An argument for why the answer to AI in university education is neither to ban it nor to surrender to it, but to change both what universities teach and how students use it, so that understanding remains the student’s, and why I think Sub‑Saharan Africa can reach that state before the West.
A note on where I’m writing from: I am building Orient, an offline-first AI study companion for university students and lecturers in Sierra Leone. It helps a student work through their own course materials so the understanding is theirs, and it runs on the phone without an AI subscription or internet connection, so I am not a neutral observer here. This essay is the argument underneath it. I am putting Orient forward as one attempt at a third response rather than as the answer, and I have tried to make the case stand on its own, so that it holds whether or not the particular tool does.
01 · The article that prompted this
There is a particular kind of essay that has become a genre over the last two years, and it always opens the same way; A freshman uses ChatGPT for the first time in the spring of her senior year of high school. By the time she reaches college she is running every assignment through it, watching her grades climb even as she senses that something is off, and yet, she does not stop.
The most-read version of that essay is James D. Walsh’s “Everyone Is Cheating Their Way Through College,” published in New York’s Intelligencer in May 2025, where Sarah at Wilfrid Laurier was the freshman.[3] The piece quoted students at Columbia, Berkeley, NYU, and a dozen other schools, all saying versions of the same thing: ChatGPT writes most of their essays, takes their notes, builds their study guides, debugs their code, and drafts their lab reports... One student in Utah even described college as little more than “just how well I can use ChatGPT at this point.”
It also quoted teaching assistants (TA) and professors too. A TA named Williams, grading papers in a writing class, reckoned that at least half his students were using AI by November, and he had no faith in the detectors meant to catch them. When he raised it, the professor running the course told him not to fail individual papers, even the obviously AI-smoothed ones, since the department could not really prove anything. Williams ended up grading on what he wryly called “their ability to use ChatGPT.”
An early survey, run in January 2023 only two months after ChatGPT launched, already had nearly 90% of college students using it for homework.[3] That was a small, self-selected panel, so hold the number loosely, but the structural evidence points the same way and is harder to argue with: ChatGPT’s traffic then dips when schools break for summer and rises when term resumes. By 2025, use is functionally universal. A philosophy professor at the University of Arkansas at Little Rock even caught students using AI to answer the opening prompt that asked them simply to introduce themselves and say what they hoped to get out of the class. And a computer-science major at Columbia openly bragged that AI had written 80% of his essays.
Brian Patrick Green, a tech-ethics scholar at Santa Clara, stopped assigning essays after ChatGPT was launched. Other professors also fell back on in-class blue books, oral exams, and prompts engineered so that then current models would produce nonsense. The students complained, and one even asked for an extension because ChatGPT happened to be down on the day the assignment was due.
It is not only that students use the tool. It is that the tool can match them. A study out of NYU Abu Dhabi, published in Scientific Reports, fed ChatGPT 233 real assessment questions from 32 professors across 8 disciplines and had blind graders score its answers against the students ’. The model earned equal or higher marks in 9 of the 32 courses, close to a third of them.[8] Its one consistent weakness was that it did best on fact-based questions and fell well behind the students whenever a prompt turned conceptual.
If you have been near a university in the past three years, none of this is new. If you have not, this is worth exploring. The institutional response has mostly been some combination of three things: detection that fails or proves unreliable, restriction that is locally effective but globally pointless, and a quiet adjustment of standards that is the path of least resistance and so the one most institutions go by. Meanwhile, Sarah graduates and Williams quits teaching. Wendy at NYU writes that her ideal of college as a place of intellectual growth was gone long before ChatGPT. And the only winners, as Gary Marcus put it, are “the Sam Altmans of the world,” who “laugh their way to the bank.” [6]
The situation is bleak in a specific way. The bleakness is not that AI is too powerful. It is that institutions do not have a coherent answer. In the absence of one they default to whichever response is cheapest, which most of the time is quiet capitulation and occasionally expensive theatre: the detectors that do not work, the blue-book exams that do not scale, and the departmental policies that reverse themselves every six months. The result of that is a system that is neither educating its students nor convincingly pretending to.
02 · Some context on how we got here
It is worth being honest that the cheating problem did not begin with ChatGPT. The genre of the Intelligencer essay is new, but the conditions underneath it are not, and they have been building for a long time.
Higher education across much of the world has been drifting away from the idea of a place of intellectual growth for decades, and the drift goes by a different name in each place it happens.
In the United States, it is credentialism. The degree has become a labour‑market signal that is increasingly unmoored from anything that was actually learned: a credential that employers use to sort applicants by trainability, status, and access to professional networks as much as by demonstrable skill.
In the United Kingdom, it is the marketisation of the university. Students are recast as customers buying a product, while faculty workloads are pushed to the point that academics themselves describe them as unmanageable, amid understaffing, audit, and administrative overload.
In Sub‑Saharan Africa, it is the brutal arithmetic of underfunded institutions trying to serve growing populations: a single lecturer might face 150 students or more in a system where the average student‑to‑lecturer ratio is roughly twice the international benchmark, and where national ratios as high as 154:1 have been reported. Lecture notes circulate on WhatsApp because formal learning management systems are inaccessible, unreliable, or simply absent, and because data costs and connectivity push both staff and students toward the platform they already have. And the gap between what is taught, how it is examined, and what graduates are expected to do in the real economy is wider than anyone wants to admit, with curricula still often described as fragmented and poorly aligned with local needs.
EdTech, broadly construed, has been promising to close that gap for twenty years, with mixed results. The Massive Open Online Course (MOOC) wave promised free university lectures for everyone, and it did reach enormous audiences, but completion rates stayed in the low single digits, around 3% of enrollees by 2018, and had not improved over 6 years, while the major providers gradually pivoted toward paid professional and credential programmes.[14] The adaptive-learning wave that followed promised AI-personalised tutoring and produced some real successes, Khan Academy’s work in K-12 mathematics among them, without delivering the broad transformation it had promised. The most recent wave, driven by generative AI, was meant to be the breakthrough. In a narrow sense it has been, since students are using it constantly, but it is not the breakthrough anyone hoped for. The clearest verdict on that wave came from its flagship. Khan Academy built Khanmigo on GPT-4 as the best-resourced AI tutor in the field, and 3 years on Sal Khan conceded that for most students it had been “a non-event”, because the tutor sat waiting to be asked and most students never asked.[15] The lesson is not that the tutor was weak. It is that a tool which waits on the student’s volition inherits the student’s volition, which for most students under pressure is exactly the problem.
Now, on cheating…
ChatGPT didn’t break a working system. It revealed that the system had been broken for a long time and was held together by the friction of essay-writing taking twelve hours instead of two.
The cheating problem is not really an AI problem. It is an alignment problem between what students are asked to do, what universities are equipped to assess, and what students actually need to learn for the world they are graduating into. Once you remove the friction, the underlying misalignment becomes visible. You will see that the students did not suddenly become cheaters so much as the system suddenly lost its ability to pretend.
This was not a fringe reading. When Scientific American canvassed education researchers in 2023, they said much around the same thing in their own words. Ian O’Byrne, who studies literacy and technology, argued that the tools were less a new crisis than a mirror held up to what was already happening in classrooms. The educational psychologist, Joe Magliano pointed to an incentive structure that rewards students for grades rather than for understanding. And Kui Xie, who studies student motivation, put it most directly: a student set on mastering a skill has no reason to cheat, so the real work is to address why students reach for the shortcut rather than to chase the shortcut itself.[9]
In Sub-Saharan Africa, the misalignment is sharper still, because the friction was already lower to begin with. A 2025 systematic review of generative AI adoption in African higher education found that institutional “preparedness for GenAI integration remains inconsistent” across the continent.[1] A separate 2025 governance review found that only a handful of universities, mostly in South Africa, Rwanda, and Nigeria, have begun aligning institutional policy with national digital strategies, while “most institutions remain in aspirational phases, limited by infrastructure and human capacity.”[2] Sierra Leonean, Liberian, and Gambian universities are functionally in that aspirational phase. There is no coherent institutional response to ChatGPT, no detection infrastructure, no AI-proofed exams, and in most departments, no policy at all.
That gap reads as the bleakest fact in this inventory, and the second half of this essay argues it might be the most promising one...
03 · What 81,000 people actually said
The argument I am about to make for Sierra Leone, and for Sub-Saharan Africa more broadly, used to rest on observation and rhetoric. As of March 2026 it rests on data, because Anthropic published a study that quietly reshaped how this conversation should be framed.[7]
In December 2025 Anthropic ran the largest qualitative study ever conducted on AI use, interviewing 80,508 Claude users across 159 countries in 70 languages. The instrument was Claude itself, configured as a conversational interviewer. It asked people what they wanted from AI, whether AI had delivered, and what they were afraid might go wrong. Several of its findings bear directly on Orient’s premise, and the rest of this section walks through what they do and do not establish.
The first is that AI sentiment is markedly more positive in lower and middle-income countries than in wealthy ones. Sub-Saharan Africa, Central Asia, and South Asia showed the highest positivity and the lowest expressed concern about AI in general. Roughly twice as many respondents in those regions said they had no concerns about AI as in North America, Oceania, or Western Europe. The strongest predictor of negative sentiment everywhere was worry about jobs and the economy, which is precisely the concern that runs lower in the regions where Orient is positioned.
[ FIG.02 · Where the optimism actually lives ]
The second is that the dominant vision for AI in Sub-Saharan Africa was entrepreneurship, framed by respondents as a way of bypassing capital they could not otherwise raise. The Ugandan entrepreneur quoted in the study put it plainly:
“Coming from Africa, not based in the US or in the UK, getting funding is very difficult. And the only way I probably have to stake a claim in the market... is building a technology that works.”
Ugandan entrepreneur · 81K study
In Central and South Asia, learning ranked disproportionately high, at 14% and 13%, against 8% globally. Respondents pointed to teacher shortages, knowledge gatekeeping, and the cost barriers of traditional education as the constraints AI might break.
The third finding is the one most relevant to the cheating argument, and it names something the original Intelligencer piece only gestured at. The firsthand experience of cognitive atrophy from AI use is concentrated among educators, at 24%, and academics, at 19%. It is least common among tradespeople, at 4%, and self-employed researchers.
That finding does not stand on its own. A longer line of research on cognitive offloading has held that handing mental work to a machine can lift performance on the task in front of you while carrying real long-term costs, among them weaker problem-solving, poorer recall, and a dulled capacity to learn something new.[13] Policy researchers have raised the sharper version of the same worry for younger learners, whose minds are still forming.[12] The 81K numbers gave measurable shape to a concern the literature had been circling for years.
[ FIG.03 · Cognitive atrophy, by occupation ]
The authors’ interpretation was direct: “AI’s benefits may be strongest when learning is volitional, compared to within institutional structures where AI is more likely to be used as a shortcut.” That single sentence does a great deal of work, because it says in measurable terms that the cheating dynamic is not a property of AI but of the institutional context AI is dropped into. A student forced through coursework they did not choose or they consider boring will use AI as a shortcut. A tradesperson learning a skill voluntarily will use the very same tool as a multiplier. The difference between the two outcomes is nothing more than the user’s relationship to the work.
[ FIG.04 · Same tool, opposite results ]
This matters for the reorientation argument specifically, because reorientation is the one response that changes that relationship: when the tool helps a student see what is worth studying and decide how to go about it, and when it becomes something they bring to their own learning rather than something handed to them, the relationship slides toward the volitional end of the spectrum. The 81K data suggests that slide is the variable that decides whether AI hollows a student out or makes them more capable.
There is a fourth finding the study foregrounds: the duality almost every user carries inside themselves. People who lean on AI for emotional support are three times more likely to also fear becoming dependent on it. Students who feel the learning benefits are the same students who worry about cognitive atrophy. The benefits and the harms are not separate populations but tensions inside individual people, which Anthropic’s researchers called “light and shade.” The upshot is that students are not naive about the trap. They are aware of it and using AI anyway, because no one has shown them an alternative that keeps the benefit without the cost. As one respondent in South Korea put it:
“I got excellent grades using AI’s answers, not what I’d actually learned. I just memorized what AI gave me. That’s when I feel the most self-reproach.”
South Korea · 81K study
A student named Cal Baker, writing in Scientific American about his own coursework, made the same point from the inside. He argued that, what mattered in an assignment is the thinking a student does on the way to finishing it, not the finished thing itself, and when the tool does that thinking the student never builds the capacity the work was meant to build.[10] It is the cognitive-offloading point again, and this time in the voice of someone it is being done to rather than someone studying it from outside.
The self-awareness is there; the alternative is not. That is the empirical foundation underneath the third response. People in Orient’s markets are more positive about AI than the West. They want it as a ladder for entrepreneurship and learning in particular. Some are aware of the cognitive cost when it is used as a shortcut, and that awareness is concentrated in the educators who are watching it happen. The institutional context decides whether AI helps or harms, and reorientation is the response that addresses that context directly. The data does not make the argument by itself, but it makes the argument much harder to dismiss.
04 · The three responses
What’s actually happening once you strip the rhetoric away, is that institutions are choosing from a menu of three responses, and they rarely name the one they pick. They simply default to it.
[ FIG.05 · The Menu Institutions Choose From ]
01 · Capitulation
This is the Williams pattern, and it is rarely chosen on purpose. Try to picture the mechanism in a single room: A lecturer sits with a stack of 40 essays and can tell that perhaps half were written with ChatGPT or Claude, but the AI-smoothed ones are not obviously worse than the honest ones, and there is no way to prove which is which, so all 40 are marked on what is on the page. The essay that would have scraped a bare pass 3 years ago now sits comfortably in the middle of the stack, and over a few semesters the middle is where the standard quietly resettles. Nobody announced a new policy. The department simply lacked the will or the resources to hold the old line, and what counted as acceptable work drifted down to meet what was being handed in. The degree still says the same thing on paper, but the skill it certifies has changed underneath it.
Capitulation happens through accumulation rather than decision. A department declines to fail clearly AI-written papers because the legal risk of a false accusation is too high. A professor stops assigning essays because grading them has become impossible. A faculty senate passes a resolution about responsible AI use that is vague enough to mean nothing. None of those moves looks like surrender in isolation, even though the aggregate plainly is. The graduate who emerges from a capitulating institution holds the certificate without the literacy, able to produce work with AI but not without it, unable to tell when the tool is wrong or to push past what it handed them. The gap does not show up in the first job so much as two or three jobs in, when the assignments stop being neatly bounded.
02 · Policing
This is the detection-and-prevention pattern: here, we are talking about AI detectors and the plagiarism scanners trained on AI signatures, the trojan-horse prompts hidden in invisible text inside assignments to confuse ChatGPT, the in-class handwritten exams brought back as a defensive measure and the blanket bans backed by honour codes. It is what institutions try when they will not capitulate but cannot reorient, and it mostly fails. Detection is a losing arms race in which every advance trains the next round of obfuscation. That is why Williams had no faith in the detectors, why the students at NYU complained that AI-proofed assignments were interfering with their learning, and why a student at Berkeley asked, not unreasonably, why would he walk from point A to point B when there was a car sitting right there? A computer-science professor in Berlin spent a semester feeding her own exam questions to ChatGPT and rewriting them to trip it up, while openly granting that she might not manage it and would be learning alongside her students, which is the arms race rendered in miniature.[9]
Policing carries a quieter cost as well, because it poisons the relationship between the teacher and student. Every assignment becomes a potential accusation, every essay is a piece of evidence to be weighed for AI signatures. And the tools doing the weighing are not neutral. Researchers who ran essays through the most widely used detectors found that more than half of the essays written by non-native English speakers were wrongly flagged as AI-generated, while essays by native speakers were cleared.[4] The burden of suspicion falls hardest on exactly the students who write English as an additional language. The same instinct shows up with no software at all. When the investor Paul Graham announced that a cold email had given itself away by using the word “delve,” he called it a sign the text was written by ChatGPT. Nigerian writers pushed back to point out that “delve” is ordinary, everyday English where they are from, and that treating it as a machine’s tell quietly punishes the international students whose vocabulary simply does not match an American ear.[5] And bias is only half of it, since every detector also produces plain false positives, and a student wrongly flagged is left carrying the stress of an accusation that is very hard to disprove.[11] In a country like Sierra Leone, where students think and write in English alongside Krio, that bias is not an edge case, it is the default condition. A policing regime built on it would land first and hardest on the very people the system is meant to serve.
In places that lack any detection infrastructure, policing usually takes a blunter form; the outright ban. ChatGPT, Claude and other AI tools are officially prohibited and students are told not to use them, which produces no enforcement, no detection, and no actual change in behaviour, only hypocrisy. The lecturers know the students are using AI, the students know the lecturers know, and the ban becomes a polite fiction everyone observes for the length of the lecture and ignores after class. It is the same dynamic as the “no-phones” rule from a decade ago, except this time the phone is doing the homework.
03 · Reorientation
The third response is the one barely present in the conversation, and it is the one Orient is built around. It says: do not ban AI and do not capitulate to it. Change what is taught and assessed instead, so that AI becomes a multiplier rather than a substitute. That turns the question from “how do we stop students using AI” into “how should students use AI responsibly and what should they learn so that AI makes them better.” There is even a structural hint in the data, since the NYU Abu Dhabi work found the model weakest exactly where good assessment lives, in analysis and synthesis rather than recall, which means an assessment that rewards real understanding is already the harder one for a machine to fake.[8]
Part of the turn is aimed at the student’s relationship with the tool itself, not only at the syllabus. AI is here and is not going away, and a relationship left to form in the dark defaults to the shortcut, which is what the atrophy numbers earlier look like from the inside. Reorientation means informing that relationship deliberately: what the tool gives, what it quietly takes when it is used as a substitute, and how to tell the difference while it is happening. The self-awareness, as the study showed, is already there. What students have not been given is the practice that turns awareness into behaviour, and that is teachable.
The answer turns out to be old and well-evidenced, and it has very little to do with the model writing your essay for you. There is a body of work in cognitive science, decades deep, on how people actually come to understand and hold onto something, and almost no university teaches it explicitly. What is new is not the knowledge but the disappearance of the friction that used to keep students from acting on it. The very model that will hand a student a finished essay will, asked differently, turn their own lecture notes into a quiz they have to take, expand a thin outline into a structured one they then have to fill in, summarise a dense reading so they know where to spend their attention, draw a concept map that forces the connections into the open, transcribe a voice note from class into something searchable, show them how to study, studying the notes with them or hold them to a study plan with reminders. Some of that rests on the student’s volition but none of it is the model doing the learning for them. It is the model setting the work up so that the student does the learning, which is the whole distinction between outsourcing the work and outsourcing the understanding.
It is worth being precise here, because reorientation is easy to counterfeit. A law professor at UCLA, writing early on, decided the honest response was to let students use ChatGPT freely and simply hold them responsible for whatever they submitted, on the reasoning that the era when you had to write well to produce good writing had ended, and that writing from scratch was going the way of penmanship and long division.[11] It is the same move as the familiar claim that AI is just the next calculator, and the analogy does more harm than good. A calculator does the rote step so the student can spend their attention on the reasoning, whereas a general writing tool does the reasoning and leaves the student the formatting, which is the opposite trade. That sounds like reorientation, but it is closer to a careful capitulation, because it reorients by declaring the underlying skill obsolete. The position this piece takes is the opposite one. The skill is not obsolete but the entire point, and the tool is there to build it rather than to excuse its absence. Reorientation means changing the method while protecting the understanding, not lowering what counts as understanding until the tool can clear the bar.
That is also the more interesting thing a model can do, and it is bigger than answering questions. Pointed at a chapter, a past paper, or a recording, it can break the material into small interactive pieces a student works through rather than reads past, behaving less like a vending machine for answers and more like a curriculum developer, study guide and a coach. The crucial difference from the generic version of that promise is whose material it works on. This is not an infinite supply of generated lessons about woodworking or astrophysics conjured out of nothing. It is the student’s own notes, their lecturer’s own past papers, the voice note from this morning’s class, turned into something to do. That specificity, the fact that the content is theirs rather than generic, is most of why it works for a student in Freetown when a generic learn-anything app does not.
There is a further step in this, and it is the one that turns a collection of exercises into something closer to a practice. The weaker version of reorientation hands the student a set of things to do and then steps back, a supplier of worksheets that happens to run on a phone. The stronger version is present in the studying itself. It does not wait at the door for the student to come and ask for a quiz. It sits inside the reading, so that reading the chapter and studying it stop being two separate jobs and become one. The material the student reads becomes, in the reading, the material they study. That is the outsourcing distinction carried to its limit: the tool is not reading on the student’s behalf. It is reading alongside them, and refusing to let the reading stay passive by asking questions at recurring intervals.
And it can teach as it goes. A tool in that position has no reason to keep its methods invisible: it can name what it is asking for and why, walking the student through the way of studying while the studying happens, so that what transfers is the method rather than a dependence on the tool. A student who has been walked through the practice enough times owns the practice, and needs the walking a little less each time, which is the direction everything in this section is trying to point.
None of these ideas is new, and some of the evidence behind them is 40 years old. What changed is that the setup cost collapsed. Generating practice questions used to need a teaching assistant. Building a deck of cards used to cost an evening. Getting feedback on a spoken explanation used to need a study partner who already knew the material better than you did. The model erases all of those costs at once, which is how an approach that was always sound has suddenly become something a single student with a mid-range phone can actually sustain.
Under reorientation, the cheating problem does not vanish; it becomes architecturally less attractive. Sarah at Wilfrid Laurier was not cheating because she was lazy. She was cheating because the assessment in front of her gave her no reason to engage. Every one of these uses gives her a reason, because each returns something she can only get by having done the work herself. The prevention is not a detector or a ban. It is the simple fact that the honest path has been made more rewarding than the shortcut. Even Gary Marcus, who is no optimist about any of this, responded to the same Intelligencer piece by writing that the curriculum obviously needs a redesign, a more thoroughgoing one than any single fix, and that nothing improvised holds at the scale of a class of 100 or 500,[6] which is the same conclusion arrived at from the opposite temperament.
There is a system-level version of the same move, available to a ministry before a single classroom changes. If the incentive structure is the disease, as the motivation research says, the incentives are also the lever: prizes, competitions, and recognition that reward demonstrated understanding and its application rather than the polish of what gets handed in. Students look at learning the way the system visibly looks at it, and rewards are cheap to re-point compared with rebuilding assessment wholesale. It does not replace the classroom work. Rather, it just changes the weather around it.
I think one more line separates the genuine article from the counterfeit, and it is the sharpest. It is a question of which way the tool is trying to move the student. A tool built to capitulate wants to be indispensable, because its business is the student’s continued dependence on it. A tool built to reorient wants the opposite, the way a good teacher does. Its measure of success is the student needing it a little less, for any given thing, over time, while the understanding it helped build becomes the part that lasts. A tool that left students permanently unable to study without it would be a capitulation in better clothing, however much it dressed its prompts in the language of learning. The test is not how good the tool looks while the student leans on it. It is whether the student walks away more able to do the work alone.
05 · Where Sierra Leone fits
The argument for Sierra Leone, and for Sub-Saharan Africa more broadly, is not that the cheating problem is somehow worse here. It is that the institutional path of least resistance has not yet hardened. Western universities and academics are arguing about how to defend a 30 year old assessment model against a tool that broke its founding assumptions. Sierra Leonean universities have no such model to defend, because the infrastructure, the policy, and the detection capacity were never built in the first place.
Now, that sounds like a deficiency, and from one angle it is. But, from another it is an opening. The country that gets reorientation right at the university level, before capitulation has set as the default, will produce graduates who can compete with peers anywhere. The countries that stay trapped in the detect-or-capitulate binary will mostly produce graduates who cannot.
This is a different argument from the usual digital-equity framing, which positions African universities as behind, needing to catch up, requiring help to reach tools the West already has. The reorientation framing inverts that. These universities are not behind so much as early, with a chance to skip a phase the West is currently struggling through. The only real question is whether the alternative gets built before the default sets.
There is a parallel worth noting, though the usual shorthand for it is a little too clean. Africa did not so much skip landlines as never build them out to scale, and most people came online directly through mobile instead. The same is happening with banking, where branch networks stayed thin and mobile money became the default rather than the upgrade. The pattern holds wherever the older infrastructure never reached saturation, because there is little legacy to retrofit and little entrenched system to defend.
[ FIG.07 · The leapfrog pattern ]
Education is no different. The Western universities have a system to dismantle. The Sierra Leonean ones, given a credible alternative, can simply build. The cost of not yet having a system turns out to be far lower than the cost of having to tear one down.
There is one honest limit to draw around this argument before going on, and Sarah is the reason to draw it. Her habit was not formed at university. Walsh’s reporting records that she first used ChatGPT to cheat in the spring of her final year of high school,[3] which means that by the time reorientation reaches an undergraduate, it is already catching the problem late. The deeper version of this argument lives in the secondary schools, where the relationship with AI is being formed right now, in students who will arrive at university in a few years already being Sarah, or already being something better. But the lever there is a different one, and it would be dishonest to let the leapfrog reasoning simply slide downward, because at that level there is an entrenched system after all: the school-leaving examination is set by a regional council, is decades old, and is not going anywhere soon. A university owns its own examinations, so reorientation at university can change what is assessed. A secondary school answers to an examination it does not set, so reorientation there means changing how students prepare rather than what they are examined on, and it travels through different hands, the teacher, the school, the lessons class after hours, rather than through a lecturer who writes her own paper. That is a different design problem, and a different essay. This one stays with the university, where the lever is nearest to hand.
What a credible alternative means here is specific. It means a tool that respects the conditions on the ground, and a general assistant like ChatGPT is the wrong shape for the job, not because it is weak but because it is built for a different problem. It is built for general intelligence and for the average user somewhere in the world, which means it was never trained for this environment, and it answers in the general case. Ask it about a topic on a Sierra Leonean syllabus and it returns what it knows about the topic at large or what it got on the internet (to which there are limited data), not what this course covers, not how this examiner sets the question, not what sits in the student’s own notes or the lecturer’s past papers. It cannot be held to the material the student will actually be marked against, because it was never pointed at it. On top of that it assumes an always-on connection where data is metered and drops, and its useful tier is a monthly subscription billed in dollars, which is real money against a local wage. None of this is a flaw it grows out of as the models improve, because none of it is about the model. A frontier assistant gets better at being general. The job here is to be specific. A tool built for these markets has to run on the phones students already own, work when the connection does not, and live inside the workflows people already use, which means WhatsApp rather than Teams, voice notes rather than typed essays, and photographed past papers rather than tidy digital question banks.
Honestly, this is not a feature list but the design constraint that makes reorientation possible here at all. Without it, the conversation about AI in African universities collapses back into capitulation, where students use whatever works while lecturers quietly grade what comes in, or into policing, where a few well-funded institutions buy detection tools that do not work. With it, there is a third option.
06 · Orient
Orient is one attempt at building that third option. It is an Android (for now) study companion for university students and their lecturers in Sierra Leone, that lives on their phone and works without a subscription or an internet connection. It is built to make AI useful inside the workflows people already have rather than asking them to move to new ones.
A student brings things into it through Android’s share sheet, straight off WhatsApp: a voice note, a PDF, a photographed past paper, a lecturer’s outline. Anything that started elsewhere can come in through an in-app capture. All of it lands in a Library that holds the raw material and everything made from it side by side, composable and taggable. From any of it the student can ask for the kinds of things described earlier: a quiz drawn from their own notes, an outline expanded, a dense reading summarised, a concept map, a voice note transcribed, a plan with reminders. The consistent point is that Orient helps them work out what to study and how to go about it without ever doing the studying for them. The experience on the surface is that simple.
There is a second way in, equal to the first. Beyond bringing material into the Library and asking it for a piece of work, the student can open a source and read it inside Orient. The reading is not a dead end that the studying happens after. The tool is present while the student reads, quietly, without crowding the page. At the end of a chapter it asks, in passing, what the student took from it, and keeps the answer. The passages the student marks while reading become things the tool returns to them rather than ink that fades on the page. A reading that would otherwise have been one more passive pass through a borrowed PDF becomes, without much being asked of the student, a real session of study. None of it requires the student to know the name of a single study technique. The tool carries the technique so that the student can spend their attention on the material itself.
All of it happens on the device itself, with Gemma 4 running locally through LiteRT-LM, so daily use needs no internet and no data leaves the phone, a choice whose full case, and full costs, I have argued elsewhere.[16] The roadmap adds a connected tier on top: heavier work, a long-document synthesis or a rich audio overview, routing out to cloud tools when there is wifi, with the app making the call and showing its working rather than asking the student to choose between offline and connected.
It is worth saying plainly what makes any of this reorientation rather than the same shortcut in a friendlier wrapper is the harness and product strategy, because the difference is not the model. The model underneath is close to the one the substitution tools use, and it is becoming a commodity that improves for everyone at once. What separates the two is the harness built around it, the set of rules the tool runs under: what it demands as input, what it refuses to do, what it grounds every claim in, and where it stops and makes the student do the work instead. The same model, wrapped in two different harnesses, produces opposite results. A general assistant ships with almost no harness on purpose, because every constraint is friction and friction lowers use. Orient adds the friction deliberately, and it is not the old friction brought back. The old kind was tedium that happened to guard the learning; this kind is the learning itself with the tedium stripped away. Here the friction is the practice. It asks for the student’s own material before it will act, answers from that material rather than from its own store of generic knowledge, says so when the material does not cover a question rather than inventing an answer, and hands back something the student still has to do rather than something already done. That is what the harness is for, and it is the part a better model does not replace. A better model only gets better at being general, which was the wrong shape for this job to begin with, and the harness is what makes a tool specific to it.
[ FIG.09 · How Orient is put together ]
All of that is what a student gets on their own. What turns it from a private study aid into something closer to a change in how learning works here is the lecturer, and that deserves its own treatment.
07 · The lecturer
Almost every argument about AI in universities treats the institution as the thing that has to change. That is an assumption worth dropping. The smallest place where reorientation is already complete is not the institution but a single lecturer and their class.
Go back to Williams for a moment. He was not lazy and he was not cynical. He was trapped, because the institution handed him two moves and no third one. He could police, with detectors he did not trust, or he could capitulate and grade on the ability to use ChatGPT. When both felt like a betrayal of the job, he quit. What he lacked was not willpower but a buildable third option, and the structure around him never offered one.
Now set a different professor beside him, one who made the third move inside a real classroom. The seed of it was public: in the spring of 2023 Gary Marcus had suggested turning the flawed tool into a teachable moment, letting students use it and making the dissection of its output part of the assignment itself.[6] C.W. Howell, teaching religious studies at Elon, picked the suggestion up and built it into practice, having each student generate an essay from the model and then grade it the way a professor would, hunting for the places it had invented its sources and its arguments. Every one of the 63 essays came back with confabulations, invented quotes, invented sources, real sources mischaracterised, a rate that stunned even Howell, who had expected it to be high but not total. His students, half of whom had assumed until that week that the tool was simply reliable, came away with fears in their own vocabulary of mental atrophy and misinformation, and one of them wrote that the worry was not AI climbing to where we are but people sinking to where AI is, which is the whole cognitive case of this essay in a student’s sentence. When Marcus returned to the exercise two years later in his response to the Intelligencer piece, standing by it while granting that no single assignment matches the scope of the problem, Howell surfaced in the comments with the rest of the story.[6] It is what Howell is structurally that matters here: a single lecturer who turned a public idea into a working classroom, with no new policy, no budget, and no one’s permission.[10] He is the existence proof that the third move can be made at the level of one room.
And then he left anyway. Howell quit teaching after that semester, and in that same comment described leaving higher education as feeling like “fleeing a burning building”. But that does not weaken the lesson; if anything, it sharpens it. A lecturer can make the third move alone, but making it alone, improvising the whole craft of it in real time against the grain of everything around him, cost enough that even the person who managed it walked away. The idea had been in public for anyone to take and try out; what stayed scarce was a classroom that ran it, and a lecturer who could afford to keep running it. The difference between the Howell who stays and the Howell who leaves was never willpower alone. It is also whether the third move has to be improvised into every classroom from a bare idea, or arrives as something a lecturer can simply pick up and use.
Orient hands that third move to the individual lecturer, and the part that matters most is that it needs no one’s permission. A lecturer authors a skill, a plain instruction file encoding their own course’s exam format, their past papers, the way they mark, and the kind of question they actually set. They drop it into the class WhatsApp group as an attachment. From that afternoon their students are practising against their teaching rather than a generic chatbot’s idea of the subject. There is no learning management system in the loop, no IT department, no procurement, no faculty-senate resolution. The unit is the class, and the win is specificity. The student is now working against the exam they will actually take, set by the person who will actually mark it. That is reorientation fully realised, at the level of one room, and it happens without the administration so much as noticing.
If several lecturers in a department do the same thing, a practice emerges from the bottom. That is the honest way an institution eventually comes around, not by decree but because the practice is already there and working by the time the institution looks up. The institution, in other words, is the accelerant and not the precondition.
The more interesting scope is the one beyond the institution, and it falls out of a single technical fact: a skill is a portable file. A good organic-chemistry skill authored by a lecturer at Fourah Bay is not chained to that lecture hall. It can travel to a student at Njala, to a lecturer at another university who adapts it, to a candidate enrolled nowhere in particular, to a student in Monrovia or Banjul facing the same kind of paper. The lecturer stops being bounded by the room they teach in and becomes a node in a distributed, cross-institutional network of pedagogy. Authoring a skill that many people come to rely on becomes a kind of teaching reach that no institution granted and none can revoke. In the policing-and-capitulation world, a great teacher’s influence ends at the classroom door. In this one it is an artefact that can outrun their employer entirely.
And because the tool gives lecturers and students composable pieces rather than one prescribed workflow, they will compose uses no one designed for: a cohort co-authoring a revision pack between them, a lecturer encoding the quirks of a particular past examiner, a skill that turns out to be less a study aid than a way of coordinating a study group. You cannot really predict users, and in an emerging market that unpredictability is a feature rather than a risk. The person building in Freetown cannot anticipate every constraint a user will hit in a different town, and a permissive tool lets the users route around the gaps the designer never saw.
Which brings me to the worry sitting underneath all of this. The piece is proposing that this could change how people here approach learning if it is adopted, and the obvious objection is this: what if the institutions are not as open to it as the argument assumes? The answer is that the thesis does not depend on them. Picture it as three rungs, of which only the bottom two have to hold. A student using Orient alone still gains, and gains for real, because the tool works whether or not anyone in authority approves of it. A lecturer and their students using it together, the way it is meant to be used, gain more, and gain something curriculum-specific the student could never assemble alone. The institution adopting it sits on top as an accelerant. If the institution never moves, you are not left with a failure. You are left with thousands of students who learned better and hundreds of lecturers who refused the detect-or-capitulate binary, which is the base case rather than the disappointment.
[ FIG.10 · Three rungs, only two of which have to hold ]
This is why the product is built on the side of the lecturer rather than aimed over their heads at the institution. The lecturer who refuses both responses, who will neither police nor capitulate, has until now had nowhere to stand. The whole architecture exists to give them a place to stand and a thing to build. That is exactly the move the universities in the Intelligencer piece could not give the Williamses on their own faculties.
08 · What this requires
I want to be honest about the difficulty. Building the product is not the hard part. The slow part is the institutional one, and the three-rung reframing in the previous section is what keeps that slowness from being fatal to the whole idea.
Because Orient runs on a student’s own phone, the first wave of adoption needs no institutional blessing at all. A student installs it and gets better. A lecturer installs it, authors a few skills, and shares them with a class. None of that waits on permission, so the first wave can happen entirely beneath the institutional layer.
For reorientation to become the default rather than a scattering of good practice, the institutional layer does eventually help, and the rungs above already carry the logic of how the two kinds of work relate. What is worth adding is the texture of each. On the lower rungs adoption compounds on its own: a lecturer authors a skill, a student practises against it and does better in the course, the lecturer notices and authors more, and other lecturers follow. The institutional work is slower, made of ministry conversations, visible champion lecturers, and evidence gathered in this specific context, and it compounds too. It is not a gate the first kind has to pass through. It is the accelerant the top rung was always going to be.
Sierra Leone is the right place to start because its constraints force the right choices. A tool that works for a Freetown student on a mid-range Android with intermittent wifi, in a country where the scaffolding for either policing or capitulation barely exists, will work almost anywhere with the same constraints, in Lagos, Nairobi, Dhaka, or Manila. The dynamics underneath travel even when the details do not.
If this works in Sierra Leone, it works.
09 · Closing
The Intelligencer essay ended on a kind of resigned shrug. Williams quit. Wendy handed in an essay on Erving Goffman she had neither written nor read. The institutions had no answer, and the piece did not pretend to supply one.
The honest position is that the West may take a long time to reach a good answer, if it reaches one at all. The inertia is real, the assessment models are entrenched, and the public conversation is dominated by detection and capitulation, precisely because those are the responses that ask nothing of the institution.
The argument here is that Sierra Leone, and places like it, do not have to wait for that conversation to mature. The inertia is not yet in place, the models are not yet entrenched, and the default has not yet been chosen. The opening is real and it is narrow, and it closes a little further every semester that irresponsible AI use deepens without a credible alternative on offer.
Reorientation is harder to build than capitulation and harder to defend than policing. It means accepting that the old assessment model cannot be preserved, that the friction students used to study under is gone for good, and that the answer is not to restore the old friction but to teach the one thing AI cannot stand in for: the skill of knowing how to learn well and deliberately with the tools that now exist. The principle behind it compresses to a single line: outsource the work, not the understanding. That is the skill Orient is shaped to grow.
If the country gets that right at the university level, its graduates compete, and if it does not, they do not. As AI use deepens in institutions, the window for that gets shorten. So, the build is real, and the work for Orient is still underway.
Until next time, namaste.
Special thanks to Diane Taylor for proofreading earlier versions of this writing.
Researched, pressure-tested, and drafted with Claude and Perplexity. The argument, and the errors, are mine.
Sources
- Apata, O. E. et al. “Generative AI in African Higher Education: A Systematic Review of Opportunities, Ethical Concerns, and Preparedness.” SSRN, August 2025. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5530978
- Sangwa, S., Ngobi, D., Ekosse, E. & Mutabazi, P. “AI governance in African higher education: Status, challenges, and a future-proof policy framework.” Artificial Intelligence and Education 1(1): 2054, 2025. https://ojs.sin-chn.com/index.php/AIE/article/view/2054
- Walsh, J. “Everyone Is Cheating Their Way Through College.” Intelligencer / New York Magazine, May 2025. Underlying survey: Study.com, January 2023, 1,000 college students. https://nymag.com/intelligencer/article/openai-chatgpt-ai-cheating-education-college-students-school.html
- Liang, W. et al. “GPT detectors are biased against non-native English writers.” Patterns (Cell Press), July 2023. https://arxiv.org/abs/2304.02819
- Vanguard. “Nigerians tackle American author for claiming ‘delve’ is only used by ChatGPT.” April 2024. https://www.vanguardngr.com/2024/04/nigerians-tackle-american-author-for-claiming-delve-is-only-used-by-chatgpt/
- Marcus, G. "Everyone is cheating their way through college." Marcus on AI (Substack), May 8, 2025. Howell's account of leaving teaching appears in the comments of the same post. Marcus on AI The big story making rounds today is New York Magazine’s Everyone is cheating their way through college, using GenAI, which concludes, inter alia, and probably correctly that…
- Anthropic. “What 81,000 People Want from AI.” March 2026. Interviews with 80,508 Claude users across 159 countries and 70 languages. https://www.anthropic.com/features/81k-interviews
- Zaki, Y., Rahwan, T., et al. Study comparing ChatGPT and student performance across 233 questions from 32 university courses, Scientific Reports, 2023. https://www.nature.com/articles/s41598-023-38964-3
- Leffer, L. “ChatGPT Can Get Good Grades. What Should Educators Do about It?” Scientific American, August 2023. https://www.scientificamerican.com/article/chatgpt-can-get-good-grades-what-should-educators-do-about-it/
- Howell, C.W., Baker, C. & Stylianopoulos, F. “To Educate Students about AI, Make Them Use It.” Scientific American, November 2023. https://www.scientificamerican.com/article/to-educate-students-about-ai-make-them-use-it/
- Villasenor, J. “How ChatGPT Can Improve Education, Not Threaten It.” Scientific American, February 2023. https://www.scientificamerican.com/article/how-chatgpt-can-improve-education-not-threaten-it/
- Tuomi, I. The Impact of Artificial Intelligence on Learning, Teaching, and Education. European Commission, Joint Research Centre (JRC113226), 2018. https://publications.jrc.ec.europa.eu/repository/handle/JRC113226
- Grinschgl, S. & Neubauer, A. C. “Supporting Cognition With Modern Technology: Distributed Cognition Today and in an AI-Enhanced Future.” Frontiers in Artificial Intelligence 5: 908261, 2022. https://www.frontiersin.org/articles/10.3389/frai.2022.908261/full
- Reich, J. & Ruipérez-Valiente, J. A. “The MOOC Pivot.” Science 363(6423): 130-131, 2019. https://www.science.org/doi/10.1126/science.aav7958
- Barnum, M. “Why Sal Khan Is Rethinking How AI Will Change Schools.” Chalkbeat, April 2026. https://www.chalkbeat.org/2026/04/09/sal-khan-reflects-on-ai-in-schools-and-khanmigo/
- Rogers, T. “Open Models and the Sub-Saharan Region.” Dispatches from building with LLMs, dev.to, May 2026. Winner of Google’s Gemma 4 writing challenge. https://dev.to/thelonewulf_wa/open-models-and-the-sub-saharan-region-29nh
Top comments (0)