This is Part 2 of a 3 part series. You can find Post 1 here.
It's Not AI That Broke Assessment #
In the last post I told you about a room full of higher-ed educators who went quiet when I suggested assignments shouldn't have a word count. I'm starting here with the same story turned a different direction, because I think it's the key to the whole cheating panic: it's not AI that broke assessment. Assessment was already broken. AI just pointed a spotlight on the issue.
Word counts, page minimums, five-paragraph essays, take-home exams nobody proctors: all of it was built assuming a certain kind of friction, that producing the words was hard, so requiring a lot of them proved effort. AI removed the friction. What's left is a bunch of assignments that were never actually measuring the thing they claimed to measure, and now everyone can see it.
That's not a new problem AI created, either. Grading the outcome instead of the process has always pushed students toward the wrong behavior. It might produce fine-looking results for the student who tests well or writes fast, but it was never built to capture what actually matters: how a student got to the answer, what they revised along the way, what feedback they used and what they ignored, how many attempts it took before something clicked. Outcome-only grading was never measuring any of that, for any student, and it definitely wasn't measuring it evenly across all of them.
Cheating and Plagiarism #
Let's start with the loudest fear, because it deserves to be taken seriously: cheating.
The numbers back up the alarm. In a national survey of college faculty, 78% said cheating has increased since generative AI became widely available, and 73% have personally dealt with an academic integrity case involving it. A Brown professor recently said, flatly, that he suspects most of his class used AI to cheat. This is where the panic is loudest, and for good reason: a college degree is supposed to certify that a person can actually do something, and there's a lot riding on that certification staying honest.
The solution? Detectors. Except for the fact that, as it turns out, the detectors themselves are bad. A Stanford study tested seven of the most widely used AI detectors against real student writing and found they flagged 61.3% of essays written by non-native English speakers as AI-generated, against a 3.2% false-positive rate for native speakers. Simpler vocabulary and more formulaic sentence structure (completely normal for someone writing in a second language) reads to these tools exactly like a machine. So the "solution" a lot of schools reach for first, install a detector, is actually a machine for punishing English-language learners specifically. That's not a minor bug. That's disqualifying, and it's probably why most professors don't actually rely on these tools in practice. They trust their own read of a student's writing instead, and given the alternative, that's the right instinct.
Plagiarism belongs in this same conversation, and it's the same question wearing different clothes: whose words are these, actually? That's not a new problem AI invented, either. We've had no shortage of high-profile plagiarism scandals in recent years, university presidents, authors, journalists, none of it needed generative AI to happen. AI didn't create academic dishonesty. It gave an old problem a faster way to happen and a bigger microphone.
If detection isn't the answer to either problem, what is? I've landed on a version of something math teachers have required forever: show your work. In my own high school coding classes, I let students use AI to write code, on one condition: they have to show me every prompt they used, explain why they wrote it that way, how they tweaked it, how they validated the output actually worked, and how they made it theirs rather than just accepting it. That's not a workaround for AI. That's just what "did you actually learn this" looks like once the answer alone stopped being proof of anything. Document the process instead of just the product, and both cheating and plagiarism get a lot harder to hide. I want to add a caveat here, because I've been burned by the flip side of this exact idea. In school, "show your work" frequently got me in trouble, because I did a lot of math in my head, and a teacher who assumed there was one correct method to arrive at an answer marked me wrong for not showing steps I hadn't actually taken. That's the old-mindset failure from the top of this series, recurring in a new form. If "show your prompts" turns into "there's one approved way to prompt," we'll have rebuilt the same trap with different materials. The point isn't to mandate a process. It's to require honesty about whatever process actually happened.
Critical Thinking #
The next fear on the list is critical thinking: hand a kid an AI that thinks for them, and they never learn to think for themselves.
This is the calculator panic, running the same argument on new hardware. Handheld calculators showed up in classrooms in the 1970s to a chorus of nearly identical warnings: kids would forget how to do math, number sense would atrophy, a whole generation would grow up mathematically illiterate. Fifty years later we still teach arithmetic by hand before anyone gets a calculator, and nobody seriously argues schools should have kept them out. The tool didn't replace the thinking. It changed what the thinking got spent on.
Show-your-work is the answer here too, and it's worth saying outright: explaining why you wrote a prompt the way you did, evaluating what came back, how you iterated and fine-tuned, and knowing when to trust it and when not to, is itself a critical thinking exercise, arguably a harder one than solving a problem that already has one correct answer waiting in the back of the book.
Screen Time #
Another concern that is popping up with increased frequency: screen time. Kids are already in front of screens more than any generation before them, and piling AI-assisted schoolwork on top of that looks like it can only make things worse.
I don't think it does, for the same reason I laid out in the first post: the pendulum's already swinging back. Schools are reversing course on screens generally, not because of AI specifically. Maine's laptop initiative went fifteen years without moving test scores. A Kansas middle school quietly took its Chromebooks back. Sweden is spending a small fortune replacing tablets with textbooks. That correction started before generative AI showed up.
I'm not arguing for more screen time. The screen time already exists, mandated by the same one-to-one programs schools are now walking back. The real question was never whether a kid is in front of a screen, they already are. It's what they're doing there. AI doesn't add a screen to anyone's day. Used well, it's one of the few things that can make the screen time already sitting in a kid's backpack actually worth something.
Human Connection #
I am not going to argue with this one, because it isn't wrong: more AI in a classroom could mean less human connection, and connection is a big part of what makes school work at all.
Here's the context that's missing, though. Only 22% of secondary students say most or all of their teachers make an effort to understand their life outside of school, an all-time low, down from 43% right after the pandemic hit. That decline happened between 2020 and 2022, years before generative AI showed up in a classroom. The relationship problem already existed. Blaming AI for a shortage that predates it lets the actual cause off the hook.
I already made this concession once, back in the first post: Dan Meyer's line that Khanmigo doesn't love a kid, it can't, is true and worth taking seriously. But the honest version of that argument cuts the other way too. Every hour AI gives a teacher back from grading and paperwork is an hour that can go toward the one thing AI genuinely can't do: sitting with a kid and paying attention. More on that in the next post. Whether that hour actually gets spent on connection is a choice, not a guarantee, but it's a choice most teachers don't currently have the time to make at all.
I'm not sweeping this one under the rug, either. In-school connectedness has measurable effects on resilience, stress, and academic outcomes, and as this technology reshapes the classroom, it's one thread we can't afford to lose.
Equity #
Equity is a fear I hear constantly, especially from people who consider themselves progressive on education: AI will help kids who already have advantages and leave everyone else further behind. I respect the instinct, but I have to disagree, at least partially.
There's a real academic debate here, and I don't want to pretend there isn't: plenty of serious people are worried that AI adoption in well-resourced schools is widening the gap for rural and high-poverty districts that lack the devices or connectivity to keep up. That specific problem, raw access, is completely legitimate, and it's the one I take seriously, more than cheating, more than privacy. Implementation isn't free. The lightweight tools are cheap, but the adaptive, personalized systems this series keeps pointing to can run schools tens of thousands of dollars, before training staff and maintaining the thing, and the pricing is still moving, fast, and not always down; even the AI companies themselves are still adjusting their own cost models in real time. If the gap between what a wealthy district can afford and what a poor one can afford ends up deciding which kids get the benefit, that's a real failure. It's a policy and funding problem, and one worth watching closely, not waving away.
But the argument I actually hear most often isn't about access. It's the claim that using AI at all is somehow unfair, because it helps kids who already have advantages. Even if that premise were true, I'd still reject it: "I'd rather nobody get the benefit than let some kids get it before everybody can" is a great way to guarantee a permanent race to the bottom. But I don't think I even need that argument, because I think the premise itself gets it exactly backward. AI is one of the only things I've seen that lets a kid without a lot of resources, or without a grownup who can help polish an essay, or without any perceived natural talent for drawing, produce something they're genuinely proud of. It's a bridge, not a moat.
I got a real-world version of this argument recently. Like many countries, Brazil struggles with the ever-widening education gap between the haves and the have-nots. Unlike many countries, however, Brazil is prioritizing addressing this need and has created laws and frameworks to support the effort. I had the honor of meeting with Brazil's Secretary of Education, and one of the points he made was that part of the Ministry's interest in AI is in how it can function as a democratizer: a way to narrow the equity gap, not widen it. It was thrilling to hear a government official responsible for tens of millions of students recognize and articulate the opportunity.
I'll come back to where I specifically think AI's benefits shine the most in the next post.
Privacy #
Then there's privacy, the fear that deserves the most nuance. This one's real. Kids' data ending up in places it shouldn't, FERPA and COPPA both straining against how these tools actually collect and process information, a startling number of teachers, under half, having received any training on using AI safely with student data. That's a genuine gap.
What I won't accept is treating this as an AI-specific problem, because it isn't one. Every school already handed a mountain of student data to Google Classroom, to whatever LMS the district picked, to every ed-tech vendor with a signed data-sharing agreement, and largely didn't ask hard enough questions about any of it. If we're suddenly worried about where student data goes, good. We should have been worried the whole time. Singling out AI as the uniquely dangerous one, while treating everything else as settled, isn't caution. It's a dodge. The fix is the same fix that was always needed: real vendor vetting, real data governance, real training, and all applied to AI the same as everything else it's sitting next to in the tech stack.
The Honest Cost #
I'll end this post with the concern I take most seriously, because it's the one that actually complicates my own argument. Grading an exam is fast. Grading a process (reviewing someone's prompts, their revisions, their progress and workflows, their reasoning about why they made the choices they made) is slow, and it asks more of a teacher who's already stretched thin. And the last thing I want to do is to burden already overworked, overstressed educators with additional tasks. I spend a lot of time thinking about this one because it's a real cost and I'm not going to pretend it isn't, and I don't have a good answer, not yet.
But a thing being harder to grade isn't an argument that it's the wrong thing to grade. It just means we've been taking the easy measurement for a long time, and it's about to stop being available to us. Good.
Which brings us to where I want to see AI used, and that's the subject for the final post in this series.