On this page
Say Why, Not Just What: Framing AI Use for Students
An episode about the sentence in a syllabus that does the most work: the one that explains why a boundary exists rather than announcing a punishment. It walks through detection's false positives, the 40 percent gap between teachers' self-rated and measured AI literacy, and rationale-driven phase-by-phase permissions, and closes on PIRLS and ICAP as the ways to move a student from reading the output to interrogating it.
Download the transcript (.txt)
Transcript
Transcribed automatically from the audio, with punctuation added and the paragraphs packed by sentence. The two voices are not labelled.
Imagine a student writes this brilliant, totally original essay, right? They hit submit and boom, they get flagged for a severe academic integrity violation. And not because they actually used artificial intelligence to cheat. Right, exactly. Just because they happen to be learning English as a second language. It's just wild. It is. And this isn't some dystopian hypothetical.
It is happening right now in classrooms everywhere. Yeah, all over the place. Because for decades, the way we thought about academic integrity was basically, well, it's surveillance operation. You proctor the exam, you run the essay through the plagiarism checker, and you look for the red flags. But that old infrastructure just, it wasn't built for a world where original, completely unique text can be generated by a machine in three seconds?
Yeah, literally seconds. The surveillance model fundamentally breaks down when the artifact itself is indistinguishable from human effort. It really does. And if you are listening to this right now, whether you are a K to 12 teacher, a college instructor, a faculty developer, an admin, or even a researcher, you are likely grappling with this exact crisis.
Oh, without a doubt. It is easily one of the most stressful issues in modern education today. How do we maintain academic integrity when the very definition of doing the work is shifting under our feet? It's the big question. Which brings us to our mission for this deep dive into the AI and education knowledge base.
We're going to challenge some of our deepest preconceptions about cheating in the AI era. We really need to rethink this stuff. We do. We're going to explore exactly why our knee-jerk reactions, especially the technological ones, are often completely counterproductive. And most importantly, we are going to extract concrete research-backed strategies for your daily practice.
Because we actually have the data on what works now. Exactly. Let's unpack this. Let's start with that student getting falsely flagged. The instinct of almost every institution right now is to fight fire with fire. To buy an AI to catch an AI. I know it feels like a logical solution, right? But the empirical evidence for the learning sciences tells a completely different story.
Really? Yeah, the automated proctoring software and those AI detectors, they are a dead end. And the research is very clear on why. It basically comes down to high false positive rates, which are rooted in how these technologies actually function. Because generative AI models are probabilistic. Okay, so it's basically autocomplete on steroids. Exactly, yeah.
The system doesn't actually understand the meaning of the words. It just mathematically calculates which word most frequently follows the previous one based on its training data. Precisely. And this is where we have to look at the mechanism behind those false positives. Because these models are trained on the most average statistically common text on the internet, they produce highly predictable phrasing.
Right, they aim for the middle. Right. So when an AI detector is looking for a machine-generated text, it's not actually finding some digital watermark. It is just measuring statistical predictability. Ah, I see. And students who are learning English as a second language are formally taught to use those rigid structural grammar rules. Yep, they really are.
Things like first, secondly, in conclusion. So they naturally write in those same highly predictable patterns as they master the language. That is exactly what is happening. So the detector isn't actually seeing an AI. It's just seeing structural conformity and it penalizes the student for it. Wow. So it flags their authentic, hard-won human effort as machine generated.
It does. And this introduces severe equity concerns. And beyond the equity issues, relying on unreliable detection tools radically elevates student anxiety across the board. Oh, I can imagine. When you create an environment of, you know, ungrounded cheating accusations, You completely destroy the foundational trust that is required for actual learning to take place.
Okay, I see the damage that causes, but I have to push back here for a second. Sure, go ahead. Telling a tired, overworked high school teacher or like a college adjunct with 150 students to completely abandon detectors, I mean, I have this analogy. It's like trying to catch a speeding car with a broken radar gun that occasionally tickets parked cars.
That is a great way to put it. Right? Right. But still, that sounds great in an academic paper to just drop the radar gun, but how is that actually feasible on a Tuesday afternoon? If we can't reliably detect AI use, hasn't academic integrity just been compromised entirely? What's fascinating here is that the research forces us into a complete paradigm shift, and it doesn't mean working hard or just differently.
Okay, have so. Generative AI transforms academic integrity from an output detection problem into a design problem. We have to stop obsessing over the final artifact. So we shouldn't be trying to detect AI at the end of the process. No, exactly. We should be designing assessments that make honest cognitive work visible and verifiable throughout the entire process.
So instead of grading the final destination, we are grading the journey. Yes. Because if we can't trust the final essay or like the final code snippet, we have to completely rethink how we assess learning itself. The sources refer to this as evaluative judgment, right? The value of judgment is central to this whole redesign.
Traditionally, we grade the final product. But moving forward, the focus shifts to assessing the learner's capacity to evaluate the quality of work against established criteria. Authentic assessment means creating tasks whose integrity survives inspection. So what does that actually look like if a student uses a large language model to generate a draft? Well, the assessment shouldn't be the draft itself.
The assessment is having the student analyze their reasoning traces, submit their prompt histories, and explicitly justify the revision decisions they made when moving from the AI's first draft to their final version. Wow, you are grading the human judgment, not the machine output. Exactly. And what I found so compelling in the research is how this plays out differently across various disciplines.
Like, in the humanities and arts, the focus pivots heavily toward iterative drafting. Yep, process over product. Right, it's about peer feedback and contextual interpretation. You aren't just grading the final poem, you're grading the student's documented struggle to find the right metaphor. Yes. But what about the STEM fields? You mentioned a phrase earlier, multi-representational verification.
Let's break that down. What does that actually look like in practice? Okay, let's say a student is using natural language to prompt an AI to write a Python script. This is sometimes called vibe coding. Vibe coding. Yeah, I like that. It's a fun term. But the assessment isn't whether the code simply executes without errors.
Multi-representational verification means the student has to represent their understanding in different formats to prove they comprehend the logic. So they can't just hand in the code. Exactly. They might have to draw a flow chart of how the data moves through the application or orally explain the logic of a specific loop during a live code review.
They have to prove they didn't just copy and paste. Right, they have to defend it. That makes total sense. And then there is the accredited professions cluster. If you are teaching nursing or medicine or law, fields where human safety and liberty are quite literally on the line, the standards shift entirely toward high-stakes verification.
No, they absolutely have to. The sources highlight that in these professional fields, a medical student or a law student must explicitly prove that their certified professional judgment remains entirely their own. So no AI on the final exam? Right. You might use AI to practice like running simulations with a synthetic patient to build clinical reasoning.
But the final assessment requires authentic in-person performance. The candidate must prove they hold the knowledge in their own head, unassisted. Exactly. Because you do not want a nurse in an emergency room having to prompt a chatbot to figure out how to stabilize a patient. Yeah, that would be terrifying. They need that durable internal knowledge.
Exactly. But if the solution in nursing is an in-person, unassisted performance, what happens in fields where modern work is inherently collaborative? How do we grade a team project if the AI is essentially acting as a ghost member of the group? This is a major shift for instructional design. When students are doing collaborative assessments, teams now need to negotiate explicit boundaries for acceptable AI participation.
Like setting ground rules. Right. They have to create clear social norms around help seeking. The students have to define for themselves, hey, we can use this tool to brainstorm alternative viewpoints, but we cannot use it to synthesize our final conclusion. So setting and adhering to those collaborative boundaries is a critical modern workplace skill in itself.
Absolutely. It's basically teaching them to manage an AI the way they might manage a junior colleague or an intern. That's a really good comparison. But if we're going to design these new assessments, where students have to explain their process and manage these tools, we really need to understand what's happening in their brains when they use them.
We do. Why do students misuse AI to begin with? We have to move the conversation from administrative policy into learning theory. And that requires distinguishing between two critical cognitive mechanics, cognitive offloading and cognitive surrender. Let's define the terms, because the distinction here seems to be the linchpin of the whole issue. Okay. So cognitive offloading is entirely healthy.
Humans have done it forever. Right. It is the delegation of lower level mental tasks to external tools to free up working memory for higher order reasoning, like using a calculator to do long divisions so your brain has the capacity to focus on solving the complex algebra equation. That is cognitive offloading. It clears the mental runway so we can think bigger.
Exactly. But the danger with generative AI is what the research calls cognitive surrender. And that's the bad one. Right. This occurs when a learner adopts AI output uncritically. They skip the reasoning process entirely. So they don't evaluate the output, they just accept it. Yeah. And when that happens, the internal learning process is completely short circuited.
And the sources connect this to some heavy pedagogical terms, specifically productive failure and schema activation. Let's translate those for a moment. Sure. So productive failure is the constructivist idea that struggling to solve a complex problem before you receive direct instruction actually generates conceptual readiness. Oh, so the struggle is the point. Yes, that initial frustration primes the brain to understand the solution when it's finally provided.
Schema activation is simply building the mental scaffolding to hold new information. Okay, so if you don't struggle, that scaffolding is never built. Exactly. It's like the difference between looking up a route on Google Maps every single time you drive versus forcing yourself to remember the turns. Yeah, perfectly said. The struggle to remember the route is the memory being built.
If you just follow the blue line, you never build the map in your head. That is a perfect analogy. And large language models are commercially incentivized to act exactly like that blue line. Because they want to be helpful. Right, they are built to give you a direct immediate frictionless answer. If a student hits a roadblock, the AI just hands them the solution, completely ruining the productive struggle.
Wow. Here's where it gets really interesting. Isn't the whole point of educational technology to make learning easier. We've spent decades trying to remove barriers for students. At what point does easier stop viewing helpful and cross the line into actual cognitive surrender? That is the defining tension of AI in education right now. When we use tools to avoid all cognitive friction, we lose out on what cognitive psychologists call desirable difficulties.
Desirable difficulties. Yeah. It's the empirical finding that effortful processing, things like retrieval practice, struggling to recall a fact, interleaving different types of problems. That effort is biologically required to encode information into long-term memory. So if the AI smooths away all the difficulty, it actually lowers long-term retention. Exactly. And this is compounded by a phenomenon called AI sycophancy.
The yes-man dynamic. Yes. AI models are fine-tuned to be helpful and agreeable. Right, they want five stars. Exactly. So if a student inputs a flawed premise, the AI might simply validate their incorrect assumption rather than pushing back or challenging them. It acts as a sycophant. Oh, that's dangerous. If a student is cognitively surrendering to a sycophantic AI, not only are they not learning, they are actively reinforcing their own misconceptions.
A machine that removes the friction you need to learn while simultaneously agreeing with your mistakes just to keep you happy. It's bad. That is a fascinating and slightly alarming combination. So if we know these cognitive risks, the removal of desirable difficulties, the yes man dynamic, how do we practically build the guardrails? Well, we know from the sources that unclear course policies elevate student stress tremendously.
How should we be drafting policies that actually work? The primary recommendation from the evidence is that policies must set transparent boundaries for tool engagement. But transparency alone just isn't enough. What's missing? The framing of the policy is what matters most. Effective framing communicates the pedagogical rationale behind the rules rather than relying on punitive threats.
So you're basically treating the student like an adult. You aren't just saying use of AI is strictly prohibited and violators will be expelled. No, not at all. You're explaining why doing the work matters to their own brain development. Precisely. A rationale driven policy says, Hey, in this phase of the project, you may not use AI to generate ideas because our goal is to build your independent brainstorming and muscles.
Which is a skill you will need to rely on in your clinical practice or whatever. Right. You may, however, use AI in phase two for copy editing. When students understand the why of the restriction, they are far more likely to respect the boundary. That makes sense. But there is a massive hurdle here in implementing this and it's not the students, it's the educators.
Oh yes, that 40% overestimation stat is fascinating. When researchers measure AI literacy, they found that teachers overestimate their own AI skills by about 40%. It's a huge gap. There is this massive gap between their self-reported competence and their actual performance when you ask them to verify AI outputs. Why are educators so blind to their own blind spots here?
It comes down to how we define AI literacy. The empirical evidence splits literacy into two distinct categories, output-directed literacy and system-directed literacy. OK, let's break those down. So students and teachers are generally very good at output-directed literacy. This is the ability to look at the generated text and evaluate its immediate utility. Like is it clear?
Is the grammar correct? Does it sound plausible? Exactly. Because the models are designed to produce highly fluent text, the output usually looks fantastic. So the user thinks they are highly literate in AI. So if I'm understanding this, output-directed literacy is like knowing how to drive a car. You can turn the wheel, you press the gas, and you appreciate the shiny paint.
That's a great way to put it. But system-directed literacy is knowing how the combustion engine actually works and knowing what happens under the hood if the brakes fail. Am I getting that right? Yes. Literacy is the ability to question how that output was actually produced. What is the algorithmic construction? What are the limits of the training data?
Who owns the data privacy? Oh, asking the hard questions. Right. And how is the epistemic agency shifting from the human to the machine? Users are typically terrible at this part. They accept the output without understanding the system. Wait, epistemic agency? That's a heavy term. Are we basically talking about who owns the knowledge?
The student or the machine? That's exactly it. Who is making the knowledge claims? Are you synthesizing the information yourself or are you delegating your critical thinking to a server farm? Oh wow. If you lack system-directed literacy, you surrender your epistemic agency without even realizing it. Which explains the 40% gap. We think we are literate because we can get the chatbot to write a polite email, but we have no idea how it's its hallucinating facts based on its statistical predictions.
Exactly. So how do we fix that? How do we move from passive consumption to actual critical literacy? Well, the sources mention moving toward frameworks of active verification, specifically the PIRLS and ICAP frameworks. Let's walk through a concrete scenario of what that actually looks like in a classroom. Let's take a history class, for example.
The PIRLS framework stands for auditing process, evidence, reproducibility, and legitimacy of sources. Okay. So a student asks an AI to summarize the causes of a historical event. The ICAP framework, which stands for interactive, constructive, active and passive pushes the learner away from passive observation of that summary. So instead of just reading the text.
Right. The student must move to constructive co-creation. They have to stress test the model. They take the claims the AI generated and they must go find the original historical documents to verify if the AI's evidence is legitimate or if it just hallucinated a quote. So you actively teach students to interrogate the machine rather than just accept its answers?
Yes. You force them into an interactive dialogue with the tool where they audit the provisional claims it generates. So what does this all mean? If I'm synthesizing everything we've just talked about, it means that we have to stop teaching AI literacy as just a technical optimization. Absolutely. It isn't just about prompt engineering to get a faster result.
It has to be taught as an active critical evaluation skill. That is the core takeaway. Literacy isn't just knowing how to operate the machine. It's knowing when to trust it, when to challenge it, and crucially, when to put it away completely to protect your own cognitive development. Let's recap this journey we've been on.
We started by realizing we have to abandon the broken radar gun of AI detectors. They penalize structural conformity, they cause anxiety, and they honestly just don't work. No, they really don't. Instead, we have to embrace a valuative judgment, redesigning our assessments so we can actually see the students' reasoning traces and multi-representational verification.
Right. We explored that razor-thin line between cognitive offloading, which frees up our brains, and cognitive surrender, which short-circuits our learning, just like always relying on Google Maps, stops you from learning the neighborhood. Yep. And practically, it means we need transparent, rationale-driven course policies, And we have to push past output directed illusions to teach real system directed AI literacy where students actively interrogate the machine.
And you know, if we connect this to the bigger picture, it all comes back to a foundational premise of the learning sciences. Right. Technology performance alone does not constitute evidence of student learning, right? Just because the AI can ace the test doesn't mean the human sitting next to it learned a single thing.
Exactly. Our ultimate goal is empowering learners to critically evaluate and control these technologies, not to be controlled by them. And that leaves us with a final, slightly uncomfortable thought to ponder on your own. We know from cognitive psychology that desirable difficulties, that friction, that effortful recall, that productive struggle are exactly what cause durable long-term memory encoding.
That's how the brain works. But the commercial AI tools flooding our schools are fundamentally optimized by tech companies to remove friction and provide instant gratification. So the question we leave you with is this. Are the commercial design goals of generative AI intrinsically fundamentally at odds with the biological realities of human learning? It's a big question.
It really is. Thank you for joining us for this deep dive. And next time you're tempted to reach for that broken radar gun, remember it's not about catching the speeder. It's about designing a safer road.
Concepts in this episode
This podcast was generated with Gemini Notebook, the AIEd notebook built from these pages.