I'm three attempts into asking a couple of AI-generated podcast hosts a question about letters I am trying to untangle in Old English—thorn and ash—and I'm getting ignored. Not "I don't understand you" ignored. Just nothing. I say "excuse me" for the fourth time like an idiot, feeling insane for negotiating with software, and then it works. The hosts stop and take my question like I interrupted a real conversation instead of a recording. Then one of them asks me a follow-up back.

I sat there for a second like, okay. That's not nothing.

We'll get back to that. First I need to back up, because the ash-and-thorn moment isn't where this story starts. It's not even close to the most important part.

Before I knew what I was doing

For most of college I studied the hard way, because it was the only way I knew. Detailed notes in every class. Reading chapters before we even covered them so I wasn't walking in blind. Flashcards. Actual posters taped to my actual walls. I'd remember information by where it was (which wall, which poster, top-left or bottom-right) because if I could see it in my head, I could get to it. It worked, mostly. It also took forever, and it only worked as long as I could keep rebuilding that mental room every single time.

Then a couple months into spring semester, I found out what NotebookLM could do. And the class where I found out was Intro to Human Origins, the class I failed. That's not a flex, that's the honest timeline. That class landed right at the tail end of a stretch where I basically didn't believe I belonged in college at all. I stopped going for almost two weeks because I couldn't make myself walk in the door. I'm past that now, all the way past it. I deserve this degree at least as much as the kids whose parents are still covering their allowance while they cruise through on autopilot. I'm not saying that to be dramatic. I'm saying it because it's true, and because it's the condition I was in when the thing that changed my whole relationship with AI happened.

The prompt that did it

It wasn't in that failing class. It was during independent research, digging into the prison-industrial complex, when I ran this prompt:

"Analyze this idea from the perspective of someone who once believed it but now feels disillusioned. What made them change their mind, and how would they reinterpret the passages they once admired?"

And it just hit me.

Not because it summarized anything faster. I didn't need faster. I needed different. That prompt let me stand somewhere else entirely and look at the same material through someone else's eyes, someone who used to believe the thing and stopped. Rereading harder wasn't going to get me there. Neither was a fourth flashcard. What got me there was making the AI hold a position and then argue with it, or through it.

That's the whole thesis of this article, so I'll just say it straight: NotebookLM didn't get me anywhere because it summarizes things fast. It got me somewhere because it let me keep reframing the same hard material: a different angle, a different belief, a different level of "explain this to me like I'm five," until something finally clicked. It doesn't replace studying. It gives you more doors into the same room.

(This one has nothing to do with any actual grade. I just run it in basically every notebook I open now, purely for the joy of it:

"What's the conspiracy-theory version of [everyday process]? Not literal conspiracies — the connecting-dots thinking that surfaces incentives and power dynamics normally invisible in it."

Half the time it's just funny. The other half it surfaces some weird connection or power dynamic a normal explanation would've smoothed right over. Don't underestimate the stupid prompts.)

I also lean on this one constantly. It's older than the disillusioned-believer prompt, and it's the one that quietly taught me I could give the AI a role instead of just a question:

"You are a University Professor of [topic]. Evaluate these sources for their reliability and usefulness. Present me with a list of the 20 best resources to use in my presentation. Do not pick sources that are less than 1000 words."

I first ran it as "University Professor of Public Policy Management," about five times in one Public Policy class, before I ever had the aha moment. But swap in whatever you're studying. That's the whole point. In hindsight, it was already doing something a keyword search can't: giving the notebook a discipline and a decision to make, not just a topic to summarize.

What changed after that

I fed it my summer classes next. Got A's in every single one.

I'm not going to sit here and tell you NotebookLM personally handed me those grades. That's not how any of this works, and this brand exists specifically to call out people who oversell tools like that. What I can tell you is that studying with it felt different enough, and worked well enough, that I had time and energy left over that I didn't used to have. And I used that leftover time to start asking: if this basic-ass notebook tool can do this, what else is out there? That's the origin of what I've been calling my AI mabbithole. It didn't start as some separate interest that happened to line up in time — it started because NotebookLM freed up bandwidth I used to spend just trying to survive a chapter.

So that's the arc: a study tool that worked well enough that it turned into the thing that got me obsessed with AI generally. NotebookLM is the reason I have opinions about half a dozen other AI tools now. It's the gateway drug of study apps.

The real test: Old English

Here's the thing, though: none of that proves the method holds up on something genuinely hard. Human Origins and my summer classes weren't nothing, but I want to be honest about what tests a study system: something you're bad at, that's required, with real stakes attached.

I started Russian once and didn't finish it. Language acquisition is legitimately difficult for my brain specifically—not a humble-brag, just a fact I know about myself at this point. And right now I have three classes standing between me and graduation. So instead of picking an easy demo, I pointed the whole experiment at ENG 428, Old English I, using my actual course textbook, and recorded the whole thing so I'd have to be honest about what happened, good or bad.

I want to say up front what "success" meant for this test, because it's not what you'd assume. It was never going to be "Old English becomes easy." The actual bar was: can I keep lowering the level of explanation until something sticks, and then come back later and strengthen it? That's genuinely what learning looks like. It's not instant, and it shouldn't be.

Where it got hard: accurate answers that didn't teach me anything

I started with one textbook, nothing else. Chose PDF over EPUB because I figured it'd be easier to flip through and mark up. No grand statement there, just what worked for my brain this time. NotebookLM spit out a summary and some suggested questions right away, and honestly I side-eyed half of them because they didn't match what I needed.

Then I asked for the complete-beginner version of chapter 4. Told it to give me exactly three things to understand first. It listened. Gave me exactly three. And every single one of them still assumed I already knew grammar vocabulary I did not have. Case. Predicate. Terms stacking on terms. I said out loud, more than once, "this isn't landing," because it wasn't.

Here's the part that matters for anyone reading this hoping AI will save them time: I opened one of the citations it gave me, and it checked out. It pointed straight at the real passage in the real textbook. Accurate as hell. And I still didn't understand the concept. That's the lesson buried in this whole section: a source citing itself correctly and a source teaching you something are two completely different tests, and a tool can ace one and flunk the other in the same breath.

Throwing out the script

At some point I just stopped following my own plan. I had a prompt library sitting in my Notion that I use, built from stuff that's worked for me before, and I went and got it instead of sticking to what I'd scripted for the recording. Built a custom Audio Overview prompt: gave the two hosts opposing jobs, one explaining, one pushing back, because a friendly recap wasn't going to cut it. Tried explaining the same concept four different ways: like I'm ten, like I'm a beginner, like I already practice this stuff, like I'm an expert.

And somewhere in there I figured out the real problem. It wasn't Old English. It was that I didn't have a strong enough grip on modern English grammar to even understand the terms being used to explain Old English grammar. The floor I needed wasn't under the subject I was studying. It was under the language I already speak every day that I had no reason to know I was so out of touch with until now. Small brag here, but I can write. Knowing how to form sentences in engaging and creative ways reflects an intuitive grasp of grammar and syntax, but it definitely doesn't mean I can remember the weird English syntax jargon.

That made the realization a bit of a surprise, and it's one I don't think I would have reached without the freedom to keep reframing and re-asking instead of following a script that wasn't organically written.

When I asked it to check its own assumptions

Once I hit that wall, I tried something more aggressive: a prompt that hunts down hidden assumptions the source material is making without stating them outright. Then I immediately second-guessed whether that was even the right tool for a grammar-mechanics problem, which, in hindsight, is a fair thing to have wondered.

I want to be straight with you about what that audit actually got right and where I'm not going to vouch for it. It correctly noticed that my textbook (Baker's Introduction to Old English) is written for literary and historical readers, not linguists. It does grammar review, pushes you toward reading real texts early, and flat out says the glossary in the back isn't a substitute for actually learning the inflections. It also correctly noticed that the book commits to West Saxon dialect because that's what most of the surviving literature is actually written in, while still admitting other dialects existed. And it correctly noticed that the book is upfront about pronunciation being a reconstruction nobody can verify firsthand, while still recommending you read out loud anyway.

Where I'm drawing a hard line: it also generated some much bigger accusations, that this teaching approach would leave students with a permanently low ceiling, or unable to handle major texts, or picking up some kind of fake accent. Those are speculative. They're not something one textbook or one recording session proves. So I'm treating that whole audit for what it actually was: a tool that helped me notice assumptions and ask better questions, not a verified indictment of how my professor's book teaches the language.

Somewhere in the middle of all this, I did start forming rough working theories: dative case felt connected to what I was calling "transitive-type" nouns in my head, accusative felt possessive-adjacent, verb information seemed to hide toward the end of a sentence. I want to leave those exactly as sloppy as they actually were, because that sloppiness is the evidence. I was retaining distinctions and building hypotheses before I could recite a single rule correctly. That's real learning happening in real time, messy, not carved in yet, but not nothing either. (For the record, once I checked those hunches against my actual textbook: they're not quite right as stated. Baker's real definitions are more specific than my shortcuts. I'm leaving the shortcuts in because that's what my brain actually produced, not because I'm telling you they're correct grammar.)

The best moment of the whole test

Back to the beginning of this article. That ash-and-thorn exchange was the clearest single proof of what this tool can do that a static explanation can't.

I'd just sat through a full generated Deep Dive, a real wait, nothing instant about it, and switched it into interactive mode so I could jump in instead of just listening. Took a few tries before it recognized I was actually trying to ask something instead of just making noise, which was genuinely annoying in the moment. But then it landed: a real answer about thorn and ash, and it let me ask a follow-up right after, like an actual conversation instead of a lecture.

What caught me off guard afterward was that it gave me historical context I'd been told earlier in my own process that I didn't strictly need, and it turned out to be useful anyway. Context I thought was optional ended up doing real pedagogical work. I'll take that trade every time.

One honest wrinkle: when I stopped asking questions, the session acted like it was over and I had to restart it. I don't know if that's how it's supposed to work, some weird state it got stuck in, or just me misreading the interface. I'm not going to pretend I know which. What I do know is I restarted it into a completely different framing, a medieval-scriptorium scene, and that creative reframe did a lot to make an intimidating dead language feel less like a wall. It didn't teach me case endings by itself. Atmosphere and comprehension are two different jobs, and I'm not going to pretend one did the other's work.

Worth saying too: interactive mode wasn't something I had access to in my own NotebookLM workflow last semester. This was genuinely new territory for me, not an old habit I was just revisiting on camera.

The Gemini detour

I also poked around in Gemini's own Study Notebook setup, just to compare. It had something that looked like a study plan with chapter priorities already laid out, and I actually stopped and asked: wait, did this get built before or after I added the textbook? I genuinely couldn't tell from what was in front of me, and I'm not going to pretend I could. An organized-looking plan isn't automatically a trustworthy one. You still have to know where it came from.

One thing did click there: framing Old English word order as "scan toward the end of the sentence for the action" made something snap into place for a second. It's not an actual grammar rule. My textbook is pretty clear that Old English word order isn't as loose as people assume; there are real patterns, just not the one I invented. But as a beginner's crutch, it worked in the moment. Then the examples got confusing again and I landed on the same conclusion I kept landing on all day: I should've started even more basic than I did.

So what actually happened here

Almost an hour of footage. Two tools. Half a dozen reframes. And I did not walk out of that session fluent in Old English case endings.

Here's where I need to be really clear, because it would be so easy to read all of that as the tool failing me. It didn't. NotebookLM did its job well. Humans don't learn at AI speed, and that was never going to change no matter how good the software is. What NotebookLM actually gave me were routes into the material that a static textbook and a stack of flashcards never could: different angles, different levels, a version of the material that pushed back on me, a version that turned into a conversation instead of a wall of text. I had to steer all of that myself. I had to keep asking for something more basic, keep switching formats, keep checking sources, keep backing up to the thing underneath the thing I thought I needed. That's not the tool underperforming. That's me using it well.

Old English wasn't a bad choice for this test because it was hard. It was a good choice specifically because it was hard, because it forced the tool to keep adapting instead of just summarizing a chapter I already halfway understood.

Pick the problem you actually have

The original plan called this a study loop. Accurate, but boring. Nobody who is already drowning in a chapter needs another diagram ordering them through ten mandatory steps like a productivity rosary.

So don't run every prompt. Pick the problem that is happening right now.

“I opened the chapter and understand absolutely nothing.”

Find the floor before you decorate the room:

"What are the three things I absolutely need to understand from this document before I can follow the rest? Explain each one clearly, with the page reference where it appears."

If the floor is still missing, force the material to leave its home territory:

"Explain the central topic of this notebook by borrowing structure from a completely different field. Name why the structures are similar, map the important pieces one-to-one, walk through one example, and tell me exactly where the analogy breaks. Cite the source for every mapping."

“The explanation sounds polished, but I think it is skipping something.”

Put your own notes beside the source and make the gap visible:

"Compare my lecture notes with the reference material. Sort the result into WRONG, MISSING, UNSUPPORTED, and FINE. Quote both sides, rank missing material by likely exam importance, and cite the source."

Keep the Hidden Assumptions Audit in your back pocket for the stubborn cases. It is a question generator, not a machine-certified indictment:

"Audit these sources for load-bearing assumptions the author does not fully defend. Mark each Stated, Implied, or Smuggled; explain what would have to be true for it to hold; flag conflicts between sources; and rank the five assumptions that would cause the most damage if wrong."

“The notebook is being way too agreeable.”

Turn the Audio Overview into a disagreement instead of a recap:

"Host A is the Expert who explains the material. Host B is the Challenger who questions the method, points out gaps, and plays devil's advocate. Do not agree just to be agreeable."

Once you know enough to judge the answer, ask for a direct critique of the source's methodology and argument. Before that point, a confident critique is just another impressive-looking object you are not equipped to evaluate yet.

“I think I know this, which is exactly when I should become suspicious.”

Make yourself answer before the machine does:

"Quiz me on these sources one question at a time, starting easy and getting harder. Do not give me the answer until I respond. Correct what I missed and cite the source. Keep going until I say stop."

Then flip the room:

"You are the student and I am the teacher. Let me explain the topic without interrupting. Then ask one basic question, one question testing why rather than what, and one hard edge case. Use only my sources. Score my understanding and tell me what to study next."

If you want more pressure, turn it into an oral exam. If you need application, generate a mix of definition, cause-and-effect, comparison, and scenario cards. The point is not collecting prompts. The point is making yourself perform a different kind of thinking.

“I am about to repeat this claim to another human being.”

Stop. Open the citation. Read the surrounding passage. Check whether the source actually says what the answer implied. End every route back at the evidence, not at the AI's confidence.

One rule will save you a headache: a prompt only stays in the rotation if it makes you do something different. If two prompts lead to the same mental move, cut one.

What this actually costs

You do not need my subscription to try the core method. Gemini Notebook has Standard access without a paid plan, but it now uses compute-based limits that can vary with prompt complexity, chat length, and features. Paid plans raise those limits; they do not own the basic idea of asking better questions and checking the source. Google's current limits page says the limits can change, so check the live Usage screen rather than trusting an old screenshot.

If you're an eligible U.S. college student age 18 or older, Google is currently offering one year of Google AI Pro through its Gemini for Students program, redeemable through December 31, 2026. It requires eligibility verification and a valid payment method, and it automatically becomes a paid plan unless cancelled. Set a reminder. Seriously.

One practical thing I learned the hard way: don't let your main study notebook go hunting for its own sources. Ask it to find things on its own and it'll dump forty or fifty random results in and turn your clean study space into a landfill. Keep a separate junk-drawer notebook for open-ended digging, and only move something into your real notebook once you've decided it's worth keeping.

What I don't trust yet

I don't love that open-ended AI tools can pull me away from the material I meant to study. What I value here is that Gemini Notebook answers from the notebook sources—the full set or the subset I select—so I can open the citation and inspect the passage myself. That still does not make the interpretation automatically correct. I don't love the synthetic voices in the audio overviews—that's just taste, sue me. And I still don't fully trust AI's read on statistics. That one's not going anywhere anytime soon, and it shouldn't.

That skepticism isn't in tension with how much I use this stuff. It's the other half of the same habit. Use it to make yourself think harder, not to stop thinking. Then go verify the parts that actually matter before you repeat them to anyone.

The actual point

NotebookLM didn't hand me anything. It just kept opening doors until I found one I could walk through, one letter, one grammar rule, one wrong theory I had to go correct myself. Turns out that was enough to change my whole relationship with studying, and with AI, period.

Sources checked for this draft

Product-access details checked September 14, 2026. Recheck them before publication.