the implementation like? So this was our second randomized control trial with Google DeepMind.

We ran a small study last year, last summer, 2025.

Then we ran a larger study this side of the year, so April to July, 2026.

And in a year, a lot changes in the world of AI.

I'd love to hear you talk a little bit about that potential future and how you think about it.

Um, the follow-up question is: which studies have actually shown the cognitive atrophy in young people using AI? How big was the sample? Uh, and maybe it would be amazing if you could share the chess story that I love, that you shared on the podcast to, uh, to, to, to explain and to illustrate this.

Okay, so I, I don't study cognitive atrophy specifically.

Uh, I study the application of, uh, of AI in education practically in classrooms more by building the technology and seeing how it can help improve learning outcomes.

But I read a lot of the research coming out of the e- out of the labs to help us design better systems and help us think through the guardrails that we need to put in place when we're building the technology, 'cause I'm a, I'm a builder, right? And some of the evidence that I thought was most interesting and it helped us shape our AI tutor is, uh, it's a really interesting study.

It's about, uh, chess.

Uh, it's a 12-week RCT where the researchers, uh, put students who wanted to learn chess in two groups.

They're, they, um, were learning chess, and they would either receive hints on regular intervals, so you're learning and you get a hint from time to time, or they would get those hints in regular intervals and there was a little button in the corner of the screen where they could click on, and any time they clicked on the button, they would get an extra hint.

What roadmaps do you see to, um, well, to, um, balance these two outwardly contradictory aims of, uh, of schools and universities?

So if an RCT lasts for, you know, 10 weeks or six weeks, how do you make sure that in the 20th week, this student will still know what you tried to teach them?

Yes.

This is one of the biggest gaps in the current literature on AI in education.

You had a guest on your podcast a couple of weeks ago, Wes, who was talking about the scale evidence hub who reviewed 800 papers in AI in education and only 20 of those papers had any form of causal evidence and none had an RCT in the U.S., We, Edie, was one of the 20 papers.

So I'll mention that I was very excited to be mentioned in that scale review.

That was our trial with Google DeepMind that I was just talking about.

But yes, it's a huge gap where a lot of these trials on AI and education have very, very short windows.

Yeah, so I feel very strongly

how do we know that this kind of what we're calling a constrained ai tutor is more effective than just going on to chat gpt

yes we love the word vanilla ai it's just like yeah it's good ice cream every time yeah so we there we have strong theoretical hypotheses for why it's important to build a constrained AI tutor that is problem-focused.

And actually, something that I do want to mention is we're working with Google DeepMind on this, and kudos to them for wanting to work with a small edtech organization that has build a decade-old diagnostic engine and that is committed to publishing any RCT results, right? They were interested in learning from this approach where we are building a constrained AI tutor, even though, of course, they're also building their own Gemini platform that is unconstrained.

So I think it's just something that credits to all of us trying to build something that activates when we realize that a student needs the help.

That's what we're building.

It's informed by the latest evidence on AI in education.

And I'll talk about a few studies that have kind of that really show this difference in effectiveness of unconstrained chatbots versus constrained AI tutors like the ones that we built.

yeah like rotate them or rotate them yeah yeah