Marcus: You open your inbox and, um, there it is, an invitation to, quote, learn anything, one chat at a time.
Devon: Oh, right, that one.
Marcus: Yeah, it is a promotional email from OpenAI. And it is suggesting you turn on a specific feature to get step-by-step guidance rather than just quick answers.
Devon: And they even suggest a prompt for you to use straight away, don't they?
Marcus: Exactly. It tells you to type, act as my microbiology tutor. So, OpenAI is no longer just selling you a search engine or a text drafter. They are selling you a private, endlessly patient educator.
Devon: Yes.
Marcus: But there is a massive catch embedded in that pitch.
Devon: Well, I mean, the horrifying yet mildly amusing edge to that exact framing is what sits underneath the hood.
Marcus: Right.
Devon: Because your new private microbiology tutor is a piece of software that, by the company's own explicit admission in their documentation, is still prone to inconsistent behaviour and factual mistakes.
Marcus: Yeah, which is a fairly big caveat for a tutor.
Devon: It is. You are basically handing your educational development over to a system that might occasionally, you know, hallucinate the structure of a cell, but it will do so with absolute, unshakable confidence, wearing the digital tweed jacket of an Oxford professor.
Marcus: Which sets the stakes perfectly for us.
Devon: Mhm.
Marcus: So, for this deep dive, we are looking at OpenAI's internal documents, their promotional emails, and their major deployment announcements from July 2025 onwards.
Devon: Right.
Marcus: The mission here is to explore how whole national education systems are being quietly, yet fundamentally, rewired around this technology.
Devon: And the crucial detail, the thing we really have to keep in mind here, is that this massive structural rewiring is happening while the concrete evidence that it actually improves learning is still being gathered.
Marcus: Exactly. The proof simply does not exist yet.
Devon: Yeah.
Marcus: So, to understand how we arrived at this massive global experiment, we first have to look at the specific mechanism OpenAI introduced, which they called study mode.
Devon: Because the shift from an answer machine to an educator requires a completely different software architecture, doesn't it?
Marcus: It really does.
Devon: If you do not understand the underlying mechanism of how study mode functions, the sheer scale of the global rollout that follows makes absolutely no sense.
Marcus: Spot on. So, think of a standard generative text model. If you ask it a question, its default imperative is to give you the answer as quickly and comprehensively as possible.
Devon: Which is, frankly, terrible pedagogy.
Marcus: Right. If a student asks how to solve a quadratic equation, handing them the completed formula teaches them absolutely nothing.
Devon: No, it just ends the interaction.
Marcus: So, study mode intervenes. I like to think of it as, um, bumper bowling for your brain.
Devon: Bumper bowling.
Marcus: Yeah, instead of just handing over the completed essay or the solved mathematics equation, the system keeps your thinking in the lane, but it forces you to actually throw the ball yourself.
Devon: Right, I see.
Marcus: It relies on custom system instructions, which were built in collaboration with pedagogy experts, to fundamentally restrict the model's output.
Devon: Restricting the output is the key phrase there, I think. The software is essentially being forced to work against its natural inclination to generate a complete answer, and instead use specific educational techniques.
Marcus: The primary technique being Socratic questioning.
Devon: Yes.
Marcus: So, if you ask the study mode tutor how to calculate the velocity of an object, it will not give you the formula. It answers your question with a guided question designed to spark your critical thinking.
Devon: It might ask something like, um, what variables do we already know from the text?
Marcus: Exactly, alongside using what they call scaffolded responses.
Devon: All right, the scaffolding.
Marcus: So, instead of giving you the whole ladder, scaffolded responses mean it gives you the first tiny rung of the physics problem, waits for you to solve it, and then builds the next step exclusively on top of your answer.
Devon: And it uses targeted knowledge checks to ensure you actually understand the underlying concept before allowing you to move forward.
Marcus: Which sounds excellent in theory, doesn't it?
Devon: Oh, it mimics the idealised version of a one-on-one human tutor perfectly. You know, the kind of bespoke education that historically only the wealthiest in society could ever afford.
Marcus: Right.
Devon: But we have to look at how this mechanism actually manifests in the real world when institutions get their hands on it. If we look at higher education, we can see early examples of this architectural shift.
Marcus: Yes. When ChatGPT Edu was released in May 2024, universities immediately began applying this exact tutoring logic.
Devon: And Arizona State University is a perfect case study for this dynamic.
Marcus: It really is. They built a custom tool they called a Language Buddies GPT. And the idea was to provide students with level-appropriate German conversation practice.
Devon: Which is a brilliant use case for language learning.
Marcus: Absolutely. The system is completely tailored to the individual's proficiency, and it delivers real-time, interactive feedback on their grammar and vocabulary choices.
Devon: So, a student can just sit in their dorm room and have a sprawling two-hour conversation in German.
Marcus: Right, with a system that corrects their verb conjugations without ever getting frustrated or tired.
Devon: But, um, the friction there is found in the stated motivation behind that specific Arizona State deployment.
Marcus: Yes.
Devon: Because when you read the documentation, it explicitly notes that this tool was built to save faculty time on assessments.
Marcus: Which is a very different goal than just helping students learn German.
Devon: Exactly. This surfaces a tension we have to keep central to this entire deep dive. Is this truly a revolution in pedagogy designed to create a superior learning environment for the student?
Marcus: Or is it primarily an efficiency play?
Devon: Right. An efficiency play disguised as pedagogy, designed to reduce the administrative burden and ultimately the financial cost of running a university.
Marcus: And that question becomes completely unavoidable when you look at the sheer scale of what happened next.
Devon: Because this mechanism promises such massive, quantifiable, time-saving efficiencies.
Marcus: Yes. It is not just being marketed to individual students studying German in Arizona. It is being embedded into the foundational architecture of entire nations.
Devon: Which brings us to the common story you hear in the media, right? The narrative that AI tutors will inevitably and ruthlessly replace human teachers.
Marcus: Yes, but the reality in the sources points the exact opposite way. OpenAI is leaning incredibly hard into a teacher-first enablement strategy.
Devon: They really are. They position the software as a co-pilot for the educator, and they aggressively push back against the narrative that they are building a teacher replacement engine.
Marcus: To see that in action, look at the Education for Countries initiative. This was launched at the Davos summit in early 2026.
Devon: Right.
Marcus: And the usage figures they are reporting are simply staggering. Globally, they cite 900 million weekly ChatGPT users.
Devon: 900 million. That is just, it is hard to even wrap your head around.
Marcus: It is. And by May 2026, Singapore had formally joined the program at the Education World Forum in London. In Singapore alone, nearly 43% of all usage by 18 to 24-year-olds is tied directly to learning.
Devon: See, we are looking at a fundamental demographic shift in how a generation searches for truth.
Marcus: How so?
Devon: Well, they are no longer querying a search engine for a list of links, are they? They are interrogating a conversational interface for a synthesised worldview.
Marcus: Yes, that is a massive shift. And the structural integration goes much deeper than just individual students querying an app on their phones.
Devon: Right, look at Jordan.
Marcus: Exactly. In Jordan, over 1 million students and more than 100,000 teachers are engaging with an AI assistant called Siraj.
Devon: Siraj is not just a chatbot on a website, though.
Marcus: No, it is an interface being woven into the daily classroom workflow. And then in Kazakhstan, 84,000 educators completed what they called AI-readiness training.
Devon: 84,000 educators.
Marcus: Yes, yes. And in their first month alone, those active educators sent 1.5 million prompts.
Devon: Oh, wow.
Marcus: Then you have Estonia's AI Leap program. Estonia is a historically tech-forward nation, and their program explicitly states its aim is to strengthen independent learning alongside reinforcing the role of teachers.
Devon: So, they are trying to build an ecosystem where the AI does the heavy lifting of independent study, but it is entirely guided by a human curriculum. Well, I mean, the scale is undeniable. But if we connect this to the bigger picture, we really have to cut through those massive, impressive figures.
Marcus: What do you mean?
Devon: We need to look closely at OpenAI's own remarkable honesty regarding what these numbers actually mean.
Marcus: Ah, right.
Devon: OpenAI frames these statistics purely as signals of positive impact. They very carefully do not call them proven learning outcomes.
Marcus: That is true. They highlight that in Slovakia, nine out of ten educators reported saving around five hours a week.
Devon: Right, but that is a measurement of administrative relief. It is a measurement of time saved on a Sunday afternoon.
Marcus: Yes.
Devon: It is absolutely not a measurement of whether the students in those Slovakian classrooms are retaining more information, or developing deeper critical thinking skills, or becoming better citizens.
Marcus: Which OpenAI admits. They candidly call these deployments a first step.
Devon: Exactly. They openly warn of the inconsistent behaviour and the factual hallucinations.
Marcus: And most importantly, they explicitly admit the research landscape regarding AI and education is, quote, still taking shape.
Devon: Which is why they point to the longer-term studies that were initiated, like the ones with NextGenAI and Stanford University's SCALE initiative.
Marcus: Yes, they are funding the research to find out if the product actually works as an educator.
Devon: Because educational effectiveness is an open research question. It is not a settled scientific fact.
Marcus: Right.
Devon: Yet, ministries of education in Jordan, Kazakhstan, and Singapore are rolling it out to millions of students while waiting for Stanford to tell them if the fundamental premise actually improves human learning. But waiting for those Stanford studies to finish creates a massive blind spot right now. If the only proven benefit at this stage in these initial deployments is that it saves a Slovakian teacher five hours a week—
Marcus: Yeah.
Devon: —well, we are restructuring global education around an administrative efficiency tool, not a learning tool.
Marcus: Yes. And that forces us to look at what happens when AI permanently takes over that routine work, which brings us to the genuine structural disagreement at the heart of this deployment.
Devon: The narrative of teacher-first enablement is colliding violently with the reality of how human expertise is actually built over time.
Marcus: Well, I am going to defend the mechanism here, because I believe the teacher-first framing is highly sincere and deeply necessary right now.
Devon: Okay, go on.
Marcus: If you speak to any educator across any of these national systems, the burnout rate is catastrophic.
Devon: Sure, nobody is denying the burnout crisis.
Marcus: Right. But imagine grading 80 introductory essays on the causes of the First World War on a Sunday evening.
Devon: It sounds utterly miserable.
Marcus: It is. You are exhausted. There are coffee stains on the paper. You are reading an essay from a student who seems to genuinely believe Archduke Franz Ferdinand was just, you know, an indie rock band.
Devon: Right.
Marcus: It is soul-crushing exhaustion. It is the reason teachers leave the profession in droves.
Devon: Nobody is arguing that grading 80 identical, poorly structured essays is a joyful, spiritually fulfilling experience for a human being.
Marcus: Exactly. So, freeing educators from the crushing bureaucracy of routine marking, from generating basic lesson plans, from answering the exact same foundational questions from 30 different students, that is a massive, tangible gain.
Devon: If it works.
Marcus: If an AI tutor can handle the rote assessment or guide a student through the basic grammatical errors in their German language practice, it frees the human teacher to do what software absolutely cannot replicate.
Devon: Assuming the administration allows them that free time, of course, rather than just increasing their class sizes to save money.
Marcus: Let us assume it works as intended. It allows for high-level mentoring. It allows the teacher to lift their head up from the grading rubric and actually notice the quiet student in the back of the room who is struggling with confidence.
Devon: Okay, yes, the pastoral care side of things.
Marcus: Yes. And it allows for complex, nuanced debate about the ethics of historical events rather than just marking dates on a timeline with a red pen. The AI does the heavy lifting of basic tutoring, and the human becomes the master conductor of the classroom. You cannot tell that Slovakian teacher to reject the tool that gives them their Sunday evening back.
Devon: You cannot. I agree. That is the utopian view of time-saving, and for that individual teacher in the present moment, the relief is profound. But who actually pays for this? What is the hidden structural cost of automating that, quote, soul-crushing work?
Marcus: What do you mean by structural cost?
Devon: We have to look at the human stakes over a longer timeline, and we have to look at the pipeline of expertise. You are entirely correct that an AI delivering 24/7 office hours, handling entry-level tutoring, language practice, and routine marking sounds like administrative heaven.
Marcus: It does.
Devon: But by automating the specific tasks, you quietly eradicate the junior teaching roles. You eradicate the graduate teaching assistants. You eradicate the entry-level tutors.
Marcus: Well, because the university or the school board no longer needs to pay them to do the marking. The financial incentive to hire them just evaporates.
Devon: Exactly. And this raises a highly pressing question for you to consider. How does the next generation of expert teachers ever get trained if we hollow out the very apprenticeship path through which humans learn to instruct?
Marcus: Wait, you are suggesting that marking those basic, terrible essays about the First World War is actually a necessary part of becoming a master teacher?
Devon: I am suggesting it is the absolute foundational requirement.
Marcus: Really?
Devon: Yes. Reading 100 poorly structured essays is exactly how a junior educator learns to map the cognitive failures of a student. It is how you learn to spot the precise moment a student's logic jumps the tracks.
Marcus: Okay, I see your point.
Devon: When you sit with a student as a junior tutor and painfully work through their misunderstandings in a one-on-one session, you are building your own internal database of human learning patterns. You learn how frustration looks on a student's face. You learn when to push harder and when to back off.
Marcus: So, if the Language Buddies GPT handles all the basic conversational practice and grammar correction at Arizona State—
Devon: Yes.
Marcus: —the human teaching assistants never get the experience of diagnosing those foundational errors. They never see the messy, early stages of someone trying to acquire a language.
Devon: Exactly. You cannot just spawn a master teacher out of thin air. They require years of grinding through the basics. If we hand the basics over to a Socratic study mode because it is cheaper and more efficient, we are pulling up the ladder behind the current generation of master teachers.
Marcus: That is a terrifying thought, actually.
Devon: The master conductor you described, the one leading the nuanced debate on historical ethics?
Marcus: Yeah.
Devon: They only possess that skill because they spent ten years grading the terrible essays. We are solving a short-term burnout crisis by potentially engineering a long-term structural collapse of the teaching profession.
Marcus: That is a profound tension. And, um, it is entirely unresolved in any of the documents we have reviewed for this deep dive.
Devon: It is just not addressed at all.
Marcus: On one hand, you have immediate, necessary relief for an overburdened workforce. On the other, the same tool providing that relief is positioned to systematically dismantle the entry-level academic labour market.
Devon: And honestly, whether you are a university student currently being asked to converse with a Language Buddy GPT, or a professional watching automation slowly enter the junior levels of your own corporate field, the dynamic is exactly the same.
Marcus: It applies everywhere.
Devon: It does. The tool is incredibly powerful at simulating instruction. It can deploy a Socratic method and scaffold a response beautifully. But the long-term impact on the ecosystem of human expertise is entirely unmet.
Marcus: Which brings us back to the reality of the rollout. We are looking at a technology that OpenAI intentionally pivoted. They took a model designed to generate text and forced it, via custom instructions, to withhold answers and guide students instead.
Devon: Yes.
Marcus: And because that study mode is highly efficient, it is not just staying on individual laptops. It is being deployed by ministries of education from Jordan to Kazakhstan, from Estonia to Singapore. We are seeing millions of students interacting with AI as a primary educational interface.
Devon: Millions.
Marcus: But the absolute centre of this deep dive is that OpenAI openly admits the evidence for improved learning outcomes does not yet exist. They are operating on signals of positive impact, primarily measured in time saved by educators.
Devon: They built a system to mimic the apprenticeship of learning, and they sold it to the world before proving it actually creates better learners.
Marcus: Which leaves you with a highly specific, anti-hype directive. Do not just watch the user numbers grow or marvel at the scale of the national deployments. Watch closely to see if OpenAI, or the researchers at Stanford's SCALE initiative, ever actually publish the concrete, peer-reviewed learning outcomes evidence that they admit they currently lack.
Devon: Absolutely.
Marcus: And perhaps more importantly, watch what happens to the entry-level teaching and tutoring work as these global deployments scale. Will the time saved simply lead to a richer educational experience for the students of today? Or will we find we have accidentally automated away the very path required to build our future experts?
Devon: Which leaves you with one final thought to mull over as you watch this integration unfold. Let us say the study mode works perfectly. Let us say we successfully build a flawless, endlessly patient tutor that removes all the friction, all the embarrassment, and all the frustration from the process of learning. If we remove all that friction, do we accidentally kill the resilience required to handle the real world, where answers are messy, people are impatient, and there is no study mode to keep your thinking safely in the lane?