The most criticised container in education is also the most durable. Here is what we learned from health workers participating in The Geneva Learning Foundation’s courses about why that is so difficult to change.
The obituary that keeps being written
In 2013, Sebastian Thrun looked at the completion rates for the massive open online courses that had made him famous and told Fast Company: “We have a lousy product.”
Fewer than 10 in 100 of his students finished.
He pivoted Udacity toward corporate training within weeks.
That was the moment the course was supposed to die, and it was neither the first nor the last.
By 2019, Justin Reich and José Ruipérez-Valiente had put numbers on the aftermath in Science: 52% of MOOC registrants never entered the courseware at all, and second-year retention had fallen from 38% to 7% across four cohorts.
Completion, they wrote, had “barely budged despite six years of investment in course development and learning research.”
Thirteen years after Thrun’s verdict, Lawrence Malama, an environmental health officer in Zambia, enrolled in a course on malaria.
He had tried it once before and quit.
This time he finished, and he wrote down what he got:
“am too junior in the ministry but i had a preverage to exchange feedback from Doctors and other high ranking practitionals.”
That sentence records who is permitted to be read by whom, and it is the reason the course has outlived every obituary written for it.
The question in the title is being asked again now, with more force, because generative AI can produce on demand almost everything a course used to contain.
Answering it means separating two things the debate keeps welding together: what a course delivers, and what a course is.
The first is finished.
The second is the only institutional form the world has agreed to recognise, measure, and pay for.
That is a more uncomfortable claim.
It means the course survives less on pedagogical merit than on the absence of any competitor with an address.
2. Why the learner accounts here are evidence, not decoration
This article draws on 5 corpora of end-of-course reflections written by learners in courses run by The Geneva Learning Foundation: malaria in English and French, health equity in English and French, and healthy ageing and menopause in English.
Several thousand responses, each answering the same 5 prompts about professional change, social connections, practice, influence, and worldview.
They are used here as evidence rather than as illustration, for a reason that belongs to the argument itself.
TGLF’s questions are deliberately inverted.
As Reda Sadki puts it in his account of the learning science underpinning peer learning, the system does not ask what should be done about, for example, vaccine hesitancy.
It asks a practitioner to describe a time she addressed it, what she did, and what happened.
He makes the epistemic stakes explicit in his analysis of the cognitive commons: artificial intelligence may be able to answer the first question, but only someone who was there can answer the second.
A sector that tells frontline workers their observations are “merely anecdotal,” he argues, is not just being rude.
It is degrading the only human capacity that can catch a fluent machine being wrong.
If that argument is right, then asking 4,000 health workers what a course did to them is an appropriate instrument.
It is also a limited one, and the limits should be stated before the quotes start rather than buried at the end.
- Four of the 5 files are quality-filtered, and every response quoted here comes from someone recorded as having passed.
- The learners who ran out of time, lost their connection, or were defeated by a schedule built for another time zone are absent by construction.
- There is no comparison group: no self-paced arm, no AI-tutored arm, no unstructured community running the same material.
- Everything is self-report collected at the moment of completion, before any of it had to survive a busy year.
- The evaluation column attached to each response concedes that most of them generalise rather than narrate. They gesture at experience instead of putting a date and a place on it.
What the corpus can establish is narrow and still useful: what learners credit, in their own words, when nobody is asking them to rate satisfaction.
3. The critics of the course, and what they actually established
The people who did the most to establish what is wrong with courses were not, for the most part, hostile to education.
They were trying to fix a measurement problem.
Benjamin Bloom set the terms in 1984.
His students found that learners tutored one-to-one with mastery methods performed 2 standard deviations above conventionally taught classes.
This placed them above 98% of the control group.
He did not conclude that everyone should hire a tutor.
He posed the 2-sigma problem: find a method that recovers as much of the tutoring effect as possible at class scale.
Four decades of work followed, and a 2020 meta-analysis of tutoring studies found an average effect of roughly 0.37.
It is real, valuable, and nowhere near 2 sigma.
The course, in Bloom’s framing, was the constraint you had to work around, and nobody has worked around it.
Dave Cormier attacked the container itself.
In Rhizomatic Education: Community as Curriculum (2008), he argued that when knowledge is a moving target, curriculum cannot be an input at all: “the community is the curriculum,” negotiated in real time by the people learning.
His formulation is elegant, empirically defensible in fast-moving fields, and almost entirely absent from the world’s education systems.
Ask a hiring manager to evaluate a rhizome.
Reich went after the disruption story with data, and his conclusion in Failure to Disrupt was that MOOCs were a vehicle for conventional teacher-directed instruction delivered at scale, which is why they produced conventional results.
Worse, they concentrated participation among the already-affluent rather than closing any gap.
Even the measurement infrastructure has turned on itself.
The Carnegie Foundation, which gave the world the credit hour, announced its intention to replace time as the essential measure of learning: 7,200 minutes of instruction per credit, an hour a day for 24 weeks, a standard it now calls inadequate.
That was 2022.
The credit hour is still there.
A note on lineage is owed here, because a certain set of theorists tends to appear whenever TGLF’s model is described: George Siemens on connectivism, Bill Cope and Mary Kalantzis on New Learning and knowledge production, Karen Watkins and Victoria Marsick on informal and incidental learning.
They appear because Sadki knows them, corresponded with them, and in the case of Cope and Kalantzis was redirected by them: he had built 80 e-learning modules to transmit knowledge from the International Federation of Red Cross and Red Crescent Societies before they told him that learning happens through knowledge creation rather than consumption.
That is intellectual autobiography, and it is a legitimate but different kind of citation from the course-specific record above.
Watkins and Marsick’s finding that most consequential workplace learning happens informally, outside courses, is the sharpest of the three for this argument, and it comes with the same problem as Cormier’s: it describes something real that no institution knows how to certify.
The pattern across all of it is not that the critics were wrong.
They won the argument decisively, and then watched the object of their criticism outlive them.
4. Why the better ideas have no address
Sadki’s cognitive commons essay contains the clearest available explanation, and it does not mention pedagogy at all.
He describes two health workers.
The first is a national programme manager in Nigeria or Ghana who has been through “in-service training with credit points that counted toward her career progression, mid-level management modules, a field epidemiology programme where a designated mentor reviewed her outbreak investigation and sent it back for revision, and 15 years inside an institution with a memory.”
The second is a district immunisation officer 3 administrative layers below, who “attended a two-day cascade workshop in the state capital four years ago, received a per diem and a slide deck, and then spent a decade figuring out alone why the children in three settlements kept being missed.”
She has real expertise.
“Nobody designed the pathway that produced it.”
Most of what formed the first woman is course-shaped.
Credit points.
Modules.
A programme with a mentor ratio and a completion.
The system that produced her expertise could see it, count it, and pay her more for it.
The second woman’s expertise is invisible to every instrument her employer owns.
This is what the “rhizome” cannot do.
Informal learning, communities of practice, networked knowledge production, all of them describe how professionals actually get good, and none of them has a transcript, a registrar, a credit, or a line on a promotion form.
There is no agreed measure and therefore no agreed value.
Which leaves the course holding a monopoly it did not earn on pedagogical grounds.
Learners say this out loud, and it is the least sentimental thing in the corpus.
Kingsley Kofi Nignere, in Ghana, had spent years implementing malaria projects:
“I have supported in implementing malaria projects but did not have any certification as an evidence of being a malaria advocate. Participation in [TGLF’s peer learning course] Malaria Turning the Tide has increased my identity and my portfolio by awarding me with this golden certificate. I’m greatly honored with this international certification.”
And later, on the equity course, having been sponsored:
“I was indeed recognized enrolling me into this special course HEART with a prestigious sponsoship. also reviwing my project with a feedbacks has raise my status. I woudnt have been able to afford this school fees.”
He is describing the conversion of tacit competence into a form his system can read, not learning itself.
That conversion is the course’s actual function, and no alternative container performs it.
5. What AI broke was the willingness to pay
The strongest evidence for obsolescence is about markets.
The learning evidence is genuinely impressive and genuinely narrow.
In a Harvard-designed randomised trial, physics students working with a GPT-4 tutor posted learning gains more than double those of an active-learning classroom, in less time, with 83% rating the AI’s explanations as good as or better than their instructor’s.
A meta-analysis reports effect sizes of 0.857 on achievement and 0.803 on motivation.
Note what that trial compared: not AI against a bad lecture, but AI against the intervention that beat lectures.
On the transmission of well-defined content to motivated learners, the argument is over.
Then the same literature turns.
The OECD’s Digital Education Outlook 2026 found that when learners over-rely on generative AI, metacognitive engagement measurably declines even as task performance rises.
Sadki, who attended and wrote it up, records the phrases that stuck: Dragan Gasevic’s “metacognitive laziness,” and Jason Lodge’s image of an e-bike used to get fit.
Students wrote better essays with a model and could not afterwards remember what they had written about.
A critical review of 40 studies names the same thing as a growing conflict between meaningful learning and visible academic achievement.
A discourse analysis of undergraduates learning graph theory found AI raising learners’ confidence while reducing their curiosity.
Meanwhile the money moved.
PwC, reading nearly a billion job postings, finds skills in AI-exposed occupations evolving 66% faster than elsewhere.
That is faster than any fixed-length programme can be authored, accredited, and delivered.
China’s Ministry of Education discontinued 12,200 degree programmes between 2021 and 2025, more than 30% of its total, on the reasoning that AI now performs the core tasks they taught.
In the United States, new international enrolment fell 17% in autumn 2025, costing an estimated $1.1 billion and 23,000 jobs, with projections of up to 112,000 fewer students and $3.4 billion in 2026-27.
Graduate programmes at public universities that had balanced their budgets on full-fee international students are now discovering which of their offerings anyone will pay for on merit.
Karen Watkins, whose research on informal learning is cited above and who is TGLF’s president, has watched that arithmetic from inside a US public university.
The view from there is that certification may increasingly be the only part of the package a buyer will still fund.
The corporate learning industry has drawn the obvious conclusion.
A vendor white paper titled The Course Is Dead predicts that within 5 years, learning management systems, instructional design models and completion certificates will look as quaint as a rotary phone.
Arist, which began as an SMS microlearning company, now sells “AI agents for talent and enablement” on the explicit premise that the function of the future is not 50 instructional designers building courses but a small team orchestrating agents: a research agent conducting thousands of personalised interviews to surface competency gaps, a creator agent ingesting 5,000 pages and generating cited, translated, role-specific lessons in under a day, an analytics agent tying it all to performance rather than completion.
The commercial pitch is striking: 95% adoption against under 30% for typical e-learning, 9 times the knowledge retention, launch in under 3 hours against 3 to 12 weeks.
Those figures appear only in the marketing of the company reporting them.
No independent verification exists in the literature, and the obsolescence argument leans on them anyway.
And the architecture concedes the point it is trying to win: Arist’s own materials still describe cohort creation, live experiences, and role-based communities.
The agents automate what happens inside the container.
Nobody has removed the container.
The pattern that emerges from the systematic reviews is not a verdict on courses at all.
A PRISMA review that screened 5,419 records down to 47 studies identifies the determinants of whether AI helps or harms as institutional readiness, faculty AI literacy, and governance maturity.
Meanwhile the OECD reports generative AI use approaching universal among students. 95% of UK undergraduates use it, while roughly 19% of institutions worldwide have a formal policy.
Obsolescence turns out to be conditional on implementation, and implementation is mostly absent.
6. What the learners paid with
The learners in TGLF’s courses did not pay money.
They paid attention, in a currency they have very little of, and what they say they bought is specific.
A bounded window. Sesugh Deborah Oryiman, after the malaria course:
“My dedication towards completing tasks improved greatly during and after the period of this course. It has helped me to push myself to think fast and deeply, draft responses and meet up with deadlines.”
Malama’s version is blunter.
He had failed to finish the same programme the year before, and this time, in his words, “i was bored enough to conquare it.”
Ochola Isaac Timothy measures his own change in calendar units:
“The two weeks of analysis and collaboration have strengthened my confidence.”
Being read, and having to read. Joshua Kofi Nnanchom, a health worker who never got the chance to enrol in professional health studies, arrived at a definition of learning that no curriculum handed him:
“I have learnt that learning is not completed until someone review your work and give feedback to it.”
Temidayo Olowoopejo:
“The use of the rubric changed how I read my own work.”
Fatima Abdulrahman, a pharmacist accustomed to seeing everything clinically, records what review cost her:
“It forced me to accept my views are not always perfect for a particular situation.”
Peers with rank. Jonah Chesang, who works on malaria programmes in Kween, Uganda, describes what reading other people’s real accounts built:
“a strong sense of professional identity across borders, not just as people doing similar work but as people willing to be honest and specific about what went wrong and how they learned from it.”
Oyinade Uvere states the counterfactual directly.
The peer-review process “created a sense of shared learning and collaboration that I might not have experienced through self-study alone.”
Language for something that had no name. On the ageing course, Hema Rao Dhande stopped calling what she was seeing in women in their 40s “ageing” or “stress”:
“After reading the primer, I talked to a woman in my 40s who had night sweats and mood swings. Instead of dismissing it, I asked Do you wake up at night sweating? Naming it as menopause helped her feel less alone and more confident to seek care.”
She has a plan with a number in it: for the next 4 weeks she will ask every woman aged 40 to 55 about sleep, hot flashes and mood changes, and track the answers to see whether women start feeling safe enough to talk.
And she has a systemic reading of her own silence:
“I used to think silence about menopause was normal. Now I see it as a data cliff and a sign of ageism and sexism. Women after 49 are not counted in clinic reports, and their pain is called just aging.”
Ramandeep Kaur, a nurse educator who has been through surgical menopause after a hysterectomy, describes the same shift from the other side:
“understanding menopause from textbooks is very different from understanding it through womens lived experiences.”
Nandhana M.S., a nursing student:
“Before this course, I thought menopause was simply the end of menstruation.”
Every fact any of them needed was already free.
The symptoms of perimenopause sit in a thousand open-access pages, and a chatbot would have supplied the list in 9 seconds, in Malayalam or Marathi if asked.
What the course supplied was a reason to ask a specific woman a specific question on a specific afternoon, and 3 strangers obliged to read the answer.
7. What a model can simulate, and what it cannot
It is tempting to list the things a chatbot cannot do.
The list is shorter than it looks.
A model can hold a deadline.
It can nudge, remind, escalate, and schedule better than any course administrator.
It can write a rubric.
Reda Sadki’s own 2026 vision for an AI-supported platform has AI diagnosing a learner’s knowledge, researching current evidence in minutes, generating a project tailored to her context and the rubric by which it will be judged, then translating everything so everyone works in their own language.
It can give feedback that is more detailed, more patient and more available at 4 a.m. than any peer reviewer.
It can issue a certificate.
It can simulate encouragement convincingly enough that most learners will not test it.
What it cannot do is be someone whose regard costs something.
This is worth unpacking, because it is not a mystical claim about human warmth.
Consider what actually happened to Malama.
A doctor, several ranks above him in a ministry, spent 20 minutes reading the written work of a junior environmental health officer and wrote back.
That doctor’s attention is scarce, rationed, and normally allocated by hierarchy.
The course reallocated it.
The information in the feedback may have been ordinary.
The expenditure was not, and it is the expenditure that carries the signal.
Malama now knows something he could not have learned from any tutor, however capable: that his account of his own work is worth a senior colleague’s time.
A model’s attention is infinite, and therefore worth nothing as a signal.
It cannot confer standing for the same reason a printing press cannot confer a knighthood.
It lacks neither capability nor fluency.
Standing is a claim on other people’s scarce regard, and a system that can produce infinite regard produces none.
Recognition works the same way.
Nignere’s “golden certificate” means something because a named institution attached its reputation to it and could be embarrassed by his failure.
A credential a machine can mint on request is a receipt, not a recognition.
Sadki’s design decisions follow this logic more closely than his AI writing sometimes suggests.
His framework for AI as co-worker puts the machine inside the structure and tethered: TGLF’s first agentic hire has a role description, a supervisor, a bounded scope and no authority to send anything externally without human review.
And the integrity layer he describes exists to answer one question at scale. It is not “is this text good?” but did this learner actually keep her commitments to her peers.
That is a system built to protect the scarcity of human attention, because the scarcity is the product.
8. The paragraph pasted 5 times
On the ageing course, one nursing student wrote a competent paragraph about menopause and quality of life, then pasted the identical paragraph into the second prompt, the third, the fourth and the fifth.
Another answered 4 different questions with one block of text about social connections.
Both passed at 100%.
On the equity course, Maureen Asembo closed an unusually strong reflection with a note: “NB Used ChatGPT to reorganize my thoughts and remove grammatical errors.”
In Sankuru, a technical assistant records using Copilot to sharpen a project narrative after a reviewer complained it lacked lived experience.
This is where the obsolescence question actually bites, and it has nothing to do with cheating.
For roughly a century, the written artefact has done two jobs inside a course.
It stood as a proxy for thinking.
You could not produce the essay without having done the cognition.
It also served as proof of participation, the evidence that entitled you to the credit.
Both jobs depended on a scarcity that no longer exists.
When fluent prose costs nothing, a course whose only evidence is text cannot distinguish a learner who thought from a learner who prompted, which means it cannot honestly certify anything.
Slimi’s synthesis of studies puts it precisely: AI exposed weaknesses already present in assessment design rather than creating a new crisis.
The lock had been broken for years.
The machine simply walked in and demonstrated it.
Sadki has written about the human cost of this more carefully than most.
In the cognitive commons essay, he describes Joseph Ngugi, a diligent Kenyan Scholar who began submitting generic AI-generated prose and refuses to read it as disengagement.
He calls it the transparency trap: disclose AI use and risk having your knowledge dismissed, conceal it and carry the ethical weight alone.
Ngugi never got to write the next chapter to that story, as he passed away last year.
May he rest in peace.
Sadki’s conclusion is that prohibition cannot fix this and only conditions in which transparency is safe will.
If TGLF’s evidence rests on first-person practitioner accounts, and those accounts increasingly pass through an AI model before they arrive, the foundation of everything it claims is vulnerable.
So one obsolescence in the debate is real and specific.
Not the course.
The essay.
9. The door
Here is the reconciliation, and it is a strategic position rather than a theoretical one.
TGLF’s own theory argues against its product line.
Reda Sadki’s draft learning theory states that under conditions of complexity, learning and implementation are one process, that just enough structure makes that process reliable at scale, and, flatly, that “the appropriate unit of design is not the course but the persistent network.”
The internal record shows a deliberate drift away from course language toward movement and assembly terminology.
The diagnostic rubric fixes the ideal at “just enough structure,” flagging both the rigid script and the unstructured conversation as failures.
And the courses stayed.
In the vision document for an AI-supported platform, after the machine has diagnosed the learner, built her project, written her rubric and found her peers, the system invites those people “to join a starting ’course’.”
The word is preserved, in scare quotes, by an organisation that has spent a decade explaining why it is the wrong word.
The reason is the one in section 4.
A working nurse in Kerala knows how to enrol in a course on her lunch break.
She does not know how to enrol in a “rhizome” or even a “movement”, and neither does her supervisor, her regulator, or her promotion panel.
The course is the only door in the building with a handle everyone recognises.
Behind it, TGLF builds the thing that has no name and no measure: a 16-day cycle of writing, reviewing 3 peers against a rubric, and revising.
Then an Impact Accelerator whose rhythm is a commitment on Monday to one action finishable by Friday, comparison with peers midweek, a Friday report that includes the failure, and a new commitment the following Monday with that behind you.
TGLF’s own pedagogy matrix declines to oversell the middle of that sequence.
It scores the 16-day peer learning exercise as only moderate on praxis, “plan-making is preparatory, not yet action,” and weak on scale, “capped at 3-peer review group.”
The Accelerator scores strong on praxis and weak on scale.
Teach to Reach scores strong on scale and moderate on praxis.
They are not interchangeable, and no single one of them is the answer.
The course is the door.
The scaffolding behind it is what the door is for.
This is also why the scaffolding cannot be replaced by continuous availability, however attractive that sounds.
Reflections from learners who wanted more are among the most instructive in the corpus.
From Chad, on the equity course:
“Support after the course would have helped me to consolidate what I gained.”
From a participant travelling when it ended:
“I haven’t yet had the time to apply what I learned.”
From the ageing cohort, flatly:
“My current professional skills and core attitudes have mostly remained the same.”
The boundary is what produced the work, and the boundary is where the work is abandoned.
Both statements are true, and the design problem is to build a second bounded thing that opens as the first one closes rather than an always-open door nobody walks through.
10. Sankuru
Dr Mathieu Kalemayi Ndjibu works as a technical assistant on immunisation in Sankuru province, in the Democratic Republic of the Congo.
His subject is children who have received zero doses of any vaccine.
He entered the equity course with energy and hit friction immediately: the time allotted for the breakout discussions was very short, and his connection kept dropping.
He volunteered to take notes anyway, and spoke first.
His group applied the 5 whys to the death of a young woman in a village, rejected by her community over a pregnancy judged culturally unacceptable, and could not reach consensus on the root cause.
The plenary rescued the analysis.
Then service obligations took him out for 3 days.
He missed the general assembly and the remote coffee.
He describes a period of stress that broke only when a facilitator told him every session had been recorded.
He watched them all and caught up.
His attempts to reach his assigned paired peers failed.
Every one of them.
His project came back with reviews.
One reviewer said the account lacked personal lived experience, which he judged simply wrong, since it was drawn from his own work, and he says so.
The misreading taught him something anyway about legibility in a shared document.
A second reviewer told him to define SMART objectives, so he rebuilt his action plan as a planning table with steps and named owners.
He left with a WhatsApp group and a set of phone numbers that, he writes, still connect them today.
Count what failed there: bandwidth, pairing, scheduling, and one reviewer’s reading comprehension.
Count what survived: an obligation to produce something, 3 strangers who read it, a disagreement he was permitted to win, a table with named owners, and a group chat that outlived the certificate.
11. Coda
Are courses obsolete?
As a way to move information into a person, yes, and the market has already priced it.
As the only container the world has agreed to recognise, measure and reward, no.
This is not because it works well, but because a century of better ideas never built an institution.
That is not a comfortable defence of the course.
It is an indictment of everything around it, and it locates the real work: not proving that peer learning is better than a lecture, which Bloom settled in 1984, but building the measures, the recognition and the pathways that would let the district immunisation officer’s decade of hard-won judgement count for as much as the programme manager’s credit points.
Until then, somewhere in Maharashtra this month, a woman in her 40s will mention that she has been tired lately, and be asked whether she wakes up at night sweating.
References
Wilson-Trollip, S. A. (2024). Harnessing AI for peer-to-peer learning support: Insights from a bibliometric analysis. Perspectives in Education, 42(4). https://doi.org/10.38140/pie.v42i4.8431
Sichterman, B., Noroozi, O., & Boetje, J. (2025). Supporting peer learning with artificial intelligence: A systematic literature review. Innovations in Education and Teaching International, 1648-1664. https://doi.org/10.1080/14703297.2025.2530118
Morris, C., & Maes, P. (2026). When peers outperform AI, and when they do not: Interaction quality over modality. arXiv preprint. https://arxiv.org/abs/2601.11777
Bacalso, F. D., Villamor, F. E., Paclipan, C. P., Orquia, C. J. B., Rendon, I. P., Albiso, A. T., et al. (2026). Institutional approaches to generative AI management in higher education: A systematic review. Frontiers in Education, 11, 1814426. https://doi.org/10.3389/feduc.2026.1814426
Bittle, K., & El-Gayar, O. (2025). Generative AI and academic integrity in higher education: A systematic review and research agenda. Information, 16(4), 296. https://doi.org/10.3390/info16040296
Bloom, B. S. (1984). The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13(6), 4-16. https://doi.org/10.3102/0013189X013006004
Burns, M. (2026). What the research shows about generative AI in tutoring. Brookings Institution. https://www.brookings.edu/articles/what-the-research-shows-about-generative-ai-in-tutoring/
Burns, M., Winthrop, R., Luther, N., Venetis, E., & Karim, R. (2026). A new direction for students in an AI world: Prosper, prepare, protect. Brookings Institution, Center for Universal Education. https://www.brookings.edu/wp-content/uploads/2026/01/A-New-Direction-for-Students-in-an-AI-World-FULL-REPORT.pdf
Cope, B., & Kalantzis, M. (2009). “Multiliteracies”: New literacies, new learning. Pedagogies: An International Journal, 4(3), 164-195. https://doi.org/10.1080/15544800903076044
Cormier, D. (2008). Rhizomatic education: Community as curriculum. Innovate: Journal of Online Education, 4(5), Article 2. https://nsuworks.nova.edu/innovate/vol4/iss5/2/
Fricker, M. (2007). Epistemic injustice: Power and the ethics of knowing. Oxford University Press. https://doi.org/10.1093/acprof:oso/9780198237907.001.0001
Hon, K. (2026). Generative AI in higher education: A systematic review of its effects on learning outcomes and academic performance. Journal of Educational Technology Systems, 54(3), 537-560. https://doi.org/10.1177/00472395251400089
Ikram, M., Hanefar, S. B. M., Saleem, S. M. U., & Zulfiqar, F. (2026). Artificial intelligence in education: A systematic review of personalized learning trends and future directions. Frontiers in Education, 11, 1782626. https://doi.org/10.3389/feduc.2026.1782626
Kamenetz, A. (2013, December 31). The online education revolution drifts off course. NPR. https://www.npr.org/2013/12/31/258420151/the-online-education-revolution-drifts-off-course
Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15(1), 17458. https://doi.org/10.1038/s41598-025-97652-6
Kim, J., & Lee, S.-S. (2022). Are two heads better than one? The effect of student-AI collaboration on students’ learning task performance. TechTrends, 67(2), 365-375. https://doi.org/10.1007/s11528-022-00788-9
Kumar, M., & Babu, A. (2026). A systematic review and meta-analysis on the use of artificial intelligence in education: Trends, impacts, and future directions. Innovare Journal of Education, 14(4), 18-23. https://doi.org/10.22159/ijoe.2026v14i4.59916
NAFSA: Association of International Educators, & JB International. (2026). Fall 2026 international student enrollment outlook and economic impact. https://www.nafsa.org/sites/default/files/media/document/Fall_2026_International_Student_Economic_Snapshot_NAFSA_JBI.pdf
NAFSA: Association of International Educators. (2025). Fall 2025 international student enrollment snapshot and economic impact. https://www.nafsa.org/fall-2025-international-student-enrollment-snapshot-economic-impact
National Education Policy Center. (2026). NEPC review: A new direction for students in an AI world. https://nepc.colorado.edu/sites/default/files/reviews/NR%20Penuel.pdf
Nickow, A., Oreopoulos, P., & Quan, V. (2020). The impressive effects of tutoring on preK-12 learning: A systematic review and meta-analysis of the experimental evidence (NBER Working Paper No. 27476). National Bureau of Economic Research. https://doi.org/10.3386/w27476
Nietzel, M. T. (2026, August 12). Projected loss of international students could bring a $3.4 billion hit. Forbes. https://www.forbes.com/sites/michaeltnietzel/2026/08/12/projected-loss-of-international-students-could-bring-a-34-billion-hit/
OECD. (2026a). Digital Education Outlook 2026. OECD Publishing. https://www.oecd.org/en/publications/oecd-digital-education-outlook-2026_e0e45a57-en.html
OECD. (2026b). Policies supporting responsible and systematic GenAI adoption in higher education. OECD Publishing.
Poquet, O. (n.d.). Social network analysis of TGLF learning environments [Unpublished analysis]. Cited in and summarised by the Structured Peer Praxis draft (2026).
PwC. (2025). The fearless future: 2025 global AI jobs barometer. https://www.pwc.com/gx/en/issues/artificial-intelligence/job-barometer/2025/report.pdf
PwC. (2026). 2026 global AI jobs barometer: Global report findings. https://www.pwc.com/gx/en/issues/artificial-intelligence/job-barometer/2026/2026-global-ai-jobs-barometer-global-findings.pdf
Qian, Y. (2025). Pedagogical applications of generative AI in higher education: A systematic review of the field. TechTrends, 69(5), 1105-1120. https://doi.org/10.1007/s11528-025-01100-1
Reich, J. (2020). Failure to disrupt: Why technology alone cannot transform education. Harvard University Press.
Reich, J., & Ruipérez-Valiente, J. A. (2019). The MOOC pivot. Science, 363(6423), 130-131. https://doi.org/10.1126/science.aav7958
Sadki, R. (2023, October 19). What learning science underpins peer learning for global health? https://redasadki.me/2023/10/19/what-learning-science-underpins-peer-learning-for-global-health/
Sadki, R. (2025a, April 6). Why YouTube is obsolete: From linear video content consumption to AI-mediated multimodal knowledge production. https://redasadki.me/2025/04/06/why-youtube-is-obsolete-from-linear-video-content-to-ai-mediated-multimodal-knowledge/
Sadki, R. (2025b, July 16). Eric Schmidt’s San Francisco Consensus about the impact of artificial intelligence. https://redasadki.me/2025/07/16/eric-schmidts-san-francisco-consensus-about-the-impact-of-artificial-intelligence/
Sadki, R. (2025c). Why peer learning is critical to survive the Age of Artificial Intelligence. https://doi.org/10.59350/redasadki.21123
Sadki, R. (2025d). A global health framework for artificial intelligence as co-worker to support networked learning and local action. https://doi.org/10.59350/gr56c-cdd51
Sadki, R. (2025e). Artificial intelligence, accountability, and authenticity: Knowledge production in the age of generative AI. https://doi.org/10.59350/w1ydf-gd85
Sadki, R. (2025f). The business of artificial intelligence and the equity challenge. https://doi.org/10.59350/redasadki.20984
Sadki, R. (2025g). What the 2025 State of AI Report means for global health and humanitarian action. https://doi.org/10.59350/dpjw3-vgp93
Sadki, R. (2025h). The great unlearning: Notes on the Empower Learners for the Age of AI conference. https://doi.org/10.59350/859ed-e8148
Sadki, R. (2025i, November 11). The future of work: Remarks at the 9th 1M1B Impact Summit held at the United Nations in Geneva. https://redasadki.me/2025/11/11/the-future-of-work-remarks-at-the-united-nations-in-geneva/
Sadki, R. (2026g, January 31). 5 reasons why our current systems of learning are broken, and how to fix them. https://redasadki.me/2026/01/31/5-reasons-why-our-current-systems-of-learning-are-broken-and-how-to-fix-them/
Sadki, R. (2026h, March 24). OECD Digital Education Outlook 2026: How can AI help human beings learn and grow? https://redasadki.me/2026/03/24/oecd-digital-education-outlook-2026-how-can-ai-help-human-beings-learn-and-grow/
Sadki, R. (2026i, May 6). When we get health wrong, people die: Designing artificial intelligence to serve community health. https://redasadki.me/2026/05/06/when-we-get-health-wrong-people-die-designing-artificial-intelligence-to-serve-community-health/
Sadki, R. (2026j). The integrity layer: Using assessment to scaffold peer learning at scale. https://doi.org/10.59350/ympfp-gzg71
Sadki, R. (2026k). Artificial intelligence for global health: A tragedy of the commons brewing in Geneva’s basement. https://doi.org/10.59350/zcnvp-yc278
Sadki, R. (2026l, August 14). Health workforce development in the Age of Intelligence: A tragedy of the cognitive commons. https://redasadki.me/2026/08/14/health-workforce-development-in-the-age-of-intelligence-a-tragedy-of-the-cognitive-commons/
Schön, D. A. (1983). The reflective practitioner: How professionals think in action. Basic Books.
Scott, J. C. (1998). Seeing like a state: How certain schemes to improve the human condition have failed. Yale University Press.
Sen, S. (2026). Why did China just junk 12,000 degree courses? They were “obsolete.” ThePrint. https://theprint.in/feature/china-just-junk-12000-degree-courses-obsolete/2960209/
Siemens, G. (2005). Connectivism: A learning theory for the digital age. International Journal of Instructional Technology and Distance Learning, 2(1). https://web.archive.org/web/2020/http://www.itdl.org/Journal/Jan_05/article01.htm
Slimi, Z. (2026). A systematic critical review of generative AI’s impact on authorship, pedagogy, and integrity, 2023-2025. Frontiers in Education, 11, 1769680. https://doi.org/10.3389/feduc.2026.1769680
Sparks, S. D. (2022, December 8). The head of the Carnegie Foundation wants to ditch the Carnegie unit. Here’s why. Education Week. https://www.edweek.org/teaching-learning/the-head-of-the-carnegie-foundation-wants-to-ditch-the-carnegie-unit-heres-why/2022/12
Thomas, D., Lin, Y., Gatz, M., et al. (2024). Improving student learning with hybrid human-AI tutoring: A three-study quasi-experimental investigation. In Proceedings of the 14th Learning Analytics and Knowledge Conference (LAK ’24). https://doi.org/10.1145/3636555.3636896
Times Higher Education. (2013, November 28). MOOCs: edX unveils Arab initiative, Udacity rethinks “lousy product.” https://www.timeshighereducation.com/news/moocs-edx-unveils-arab-initiative-udacity-rethinks-lousy-product/2009277.article
Upadhayaya, K. K. (2026). Artificial intelligence in education: Opportunities, risks, and pedagogical implications for learning and assessment. Advances in Mobile Learning Educational Research, 6(2), 1833-1844. https://doi.org/10.25082/AMLER.2026.02.001
Watkins, K. E., & Marsick, V. J. (1990). Informal and incidental learning in the workplace. Routledge.
Watters, E., Maxham, M., & Baron, P. (2026). The use of artificial intelligence in higher education: A systematic literature review of learning effectiveness and ethical issues. IGI in Education Insight, 1(2), 79-86. https://doi.org/10.53905/edu.v1i02.10
Wells, C. (2025, November 30). Fewer international students are enrolling at U.S. colleges, which could cost the country $1 billion, reports find. CNBC. https://www.cnbc.com/2025/11/30/international-student-enrollment-decline.html
Wu, F., Dang, H., Li, Y., et al. (2025). A systematic review of responses, attitudes, and utilization behaviors on generative AI for teaching and learning in higher education. Behavioral Sciences, 15(4), 467. https://doi.org/10.3390/bs15040467