# AI @ School - I Built the Tutor. He Decided How to Use It.

By Wes Pomeroy. Published October 9, 2026. https://wespomeroy.com/writing/ai-tutor-six-week-results/

At 12:17 in the morning, the night before his first Bio exam, my son had a forty-minute podcast about cells and enzymes that hadn't existed when he sat down that evening. Two hosts, built from his own class notes and his professor's list of exam objectives. One explains, one asks the questions a student actually asks.

Six weeks ago I [set my sons up with an AI study assistant for every class](https://wespomeroy.com/writing/ai-tutor-for-students/), then [wrote up how to build one](https://wespomeroy.com/writing/how-to-set-up-an-ai-tutor/), then [reported the first week](https://wespomeroy.com/writing/ai-tutor-first-week-results/). I said I'd rather find out what happens than study it with the technology switched off.

Here's the thing six weeks taught me. You use this technology to the degree your work calls for it, and to suit yourself. That looks obvious written down. It wasn't obvious in August, when I handed my kid the setup I use and assumed he'd use it the way I do.

## He used it a seventh as much as I did

Each class has its own folder the assistant remembers between visits. Across six weeks he opened one of his 22 times. Over the same stretch I opened one of mine 150 times.

That gap isn't him falling short. My work lends itself to this technology, it suits my skills, and I like it, so I use it on almost everything. He's a college student who enjoys some of his classes for the experience of sitting in them, with a life away from the computer far more active than mine.

His 22 sessions bunch into the two or three days before something got graded and go quiet in between. Same tools, two people whose work has almost nothing in common, and the usage looks different. It should.

## Three things it did that I didn't write about

**1. It wrote down what his professor said and the syllabus doesn't.** He records his political theory lectures, and the system turns each one into written notes. A section of those notes is headed "What he said that the syllabus does not." Skip the short assignments and your grade can't go above a 60, so a student with a good paper and good exams can still fail. Save your two dropped grades for the end of the term. The short assignments aren't summaries, because the professor adds material in class that isn't in the readings. A student who misses that, or hears it in week one and forgets it by October, pays for it.

**2. It made whatever he needed before a test.** A podcast for the Bio exam. Vocabulary drills for Spanish, pulled off the textbook publisher's site. For a Foundation of Democracy quiz where handwritten notes were allowed, it checked the course site for what the quiz covered and then wrote notes designed to be copied out by hand. The format came from the quiz rules, not from a preference.

**3. It caught mistakes in three directions.** His own, in a draft: a line credited to Confucius that belongs to one of his students, and the definition of a term he had backwards. Its own, by checking its notes against the source texts and finding two passages it had labelled wrong. And his professor's, describing Confucius and Plato as contemporaries, which they aren't. It told him to ask about that rather than repeat it on an exam.

## The night it worked out how he gets things wrong

Intro biology. First exam of the term, four chapters, all multiple choice. He sat down at eight in the evening the night before and pasted in his own typed notes. The system pulled his professor's list of exam objectives off the course site, which is the actual specification for what the test would ask. Out of that it built the podcast, so he had something to listen to instead of rereading the same pages.

Then it did something a study guide can't. It started quizzing him, chapter by chapter, and kept the score. Chapter three came back 19 out of 20. Chapter two, cold, came back 14 out of 19, so it re-quizzed the misses four times and got him to 9 out of 10. One section of chapter four went 7 out of 12.

Then it looked across all those rounds and found something none of the individual scores showed. He knows each term fine on its own, but put two of them side by side and he swaps them. Hypo and hyper. Cohesion and adhesion. It happened across seven pairs of look-alike opposites, and his professor's objectives list says "compare and contrast" more than a dozen times. So it stopped drilling both definitions at him and taught him to derive one from the other: organic means it came from something living, so the organic ones are vitamins, which leaves minerals as the inorganic ones. The section that had gone 7 out of 12 came back 8 out of 8.

It also split two problems that usually get the same advice. He got one topic right twice and then wrong again a day later. Its read was that he understood it fine and wasn't holding onto it, so more explaining wouldn't help him and repetition would.

It got the teaching wrong once, too. It noticed a slide explained a rule backwards and tried to get ahead of the problem by warning him first. That backfired. He didn't know the rule yet, so a warning about a confusing version of it was all he had, and he scored 2 out of 10. The note it left itself: teach the plain rule first, mention the confusing slide later. That note sits in the folder, so the next session starts from it.

## The lectures and grades that never went in

Most of what I built never got used, and that's the other half of the story.

Three lectures got written up, all three on one afternoon in early September. Fifteen classes met over the six weeks, so twelve of them never made it in. His own project notes say it plainly: no lectures have been pulled off the recorder since September 4. One of his six classes never got a folder at all.

And no graded work ever went in. Not one marked assignment, so my most confident claim in August, that this would show the pattern in how a kid's understanding builds across a term, only ever happened in that one Bio session, off practice quizzes the system wrote itself.

I don't know why, and I haven't pushed. Some of it is a kid with six classes who went to the system when he needed it and not otherwise. Some of it is friction I built in: recording every lecture and feeding it in every day is work, and the system never made that work small enough to be automatic. One network outage killed a session right after he typed "save progress." That session's notes only existed in the chat window, so they went down with it, and three weeks later his notes on two readings are still gone. That is the kind of thing that teaches a person not to bother.

## The value he's getting

No sign his grades improved. In his situation that would be hard to see even if it were happening, because he was already scoring in the nineties.

Spanish has posted grades. Nothing else has yet. Four graded items: 92, 95, 93, 93. The first came before he'd used the system for Spanish at all. He prepped for the next two. For the fourth, the study session never happened. There's no correlation in that and I'm not going to manufacture one. It's hard to compare good with good.

So the evidence is softer, and I'd rather report it than a number I don't have. He keeps going back to it and I haven't asked him to. And he tells me, repeatedly, that it makes him feel more prepared, more confident and less stressed about school. He's also a strong student with the time and resources to try, so a kid without those things would be running a different experiment.

Outcomes are what I care about. If he's happy and doing well, I don't need the system to have caused it.

## Lessons learned

This isn't a carpenter getting power tools. It's a good set of tires. They don't make a good driver faster on a dry road and the lap time looks about the same. You find out what they were worth in the rain.

He didn't change how he works. He sharpened the way he already worked, and he was selective about where: the weak spots got the system, and everything he already had a handle on he just did.

I don't think there's necessarily a one-size-fits-all answer here. Something about this technology is bespoke in a way most of what came before it wasn't - it has to be fitted to the person and the work, every time - and that's what makes it hard to get your head around and hard to get anything out of. So you start with a specific opportunity, try things, and watch what you put in against what you get back.

The goal was never to get him using more AI. It was to get value out of it. Those things are not the same. One delivers calibrated value. The other just checks the box.