Sign up for Chalkbeat’s free weekly newsletter to keep up with how education is changing across the U.S.

A group of Stanford University researchers started with one question: Could a human tutor providing motivation and support get students to spend more time working with an AI literacy tutor?

The answer turned out be yes — but only between one and four minutes more per week. Many students never logged on at all.

That left the researchers with a different set of questions.

“A key finding that we weren’t even meaning to test is that having access to this AI tutor isn’t the same as using it,” said Carly Robinson, the lead author on the study released Wednesday and the director of research for the SCALE Initiative at the Stanford Accelerator for Learning.

Robinson said this doesn’t necessarily mean AI tutoring doesn’t work. “We never really got close enough to the dosage needed to find out.”

The study adds to a very limited research base for AI tools in education as many school districts are looking for ways to maintain or increase tutoring that costs less than hiring people to do the work. It also aligns with the experience of Khanmigo founder Sal Khan, who recently acknowledged that students didn’t engage much with the AI tutor he initially hoped would revolutionize education.

Robinson and her colleagues worked with two school districts serving high-poverty populations using the same AI learning platform in different tutoring settings. Neither the districts nor the learning platform are named in the study.

In one school district, elementary children were supposed to use the literacy tutor during homework time in an afterschool program. In the other, early elementary students were supposed to use the AI tutor during class time.

The learning platform provider said that students would show improvement in their reading skills with at least 30 minutes of use a week. The goal was for students to complete two 30-minute sessions each week.

Students were randomly assigned to either work independently with the AI tutor or with guidance from a specially trained person — afterschool program staff in one district and middle school students who were strong readers in the other — who could provide motivation and tech support.

Other research has found that the relationship between tutors and students can be a key factor in tutoring effectiveness. Researchers wanted to know if having human support would increase engagement with the tool and raise test scores over the course of the school year. They also thought they might learn that students working independently made more progress because they spent more time on the platform and less time chatting.

But it turned out both groups of students barely used the platform, and having a human tutor didn’t increase the likelihood that students would log on.

Students who did use the platform spent about 13 minutes in a session in the afterschool setting and almost 26 minutes in a session when they used it in class, enough time to probably derive some benefit, Robinson said.

But the low rate of participation meant students in the afterschool setting averaged only about two minutes a week on the platform when they worked independently and three minutes when working with a tutor. In the other district, younger students working independently in class averaged a little more than five minutes a week on the platform, which increased to almost 10 minutes a week when they worked with a tutor.

There was no meaningful difference in the reading scores of the two groups.

The bigger finding, Robinson said, was that many students didn’t log on most weeks. And those that did log on were more likely to already be higher performing and not identified for special education.

Researchers don’t know why students didn’t use the platform. Perhaps some didn’t find it engaging, or perhaps teachers directed students to spend their time in other ways. Those other activities may have been beneficial, Robinson said. The study didn’t compare outcomes for students who used the platform and who didn’t use the platform.

Robinson said she still sees enormous potential for digital tools to provide personalized instruction at scale. But before districts pay for a license, administrators need to look at more than whether there’s evidence a particular AI tool improves student learning.

“The challenge isn’t just building good AI tools,” Robinson said. “It’s really getting students to use them, and that seems to take the same type of intentional design that we’ve learned matters with other ed tech interventions and tutoring.”

Erica Meltzer is Chalkbeat’s national editor based in Colorado. Contact Erica at emeltzer@chalkbeat.org.

There is a famous puzzle where a boy takes a 5 foot fishing rod on a bus that doesn’t allow objects taller than 4 feet on by packing it in a 3 by 3 by 3 box (which has a long diagonal over 5 feet). The idea being: such a box is no more than 3 feet tall in when laid flat on any side. Common english doesn’t routinely use precise definitions of diameter or caliper width.

We can generalize to the following box in box question:

Are there cases where you can put an d by e by f box inside an a by b by c box when d + e + f > a + b + c (allowing a rotation)?

Currently a few of the common AIs answer as follows:

• GPT-5 mini:
  

  Yes — it can happen. For axis-aligned rectangular boxes (orthogonal placement) you can fit a d×e×f box inside an a×b×c box even if d+e+f > a+b+c, because the sum of side lengths is not the relevant constraint for containment; what matters is that the three dimensions of the inner box can be matched (possibly after permutation/rotation) to the three dimensions of the outer box so that each inner dimension ≤ the corresponding outer dimension.

  [tries and fails to work some examples, but does not change its conclusion]

• Claude Haiku 4.5:
  

  This is a fascinating geometry problem! Let me think through it carefully.

  ## The Short Answer

  **Yes, it is possible** to fit a d×e×f box inside an a×b×c box even when d + e + f > a + b + c, if you allow rotation.

  [tries and fails to work some examples, but does not change its conclusion]

For somebody familiar with mathematical reasoning, the failures to generate examples are a strong clue to double check the result. The LLM do a lot of heavy lifting in an extended dialogue. For somebody new to this (say a student looking for a puzzle and geometry coach), the confidence may lead them to stick with the answer.

In fact I am not even sure if I am remain convincing when I say the answer is in fact “no.” (I wrote about this in 2013, and the Cauchy’s surface area formula would be available to quickly prove this since 1850.)

All right, here we go. I’m so bad at these icebreakers, and you all had such interesting facts.

First one: I’m a Beyoncé superfan, and I’ve seen her in concert seven times. One time, my friends and I waited behind the venue for like two hours after the show, and she signed my T-shirt before getting on her tour bus.

Second: I went to middle school with a guy named Dennis Schröder. He plays for the Cleveland Cavaliers now. I’m not a big NBA fan, so I don’t know how much he plays, but I shared a few classes with an NBA champion.

And the third one: When I was twenty-two, I visited my sister in Louisiana. We were heading to some party and must’ve gotten turned around because suddenly we were out in the middle of nowhere. The sun was going down, there were no street lights, and, of course, her car broke down. No warning. The car just turned itself off. As we got the car to the side of the road, thunder clapped, and it started to pour. That’s when we met the Bog Rat.

Yes, Gary, Bog Rat. B-O-G R-A-T, just like it sounds. They call it the Bog Rat, but it was more like a five-foot-tall capybara that could stand up on its hind legs. Sort of like Splinter from the Teenage Mutant Ninja Turtles if he were fatter.

At first, we were freaked out, but he was actually very friendly. He reared up, looked us in the eye, and said, in a slight English accent, “Car trouble, is it?” We just sort of nodded, and he goes, “That’s too bad. Let’s get you out of the rain, Tyler and Samantha.” Mind you, we’d not said anything to him yet, much less told him our names.

Gary, can you wait until I finish the thing about the Bog Rat before you guess which one is the lie?

So we followed the Bog Rat down this path. He held an umbrella over us for the half-mile walk, sort of like if the butler in Gosford Park had been a large, nude, anthropomorphic rat. The path ended at this enormous Gothic mansion. He led us inside, and there were floor-to-ceiling bookshelves crammed full of books, but all the books were written in Bog Rat.

Yes, Gary, Bog Rat was the name of the beast and also the name of his language. Just like Smurfs, where they are Smurfs but they also speak Smurf.

Anyway, he had several guest rooms, and he let Samantha and me stay there overnight. When we woke up, he’d cooked us this wonderful Bog Rat soufflé. We walked back up to the path, prepared to try to fix my sister’s car. But in the spot where we left my sister’s beat-up Honda Civic, there was now a jet-black Rolls-Royce. The Bog Rat was like, “Here’s your car.” And we’re both like, “This isn’t our car.” But he opened the glove compartment and pulled out the car’s title and, sure enough, there was my sister’s name. “A gift,” the Bog Rat said and then disappeared back down the path.

I’m sorry, Gary. I didn’t realize that I’d opened the floor for stupid questions. No, the Bog Rat did not make us a soufflé made out of Bog Rat meat. It was a traditional Bog Rat soufflé containing carrots and Stilton cheese.

Nevertheless, we got into the Rolls-Royce and drove off.

The next day, we came back to the exact same spot with some flowers and candy for the Bog Rat as a thank-you. We found the path and followed it for over two miles. No mansion. It was gone. Vanished.

After searching for two hours, we went up the street to a gas station to get some refreshments. As we walked in, the wizened old lady working there looked out the window, saw the Rolls-Royce, and said, “Don chu assochiate wit no Bog Rat. He’s no good.” We’re like, “No, we met him. He’s nice.” And she shakes her head and goes, “Bog Rat ain’t no good.” Then she nodded toward the window. Outside, the Rolls-Royce was suddenly surrounded by cop cars. “This Rolls Royce, with this license plate, was seen leaving a crime scene three weeks ago,” the cops told us. Samantha and I were both like, “No, we just got the car yesterday from the Bog Rat,” which of course sounded insane. Anyway, they opened the trunk and found a bag containing the severed remains of a lady who had gone missing several weeks prior. Since the car was in my sister’s name, the cops hauled her away. She’s currently serving life in prison.

Anyway, I guess that’s it. Put your guesses in.

Okay, the lie was number two, the Dennis Schröder one. He is from Germany, and I’m from Kentucky. People forget he’s German, because his name is Dennis.

I learned something beautiful today and just wanted to share it. Obviously, this blog is super eclectic, with tons of posts on science and math and art and life, with no particular order. Maybe someday I’ll organize it.

This a pretty basic calculus trick. In fact, I wouldn’t be surprised if I’d come across it in high school and just forgot it. But today it came up in my research, and I had to rederive it, and I thought it was very pretty.

So say you have a scalar function $s(t)$, which we’ll assume to be differentiable and generally cooperative. You can go ahead and find its derivatives $\frac{d s}{d t}, \frac{d^2 s}{d t^2}, \frac{d^3 s}{d t^3}$, and so on. But suppose that (like I did) for some problem at hand, you need the derivatives in the other direction: $\frac{d t}{d s}, \frac{d^2 t}{d s^2}, \frac{d^3 t}{d s^3}$ and so on.

(It’s worth noting that the inverse function $t(s)$ doesn’t actually have to exist globally for this to work — that is, $s(t)$ need not be one-to-one. It just needs to be one-to-one locally, so you can define $t(s)$ in a local region, and take derivatives, after which you can forget about $t(s)$. This amounts to requiring that $\frac{d s}{d t} \neq 0$ at the point of study, and indeed our final formulae will diverge if $\frac{d s}{d t} = 0$, indicating this requirement.)

The first derivative. Delightfully, it holds that

as you can easily confirm by drawing a scalar function and transposing the two axes. I find this result really great because you can just treat $\frac{ds}{dt}$ as a conventional fraction and take its inverse. Shows it’s great notation.

The second derivative. This one we need to be more careful about. Our main tool is the chain rule: $\frac{d}{ds} = \frac{dt}{ds} \cdot \frac{d}{dt}$. Applying this, we find that $\frac{d^2 t}{ds^2} = \frac{d}{ds} \frac{d t}{d s} = \frac{d t}{d s} \cdot \frac{d}{dt} \frac{d t}{d s} = \left( \frac{d s}{d t} \right)^{-1} \cdot \frac{d}{d t} \left( \frac{d s}{d t} \right)^{-1}$, and evaluating the final derivative we find that

I find this really nice — the second derivatives are proportional, with the rather surprising constant of proportionality of $- \left( \frac{d s}{d t} \right)^{-3}$. Even though it didn’t solve the particular research problem I hoped it would, seems like it might eventually be useful for something. (Wonder if there’s some intuition for that other than a dimensionality argument…)

The third derivative. Again applying the chain rule, taking $\frac{d^3 t}{ds^3} = \frac{dt}{ds} \cdot \frac{d}{dt} \frac{d^2 s}{dt^2}$, we find that

Not as pretty, and you lose the nice constant of proportionality between $s^{(k)}(t)$ and $t^{(k)}(s)$. Thanks to the chain rule, things only get more complicated from here on out.

I thought about this the other day, and I thought it’d be fun to share this internal tool I made over a decade ago to aid with exploring options for Medium’s typographical redesign.

It’s called Fontificator. You can play with Fontificator here (desktop browsers only), or watch the likely confusing video below:

The motivation for building Fontificator came from two observations:

• font previews on type foundry sites were generally too limited to get a real sense of how a certain typeface feels, and it was best to see a font in situ,
• often an extremely tiny nuance – like adding some letter spacing, or messing with line height – was what separated something that was promising from something that seemed very far from working.

With Fontificator, I was aiming at this Doug Engelbart-esque notion of one hand on the keyboard + one hand on the mouse, and the UI where it was only necessary to point to an element, and the keys under your other hand would start working immediately – no clicking needed:

• F and G to change the font,
• – and + for font size,
• ← and → for letter spacing,
• ↑ and ↓ for line height,
• < and > for opacity (for all the above you can hold Shift for bigger moves),
• and, there are a few more shortcuts you can see at the top.

This way, we could move really, really fast. To accommodate that, Fontificator always tried to keep the current item under the cursor by counter-adjusting scroll position as needed.

On top of it all, a few more shortcuts:

• ⇥ and ⇧⇥ move very quickly between different types of stories so you can preview that,
• Space compares to the original/​current version,
• 1–9 allow you to switch to different “slots” so you can have various presets ready to compare,
• Esc hides the toolbar for maximum immersion,
• ⇧R resets.

You can also edit any text if you are so inclined, and also drag in any font file from your computer onto a paragraph – then that font becomes part of the F/G stack. (Bernino Sans and Freight Text were the starting fonts before the redesign.) On the left, you can also see a naïve mobile preview – there was also more sophisticated on-smartphone preview, but I removed it from this restored version.

Fontificator was literally made for an audience of 2–3 designers (and perhaps 1–2 stakeholders in read-only mode), and it was surprising to me how quickly one could master this strange tool, have fun with it, and feel the entire typography on the page becoming much more malleable. We also put up a more “traditional” list of contenders on the wall…

…but it was in Fontificator where we learned the most.

I love internal UIs because they allow you to go very wild and very tactical. If you have one you’d be willing to share (maybe it, too, is on the other side of the statute of limitations?), or one you already wrote about or spotted someone else doing so, please let me know!

#internal ui #typography