Transcript: The Greed of Man - The Hidden Cost of AI
Part 1: The Hook & The Roadmap
[SCENE START: Music fades in under the dialogue]
Seikoh: Welcome to the Newest Episode of The Greed of Man: The Hidden Cost of AI. I’m Seikoh Murakami.
Selim: And I’m Selim Liu. We are coming to you from the University of Toronto Mississauga, and today, we’re tackling a topic that has basically turned the academic world upside down over the last couple of years.
Seikoh: Exactly. We’re talking about Generative AI. But we are not here to give you the standard "AI is changing everything" lecture. We want to look at a very specific, almost invisible paradox that happens when a student opens up agents like ChatGPT to help them with their schoolwork.
Selim: Right. Because on the surface, it looks like a win-win. The student gets their work done faster, the AI provides clear explanations, and grades might even go up. But beneath that surface, there’s a massive "cognitive trade" happening.
Seikoh: That’s the perfect way to put it, a "cognitive trade." In my half of this investigation, I’ve been looking into what happens to the studying process itself. If a student uses AI to summarize every reading, generate every practice question, and plan their entire exam schedule... what is left of the student’s actual knowledge? Does an "A" on an exam still mean the same thing if the machine did all the heavy lifting in the weeks leading up to it?
Selim: And that leads directly into the second half of our show. My research picks up where Seikoh’s ends. Because if we agree that students are "offloading" their independent thinking to these models, we have to ask a much more dangerous question: What is filling that void?
Seikoh: That’s the part that really surprised me when we started comparing our research notes. It’s not just an empty void, is it?
Selim: Not at all. When you stop thinking for yourself, you don't just "stop thinking." You start inheriting the assumptions, the values, and the cultural frameworks of the machine. I’ve spent the last few weeks looking into the data science and the sociology behind this, and what I found is that these AI models aren't neutral tools. They have a very specific "Western-centric" bias hardcoded into their DNA.
Seikoh: So the student who thinks they are just getting a "neutral" study summary is actually being fed a localized, Western worldview without even realizing it.
Selim: Exactly. It’s a psychological trap. We’re going to be talking about things like "Automation Bias”, which is really just a fancy way of saying we trust a computer’s voice more than a human’s.
Seikoh: And we’ll get into the heavier stuff too, like "Cognitive Imperialism," and "Data Colonialism." It sounds like a history lesson, but we are actually talking about how AI might be colonizing the way we think, replacing our local cultures, our beliefs, and values with a generic, Western version of ‘truth’.
Selim: It is happening in real-time. Every time a student at UTM hits “enter” on a prompt toward these agents, they are inviting these invisible biases into their brains.
Seikoh: So, here is the roadmap for today’s conversation. First, we’re going to dive into the Pedagogical Crisis. I’ll be walking Selim through the research on "Decremental Learning" and how AI might be making us "better" students but "worse" thinkers.
Selim: Then, we’ll pivot to the Ideological Crisis. We’ll look at the technical "security analysis" of the datasets used to train these models,basically, I’m going to show you the "math" behind the bias. And we’ll discuss the "allure" of the AI’s voice, why it’s so easy to trust a machine even when it’s wrong.
Seikoh: And finally, because we don’t want to leave you in a state of total panic, we’re going to close with some Actionable Strategies. We’ll talk about "Critical Posthumanist Literacy", which is basically just a fancy term for learning how to treat AI as a partner you still need to keep an eye on. and a technique Selim found called "Cultural Prompting" that can actually help you reclaim your independent thought.
Selim: We’ve got about 10 peer-reviewed sources each,20 total, that we’re going to be synthesizing today to make sense of this. So, let’s get into it. Seikoh, let’s start with that student you were telling me about earlier. What happens when Chris decides to let ChatGPT handle his midterm prep?
Part 2: Seikoh’s Segment - The Assessment Crisis
Seikoh: Let’s look at a hypothetical student named Chris. Chris is facing a massive midterm, so he takes his lecture slides, those dense PDF readings, and his syllabus, and he dumps them all into ChatGPT. He asks the AI to map out a study schedule, condense the main points into bullets, and generate a 50-question practice quiz. He memorizes the answers, walks into the exam hall, and scores an 85%.
Selim: And on paper, the university registrar sees that 85% and assumes Chris is a success story.
Seikoh: Exactly. But my research pushes us to look at what Chris didn’t do. In a traditional setting, studying involves self-regulated learning. There is this really fascinating 2026 paper by Panadero and Broadbent that basically became the foundation for how I look at this.
Selim: Right, you mentioned their work on decremental research earlier, basically looking at what we’re losing in the process.
Seikoh: Exactly. Panadero and Broadbent argue that generative AI can actually displace those self-regulated learning processes. Studying isn’t just about absorbing facts; it’s about planning, monitoring your confusion, identifying weak spots, and adjusting your strategy. That friction is part of the learning. When Chris uses AI to instantly generate summaries, study plans, and practice questions, he is bypassing the effortful thinking needed to actually internalize that knowledge. Their point is that we can’t just look at what AI improves. We also have to look at what disappears underneath that improvement.
Selim: It’s like going to the gym and having a robot lift the weights for you, but still expecting your own muscles to grow.
Seikoh: That’s the perfect analogy. The task is finished, but there’s no growth. Panadero and Broadbent basically force us to separate performance from learning. Chris might perform well, but that doesn’t automatically mean he has built knowledge he can still use when the tool is gone. That’s why I keep coming back to this question: what could Chris actually do without the machine in the room?
Selim: This brings up a critical paradox. If Chris uses ChatGPT to build the schedule and summarize the text, and then he gets that “A” on the midterm... whose knowledge is that exam actually measuring? Did Chris learn the material, or is he just mirroring the machine’s understanding?
Seikoh: That’s a great question, Selim. Honestly, I think the “A” still shows he retained something, but it no longer clearly tells us how much of that learning was independently built by him. He may have memorized the machine’s synthesis instead of constructing understanding for himself. So the issue isn’t just “did he learn or not?” It’s that performance and learning are no longer the same thing.
Selim: But if that’s true, how do professors even grade this fairly without just accusing everyone of cheating?
Seikoh: That’s exactly what makes this an institutional crisis. Researchers like Corbin and colleagues argue that maybe the problem isn’t that professors need better cheating detection. Maybe the problem is that the task itself no longer makes the boundary clear enough. Their solution is really interesting,they say the boundary should be built into the assessment itself, instead of being guessed after submission. That could mean oral components, live explanation, staged drafts, or reflective follow-ups, where the professor is grading visible understanding rather than just a polished final answer. Their oral-exam example is useful because it shows that a better design can make the intended skill clearer without relying only on surveillance.
Selim: So the exams themselves are fundamentally broken?
Seikoh: In some cases, yes. Corbin and colleagues basically shift the conversation away from the panic of “Did they cheat?” toward a deeper question: “Does this assessment still measure what it was supposed to measure?” If AI can help produce outputs that do not genuinely reflect students’ own competencies, knowledge, and skills, then the grade starts to lose its meaning.
Selim: It’s an existential crisis for the university. If an “A” doesn’t mean the student understands the topic, the grade is basically an illusion.
Seikoh: Exactly. It’s like trying to measure height with a broken ruler. And this isn’t just theoretical. Walton and colleagues show that AI is already deeply embedded in how students prepare for and complete assessment tasks. What I found especially interesting in their research is that students are not only judging AI outputs, they are also making judgments about their own weaknesses while using the tool. Sometimes they rely on AI for things they could not otherwise do, adopt ideas with low criticality, or even start misjudging the machine’s contributions as their own.
Selim: I see it in the library every day.
Seikoh: Right. And because it’s becoming normal, Seeley and Cournoyea push us to ask a much bigger question: what skills are universities even trying to measure now? If AI can summarize the reading, organize the studying, and generate the practice material, then is the exam measuring knowledge, memory, tool use, or some messy combination of all three? They also warn against treating GenAI like it’s just another calculator moment. Their point is that this technology is arriving faster, with more uncertainty, and with bigger ethical stakes.
Selim: Let me play devil’s advocate. The real world uses AI for efficiency. If ChatGPT makes me study faster, isn’t that just meeting the goal of the corporate world?
Seikoh: It’s a valid pushback. We aren’t advocating for quill pens. Su and Yang help frame this by saying the level of automation has to match the learning goal. In plain English, if the task is something low-stakes like formatting or basic organization, maybe automation is fine. But if the actual goal is to develop interpretation, argument, or critical thinking, and AI takes over all of the reading and synthesis, then you’ve traded cognitive development for convenience.
Selim: So it’s a substitute versus a support.
Seikoh: Exactly. Anders and Dux Speltz argue for a human-centered model. They see AI as something that can scaffold learning without replacing student agency, but only if the student is still the one actively planning, iterating, and evaluating. That’s why I liked their example of students creating a “fluent hallucination.” It teaches them that AI can sound confident and intelligent while still being completely wrong. So the real goal is not to reject AI, but to stop surrendering judgment to it.
Selim: But during a 2:00 AM cram session, who has the energy to “actively evaluate”? You just want the answer.
Seikoh: And that’s the scariest part. Walton and colleagues found that students often adopt AI-generated ideas with low criticality. McPhee and Jerowsky push that even further by arguing that technical skill is not enough. Students need AI tool assessment, critical AI evaluation skills, and AI information literacy. Otherwise, they choose tools because they are fast and convenient, not because they are actually appropriate for the task.
Selim: They forget where their brain ends and the AI begins.
Seikoh: Exactly. And that’s why Wang and Wang are important too. They argue that AI should not be treated either like a human authority or like “just a tool.” Both extremes are misleading. Their point is that students need to interrogate AI-generated text as something co-produced, not something neutral. Once you stop seeing the output as neutral, you start asking better questions about who shaped it and what assumptions are built into it.
Selim: It sounds like we need a totally new type of literacy to survive university now.
Seikoh: Honestly, yes. Tzirides and colleagues call for reflective AI literacy. What I found useful there is that they move beyond vague slogans and actually suggest reflective practice, asking students where AI-generated ideas come from, what assumptions they carry, and whether they can be challenged. They call that epistemic provenance, and I think that’s huge. Cardon and colleagues also make this practical by framing AI literacy around authenticity, agency, accountability, and application. So the issue is not just whether students use AI, but whether they understand what the tool is doing to their learning.
Selim: So if banning it is off the table, what’s the takeaway?
Seikoh: Banning is bad pedagogy. The takeaway is that universities have to be much more intentional about what can be outsourced and what must remain human. Better assessment design matters. Reflective AI literacy matters. And student judgment matters. The real question isn’t “Did they use AI?” It’s “Are students still doing enough of the thinking for this exam result to actually mean something?”
Part 3: The Bridge - The “Trojan Horse”
Selim: Exactly! We are policing the output, but completely ignoring the input. Which leads me to a critical follow-up: considering the psychology of the students you researched, do they even realize this is happening? Or is the AI’s authoritative tone just too seductive?
Seikoh: From what I’ve read, they absolutely do not notice. Think about the physical state of a university student,they’re sleep-deprived, stressed, and overwhelmed. When a machine spits out a perfectly formatted, highly confident answer in three seconds... you don't question it. You just feel relief. The tone is designed to sound like an absolute authority.
Selim: It’s designed to sound like the smartest professor in the room.
Seikoh: Exactly. So why would a tired 19-year-old argue with it? But Selim, you actually researched the data science behind this. Why is that tone so dangerous? Why do we trust it so reflexively?
Selim: There is actually a scientific term for this phenomenon: Automation Bias. I was reading a 2025 paper by Nazaretsky and her colleagues, and they ran a massive study on how students react to feedback from machines versus humans.
Seikoh: And what did they find?
Selim: They found that students are highly susceptible to the “allure” of the machine. Even when the AI is wrong, we’re biased to trust it simply because it came from a computer. It’s a psychological trap where we assume machines are purely logical and devoid of error. It’s like following your GPS directly into a lake just because the digital voice sounded so certain.
Seikoh: That is terrifying in an academic context. If you follow a GPS into a lake, you know you made a mistake. But if you follow an AI into a biased understanding of sociology or history... you might never know. You just carry that bias as truth for the rest of your life.
Selim: Exactly. You internalize it. And because the AI is so polite and confident, Chris never stops to ask: Wait, who actually wrote this? What cultural perspective is this summary coming from? So, if the AI is a Trojan horse and Automation Bias is why we let it inside... what’s actually hiding inside the horse, Seikoh? Where is that Western-centric bias coming from? Is it intentionally programmed?
Selim: It’s not a conspiracy; it’s just pure math and data science. To understand the payload, we have to look at the “diet” the AI was fed before it ever spoke to a student. And when we look at the actual datasets training these models, the numbers reveal the truth.
Part 4: Selim’s Segment - The Bias Trap & Actionable Solutions
​
Seikoh: So, we’ve established that the AI is a Trojan horse, and that "Automation Bias" is the reason tired students willingly open the gates and let it inside their brains. But what’s actually hiding inside the horse, Selim? You mentioned a "Western-centric bias." Where is that coming from? Is there a programmer sitting in Silicon Valley intentionally typing out biased code?
Selim: That’s the most common misconception. It’s not a conspiracy, and it’s not one evil programmer. It’s just pure math and data science. To understand the payload inside the Trojan horse, we have to look at the "diet" the AI was fed before it ever spoke to a student. I was actually reading a 2024 security analysis by Inoshita and Zhou, and they looked under the hood of the massive datasets used to train Large Language Models, specifically datasets called C4 and OSCAR.
Seikoh: For our listeners who aren't computer science majors, what are C4 and OSCAR?
Selim: Basically, think of them as giant digital vacuums that scraped petabytes of text from the internet to teach the AI how to speak. But Inoshita and Zhou found a massive infrastructural flaw. Because the internet itself is heavily skewed toward Western, English-speaking servers, the data the AI learned from is mathematically tilted. Their security analysis proved that "pro-American" sentiment and negative non-Western biases aren’t just accidental glitches; they are fundamentally hardcoded into the foundational architecture of the models.
Seikoh: So, when you ask ChatGPT a question, you aren't getting an objective summary of human knowledge. You are getting a summary of the English-speaking internet's knowledge.
Selim: Exactly. It’s a localized epistemology dressed up as universal truth. And we actually have empirical proof of what that output looks like. There is a fantastic 2024 study by Tao and colleagues where they conducted a massive evaluation across the entire GPT model family. They wanted to see where the AI's "cultural alignment" naturally rested.
Seikoh: Let me guess. It didn't align with global diversity.
Selim: Not even close. Tao et al. found that all of these models consistently generate outputs that heavily resemble Protestant European and English-speaking cultural values.
Seikoh: That brings me to a question I really wanted to ask you, Selim. You use the phrase "Western-centered bias," and I think that’s a crucial point, but I want to slow it down a bit. What does that actually mean for our hypothetical student, Chris? What kinds of assumptions is he unknowingly absorbing when he treats those outputs like neutral knowledge?
Selim: Let’s put Chris in a sociology or history class. Let’s say he asks ChatGPT to summarize the impacts of the Industrial Revolution or the Cold War. Because the model defaults to Protestant European and American datasets, the summary Chris receives will likely highlight Western technological triumphs, economic growth, or Western political strategies.
Seikoh: Right, it's going to frame the narrative from the perspective of the "winners" of that specific historical dataset.
Selim: Exactly. It will quietly minimize or completely erase the perspectives of the colonized nations, the global south, or indigenous populations who experienced those exact same events in devastating ways. But because the AI writes it with such perfect grammar and authoritative confidence, Chris doesn't realize he is reading a highly localized cultural perspective. He just thinks, "Okay, these are the objective facts I need to memorize for the exam."
Seikoh: And that ties right back to my research on exam validity! If Chris goes into the exam and writes an essay using those exact Western-centric frameworks, and the professor gives him an "A"... the university is essentially validating and rewarding the erasure of diverse perspectives.
Selim: You hit the nail on the head. And that’s where my research pivots from data science into sociology. We aren't just talking about a technical glitch anymore; we are talking about a systemic pedagogical crisis. A researcher named Yaw Ofosu-Asare (2025) coined a brilliant term for this: "Cognitive Imperialism."
Seikoh: Cognitive Imperialism. That’s a heavy term. Can you break that down?
Selim: Historical imperialism was about conquering physical land and resources. Cognitive imperialism is about conquering the landscape of the mind. Ofosu-Asare argues that when AI replaces independent human thought, it doesn't just leave a blank space. It actively imposes a dominant Western epistemology that sidelines non-Western knowledge systems. It colonizes the way the student frames reality.
Seikoh: So, when Chris relies on the AI to do the thinking, he isn't just taking a shortcut. He is participating in an ideological shift.
Selim: Exactly. And it gets even deeper. I looked at another paper by Arora and colleagues (2023), and they discuss this through the lens of "Data Colonialism." They point out that this unquestioned reliance on AI replicates historical power imbalances perfectly.
Seikoh: How so? How does using ChatGPT replicate colonialism?
Selim: Think about how the AI is built. Arora et al. highlight that the raw data is extracted globally, but the models are owned, refined, and monetized by a few massive tech corporations in the Global North. Furthermore, the algorithms are often "cleaned up" and made safe by invisible, low-wage workers in the Global South, places like Kenya or the Philippines, who spend their days filtering out toxic data so Western users have a polite chatbot.
Seikoh: Wow. So the Global South provides the raw data and the cheap manual labor to clean the system, but the final product is engineered to reflect the values of the Global North.
Selim: Precisely. Data Colonialism means that intellectual wealth is extracted from the globe, processed through a Western filter, and then sold back to universities as "intelligence." When Chris uses that tool without questioning it, he is treating diverse, global narratives as invisible.
Seikoh: That is an incredibly sobering thought. It means that unmitigated use of generative AI in a diverse postsecondary classroom threatens to homogenize all student thought into a singular, Western-centric worldview. If every student uses the same bot to study, every student starts to think exactly the same way.
Selim: And that is the ultimate danger of the cognitive offloading you were talking about in your segment, Seikoh. It's not just that they aren't learning; it's that they are actively unlearning their own diverse cultural perspectives.
Part 5: Actionable Strategies & The Outro
​
Seikoh: Okay, Selim. We’ve taken our listeners on a pretty intense journey today. We’ve established that relying on AI degrades the self-regulated learning process, which completely breaks the validity of university exams. And on your end, we’ve established that the tool itself is a Trojan horse that uses a confident, authoritative tone to sneak Western-centric biases into the minds of tired students.
Selim: It’s a heavy realization. And I think the worst thing we could do right now is just leave everyone with a pure diagnostic critique. If a UTM student is listening to this while walking to the library, they need to know what to actually do about it.
Seikoh: Exactly. Because banning the technology isn't an option. It’s here to stay. But the current rules universities are using don't seem to be fixing the problem either.
Selim: They aren't. And that’s because most university AI policies are built around rudimentary academic integrity. If you look at a standard syllabus, the rule is essentially: "Do not use ChatGPT to write your essay, or you will get a zero." But as we’ve discussed for the last half hour, the real danger isn't just submitting a fake essay.
Seikoh: Right, the danger is the cognitive offloading and the automation bias happening during the process of studying, long before the essay is ever written. We are policing the output, but ignoring the input. So, what is the educational defense against this?
Selim: To build a real defense, we have to completely change how we teach students to view the machine. I synthesized two incredibly important frameworks for this. The first is from a 2024 paper by Burriss and Leander, who advocate for teaching "Critical Posthumanist Literacy." And the second is from Wang and Wang (2025), who argue for a "relational approach" to AI literacy.
Seikoh: "Critical Posthumanist Literacy." I love the sound of it, but it sounds like a concept from a graduate-level philosophy seminar. How do we explain that to Chris, our hypothetical first-year student?
Selim: It sounds intimidating, but the core idea is actually very intuitive. Traditionally, "humanism" puts the human at the absolute center of knowledge creation. Tools are just dead objects we use. A calculator doesn't have an opinion on math; it just does the math. But "posthumanism" recognizes that Generative AI is not a dead tool.
Seikoh: Because it has what researchers call "ontological agency"?
Selim: Exactly. The machine is an active participant in the room. It makes decisions about what words to use, what history to highlight, and what culture to prioritize. Burriss and Leander, along with Wang and Wang, argue that educators must train students to recognize this agency. Chris needs to stop treating ChatGPT like a magical, neutral encyclopedia, and start treating it like a highly opinionated co-author who has a very specific worldview.
Seikoh: I really like that framing. Treating it like an opinionated co-author fundamentally changes the dynamic. It breaks that "automation bias" you mentioned earlier. If I know my co-author is incredibly biased toward a Western, English-speaking perspective, I’m not just going to blindly accept their summary of the readings. I’m going to interrogate it.
Selim: Exactly! You are going to argue with it. And Seikoh, think about what that does for your pedagogical crisis. If Chris is arguing with the AI, is he still passively offloading his cognition?
Seikoh: No! He’s actively engaging with the material again. He’s stepping back into the struggle of learning. He’s monitoring, reflecting, and evaluating, which brings back the exact self-regulated learning processes that Panadero and Broadbent warned we were losing. That is a brilliant synthesis.
Selim: Right! The struggle is brought back into the studying process. But to make this incredibly practical, we have to give students a specific mechanism to do this at 2:00 AM. And this goes back to the empirical research by Tao and colleagues (2024). They present a highly tangible, user-end solution called "Cultural Prompting."
Seikoh: Cultural Prompting. Walk me through how Chris actually types that into his computer.
Selim: Since we know from the data science that the AI defaults to a Protestant European and English-speaking perspective, students can use their prompt to intentionally hijack that default architecture. Let’s say Chris asks the AI to explain the economic impacts of the global supply chain. The default answer will likely frame it around Western corporate efficiency and consumer benefits.
Seikoh: Because that’s what the C4 dataset is full of.
Selim: Exactly. But using "Cultural Prompting," Chris explicitly commands the LLM to change its lens. He types: "Explain the economic impacts of the global supply chain specifically from the perspective of manufacturing workers in Southeast Asia," or "Analyze this historical event using an Indigenous epistemological framework."
Seikoh: So he is forcing the AI out of its algorithmic comfort zone.
Selim: He is. Tao et al. found that explicitly directing the AI to adopt a specific, non-Western regional perspective actually works. It forces the model to bypass its default cultural alignment and pull from marginalized data that would have otherwise been invisible.
Seikoh: That is an incredibly powerful tool. But even beyond getting a more diverse output on the screen, I have to imagine the real value is what that prompt does to Chris's own brain.
Selim: 100 percent. It introduces what we call "critical friction." By forcing himself to stop and ask, "Wait, what cultural perspective do I actually want this information from?", Chris is no longer a passive consumer. He is stepping back into the driver’s seat of his own education. He is reclaiming his independent critical thought.
Seikoh: He is using the tool, but he is refusing to let the tool use him.
Selim: Precisely. And that is the ultimate takeaway of my research. Generative AI is structurally biased. If we don’t actively teach students the mechanics of that bias, if we don't teach them posthumanist literacy and cultural prompting, we are allowing algorithms to quietly overwrite human diversity. We have to empower students to introduce that critical friction so they can safely navigate these systems.
Seikoh: And that brings us full circle to the main question of this entire podcast: When students use AI to study, what does the exam result actually measure? If universities redesign their assessments to measure true human judgment and if students adopt cultural prompting to break the AI’s ideological trap, then that "A" on the exam actually means something again. It means the student engaged critically with the modern world, rather than just memorizing a machine's reflection of it.
Selim: That is exactly right. The machine might be able to do the reading for you, but you can never, ever let it do the thinking for you.
Seikoh: I think that is the perfect place to leave it. We hope this conversation gives everyone, students and professors alike, a new way to think about what is actually happening when we open that chat window. It’s not just about efficiency; it’s about protecting the diversity of human thought.
Selim: Thank you to everyone listening to this deep dive into the AI learning paradox. We encourage you to try out cultural prompting the next time you study, and see how it changes your perspective.
Seikoh: I’m Seikoh Murakami.
Selim: And I’m Selim Liu. Stay critical, stay curious, and we’ll see you in the next episode.
[SCENE END: Lo-fi music swells up, holds for five seconds, and fades out to silence.]