What's Actually Changing in AI Education (And What's Just Marketing)
What's Actually Changing in AI Education (And What's Just Marketing)
I work on a study product. I read trend pieces about AI in education almost every day. Most of them are useless.
They're useless because they treat "AI in education" as one thing — a wave that's coming, that will reshape learning, that students must be ready for. The actual story is messier and more interesting. Some of the changes that get the most coverage are mostly marketing. Some of the changes nobody is writing about are quietly reshaping how students study, what professors test, and where the line is between cheating and learning.
Here's what I think is actually happening. Some of it is bullish for the category I work in. Some of it isn't. I'll try to tell you which is which.
The death of "AI chatbot for homework" as a category
Two years ago, every student app pitched itself as a chatbot. Paste in a question, get an answer. The pitch was that ChatGPT was too generic and education needed a chatbot tuned to homework.
That category is dying, and it should.
Two reasons. First, the generic chatbots got really good. ChatGPT, Claude, and Gemini will solve almost any undergraduate problem you throw at them, free or close to it. A wrapper that just routes to one of these models doesn't add anything a student can't get directly. Second, students figured out — quickly, in many cases more quickly than their professors — that getting an answer isn't the same as learning. A 4.0 freshman year and a panic attack at the first closed-book midterm is not a rare story anymore.
The product shape that's replacing chatbots is integrated study systems. Course-aware. Calendar-aware. Memory across sessions. The chatbot is one feature inside a larger loop, not the whole product. Fennie is one of these. So are a few others, with different shapes. The fact that we have competitors who agree with us about the shape is, honestly, a more useful signal than any of our own marketing.
What "memory" actually means, and why it matters
If you read enough launch posts you'd think every AI product has memory. Most don't. They have a context window — the recent conversation — and they call it memory.
Real memory in a study system is something else. It's the system tracking, across days and weeks: which topics you keep getting wrong, which ones you've stopped missing, what time of day you actually study versus what time you say you'll study, which kinds of problems you skip. That information shapes tomorrow's plan. It shapes the way the chat tutor talks to you (more scaffolding on weak areas, fewer hints on strong ones). It shapes which flashcards spaced repetition surfaces.
This is the genuinely new thing. Not the AI itself — those models are increasingly commoditized. The new thing is software that knows you over time and uses that knowledge to make better decisions about what you should do next.
In Fennie this is the eighth object — Memory — that nobody pitches as a feature because it has no UI of its own. It just makes the rest of the system smarter the longer you use it.
The boring corporate version of this trend is "personalized learning." That phrase has been around for two decades and meant nothing the entire time. The non-boring version is: software that actually adapts because it actually remembers.
Calendar awareness is the thing nobody is writing about
Here's a small change that I think has bigger consequences than most of the "future of AI in education" headlines.
A study system that knows you have a midterm Friday studies you differently Monday through Thursday than a system that doesn't.
That sentence sounds obvious. It is obvious. And almost no AI study product does it. Most are session-based — you open a tab, ask a question, leave. The session has no awareness of what's coming this week, this month, this term.
Once a system has the calendar, the daily plan stops being generic. It pulls more orgo into the days before the orgo midterm. It deprioritizes the diff eq pset that's already submitted. It notices you have an essay due Tuesday and Tuesday's plan changes shape three days out.
This is mundane software. It's not flashy. It's not the demo that goes viral on Twitter. But ask any pre-med who survived an orgo final whether spaced, calendar-aware practice is more or less useful than a chatbot that explains a single mechanism, and you know which one wins.
The syllabus pipeline is doing more work than people think
Adjacent to calendar: the syllabus pipeline. PDF in, topic map out.
The reason this matters isn't that it saves you the effort of typing. It's that it grounds the rest of the system in your actual class. Your professor's chapter ordering. Your professor's exam emphasis. Your textbook, not the average textbook.
The next time you read a piece about "AI personalizes learning" — ask whether it's personalized to you, or personalized to a course you're actually taking. Almost always it's the first, and almost always the second is what students actually need.
Oral exams are coming back
This one is worth pausing on.
A few semesters ago, professors started panicking about AI cheating on essays and problem sets. The first wave of response was AI detection software. That didn't work — and I'll get to why — so the second wave is happening now: a return to in-class, oral, and proctored assessment.
I have friends who teach at law schools and engineering programs who've added oral components to courses that haven't had one in fifty years. A 1L in a contracts course who got an A on every paper but can't articulate the basics of consideration is now a problem the professor can detect in five minutes of conversation. So the conversation is happening.
This has a quiet effect on study products. If the test of whether you know something is whether you can explain it out loud to a person, then a study system that hands you answers is actively harming you. A study system that makes you explain things back — Socratic chat, "tell me how you'd approach this," generated quizzes that ask "why" rather than "what" — is doing the work the oral exam will eventually demand.
This is part of why Fennie's chat doesn't hand out answers. Not a moral stance. A practical one. The students whose chat tutors taught them to think out loud are going to do fine when their professor asks them to think out loud. The students who got used to copy-paste are going to be in trouble.
What's overhyped: AI detectors
AI writing detectors have a fundamental problem, which is that they don't work. The false positive rate is high enough to flag innocent students. The false negative rate is high enough that any half-effort prompt rewrite slips through. There's published research on this. There's anecdotal evidence in every academic-integrity hearing room in the country.
Schools are slowly figuring this out. Some have abandoned the tools quietly. Others still use them as a "signal" rather than evidence, which is more honest but also pretty close to admitting they don't work.
The deeper issue is that detectors are trying to enforce a rule — "your writing must be your own" — by inspecting the artifact. Once the artifact is indistinguishable, the only way to enforce the rule is to inspect the process or the person. Which is why oral exams. Which is why in-class essays. Which is why notebook entries that have to show your reasoning.
If you're a student worried about getting flagged, the honest answer is that the way to be safe is to actually do the work. The way to do the work efficiently is to use AI tools that make you think rather than ones that hand you the answer. There's no shortcut around this anymore.
What's overhyped: VR classrooms
People have been promising VR education for fifteen years. It is not happening. The headsets are expensive, the use cases are narrow, and the actual learning gains over a textbook for most subjects are unclear or negative. Maybe in anatomy. Maybe in some lab simulations. Mostly no.
This will continue to be promised. Mostly ignore it.
What's overhyped: agentic study buddies
Every six months someone launches an "AI agent that does your homework for you." This is a product that solves the wrong problem. The students who want this don't need a better agent — they need to either change majors or change their relationship to the work. The students who don't want this aren't going to use it.
I think we'll see this category get small and stay small.
What's actually changing: the line between cheating and studying
Five years ago the line was simple. Don't copy from a friend. Don't buy a paper. Cite your sources. Most students could mostly tell what was over the line.
The line is much messier now. Is it cheating to ask Claude to explain a concept before you do the problem? No. Is it cheating to ask Claude to do the problem and then pretend you did? Yes. What about the middle — Claude does the first part of the problem to show you the approach, you do the rest? Most professors say it depends, which is honest but unhelpful.
Where the line is actually settling, I think:
- Using AI to learn = fine, often actively encouraged
- Using AI to skip learning = cheating, regardless of whether anyone catches you
- The disclosure norm = ask your professor, write it in the assignment if you used AI substantially, don't pretend
The interesting thing is that the systems best positioned for this new line aren't the ones with the strongest "AI detection" or the most permissive chatbots. They're the ones that make the cheating shortcut harder to take. A study system whose chat refuses to hand you answers — whose quizzes are generated from your own notes, whose flashcards come out of your own misses — is, structurally, in the studying zone, not the cheating zone. That's not because anyone made it virtuous. It's because the product is built to teach, and teaching and cheating are different shapes.
What this means if you're a student right now
A few practical implications.
If you're shopping for a study product, ask whether it has memory across sessions and whether it knows your calendar. If neither answer is yes, it's a wrapper around a chatbot you can already access for free.
If you're worried about the cheating line, the safest place to be is using tools that make you do the work. Not because they'll exonerate you. Because you'll actually know the material when the oral component shows up.
If you're early in college, get used to studying as if your finals will be in-person and proctored, because they increasingly will be. The skills that fall out of "AI does my homework" workflows do not survive a three-hour closed-book exam.
If you're later in college and this all sounds exhausting, here's the honest cost-benefit: you can keep typing into chat boxes for the rest of the term and probably get fine grades on async assignments. You will get worse grades, possibly much worse, on synchronous assessments. The latter category is growing.
Where I think we'll be a year from now
Predictions are mostly stupid but here are a few I'd bet on:
The "AI chatbot for X" category mostly consolidates. Some die. A few merge with broader study systems.
Calendar-aware planning becomes a basic expectation in study products, the way mobile apps are now.
At least one major university announces a return to handwritten in-class finals across most of its undergraduate curriculum. Several follow.
AI detection software stops being marketed as evidence and starts being marketed as a "discussion starter" — a tacit admission that it doesn't work.
Memory and personalization — the actual kind, not the marketing kind — becomes the main axis of competition between study products. Whoever's best at this wins the next few years.
That last one is what I'm betting on, professionally and personally. Software that knows you, that adapts to your term, that builds you a small specific plan every morning — that's a different shape than a chatbot. It's the shape that fits how studying actually works.
The students who figure this out early have a real advantage. The students who keep typing problems into ChatGPT will eventually run into a final exam that doesn't care what ChatGPT thinks.