Microsoft's Study and Learn Whitepaper Review

Microsoft’s learning agent mistakes its prompt for an architecture

TL;DR Microsoft built an AI tutor with good learning science behind it, then shipped it in a way that doesn’t quite hold up. I’d know, I spent five years building this stuff at Microsoft and red-teamed this one for my university. The teaching ideas are solid, but they live in a prompt rather than the architecture, so the agent drifts the moment a student stops cooperating (a few polite “make this stronger” requests and it’ll write the essay). It works best for confident self-starters and struggles with the younger students (13-17) it most wants to help. My three core arguments:

1. Don’t pitch Study and Learn to children still building the judgement to use it well. Pitch the platform underneath it, so universities (and maybe some later 7-12) can give students richer experiences at scale. To win higher ed, Microsoft needs to help us safely ground agents in our facilities, knowledge, people, research and culture.

2. Chat-based study agents (or whatever we’re calling this trend) have a short shelf life. There’s far more you can do with AI and good data, and I hope Microsoft’s roadmap goes beyond prompting.

3. Microsoft has a rare chance to partner with universities on the future of teaching. But they don’t seem to see us as enterprise customers with the same value as the “E” SKU crowd, even though we’re literal cities with huge headcounts, complex systems and real problems.

The Microsoft Study and Learn white paper calls its four pillars architectural constraints. In reality though, they are instructions to a model whose default is to finish the task, and this difference surfaces the moment a student stops cooperating with the demonstration. I red-teamed the shipped agent, and the university I work with switched it off for every student. We are slowly opening it back up again, but only to select cohorts of users.

For transparency, I have spent more than 20 years between classrooms and the companies that build for them. I taught secondary in Western Sydney, I have tutored pre-service teachers at Western Sydney University since 2010, and for five years I was a product engineering lead for Microsoft’s education business across the Asia-Pacific, working on the tools that put AI in front of teachers and students. I now build memory and orchestration systems for AI agents, and I red-teamed Microsoft’s Study and Learn agent after it shipped, which is how the university I work with came to switch it off for every student (that deployment story is its own companion piece). Those vantages turn on two distinctions, and the white paper behind Study and Learn (Padte, Sarkar, Sparvell, & Jabbour, 2026) still trips up on both. The engineering one is the difference between a behaviour a model can be prompted into and a capacity the system holds. The teaching one is the difference between a lesson that looks like learning and one that produces it. The paper calls the agent’s four pedagogical pillars architectural constraints, evaluated in every interaction. That phrase fails the first distinction, and most of what follows from it fails the second.

A prompt is not an architecture

Let’s start with what the word architecture is being asked to carry. In a cognitive architecture, in the sense Sweller or Newell would use it, a constraint is a property of the system. The limits of working memory, inhibition of return, the forgetting curve. They operate whether the system wants them to or not, because they are how the thing is built. The agent’s pillars are not like that (once you take apart how it works). “Ask one question, withhold the answer, keep guiding even when the learner asks you to stop” are sentences written into a prompt and graded by an evaluation rubric, sitting on top of a base model whose trained-in disposition is to complete the task in front of it. Calling that architectural borrows the authority of the word while inverting what it really means.

This is the anthropomorphism move, and I have written before about why it is never harmless, because how you describe a system decides which problems you think you have. Believe the model remembers and you treat a context-window limit as forgetting, to be fixed by storing more. Believe it processes and you treat the same limit as information management, to be fixed by curating what reaches it. Believe its coaching behaviour is architecture and you treat a lapse as a tuning bug, to be fixed by another line in the prompt. The paper’s own development history is a record of doing the third thing. Nine rounds of evaluation and a lengthening list of constraints (ask one focused question, keep explanations brief, do not answer your own question, keep guiding when the learner tries to bypass the questions), each added because the model kept reverting to completion. It reads like the maintenance log of a leash, and that’s not a criticism, as anyone working on agents today has been there.

The competence is the case against them

The pedagogy underneath is real, and worth being clear about, because it sets up everything that follows below. The pillars map to constructs about as solid as anything in the field, scaffolding and the zone of proximal development (Wood, Bruner, & Ross, 1976), the generation effect (Slamecka & Graf, 1978), retrieval practice (Roediger & Karpicke, 2006). The strongest evidence for the approach landed after the design was fixed, when Kestin and colleagues (2025) ran a randomised trial at Harvard and a carefully scaffolded GPT-4 tutor doubled the learning gains of in-class active learning. Microsoft did not stumble into this, let me be clear. They know the science, and the build reflects it.

That is, however, what makes it an indictment of the launch, if not the intent. A team this fluent in the science did not fail to notice that an agent designed to withhold answers would be asked, repeatedly and inventively, to stop withholding. It went out anyway, to a product an institution can switch on for learners as young as 13, with the obvious vectors open (much of which the team has since moved to close, and I will come to that). When the craft is this good, the gaps at launch read less like the limits of the art and more like decisions about what to close first and what to leave for the institutions to find. And I get it, we should be treating the current version of Study and Learn as the first step in a much longer roadmap.

Learning science, used the way a pitch deck uses studies

A fair word before the hard part, and to be clear I am not intending to debunk. On the contrary I want to support this project and genuinely believe it could be awesome. This is a white paper, a piece of product marketing with a bibliography, and it was never going to be held to the standard of a peer-reviewed article. It should not be. Microsoft also put it into the world knowing that education academics and cognitive scientists would read it closely, because a document that makes claims on our turf gets read closely by the people whose turf it is. That is the reading I am giving it, although I’m doing it mostly for fun at this point.

The paper wears its citations well, and the long reference list is doing a lot of the persuading. Read a few of the numbers closely and they carry more confidence than the literature actually bears (and I’ve read the stats and numbers of each paper in the bibliography).

Take John Hattie, whose figures run through the paper. Hattie ranks teaching interventions by effect size, which is a way of putting different studies on one ruler. The common unit, Cohen’s d, measures a difference in standard deviations, where roughly 0.2 counts as small, 0.5 as moderate, and 0.8 as large. Hattie reports feedback an effect size of 0.70, large and dependable-looking, and draws a hinge point at 0.40, which he treats as the average effect of the things schools already do, the line below which an intervention is not worth the bother. The paper takes these as settled measurement. But the reality (like all research) is that they are quite a bit shakier than what appears on the surface. Simpson (2017) showed that the same intervention can score a large or a small effect size depending on which test you use and how spread out the students’ scores are. A narrow test on similar students inflates the number, a broad test on a mixed group shrinks it, so stacking these figures into a league table, which is what Hattie does, means comparing numbers that were never on the same scale. Kraft (2020) added that the 0.40 bar sits on a biased pile, because studies that find a big effect get published while studies that find nothing sit in a drawer, which pushes the published average up and flatters everything measured against it. Welcome to the academy my friends! Subtle plug, I just wrote a paper on the frontier researcher. See if you can guess which friend I am talking about.

Productive struggle, the load-bearing pillar, rests on an effect that is real and in a word “modest”. Sinha and Kapur’s (2021) meta-analysis put productive failure at an effect size of 0.36, a small effect, rising to about 0.58, a moderate one, only when the design is close to perfect (and it nearly never is). And they explicitly named the boundary conditions in the same breath. Secondary-age and older, mostly STEM, dependent on collaboration and on contrasting the students’ own attempts against the expert solution. A one-to-one agent has no peers and no rival attempts to contrast, its worked examples are basic algebra, and the effect keeps failing to replicate in the younger children who make up most of the Microsoft K-12 target. I know a guy who would happily run some efficacy studies here. Seriously, reach out!

Bloom’s Two Sigma anchors the framing, as it anchors every tutoring pitch on the planet. The name comes from Bloom’s 1984 claim that one-to-one tutoring moved the average student two standard deviations ahead of ordinary classroom teaching, which would be enormous, roughly the distance between an average student and the top few per cent. It rests on two small dissertation studies run under an unusually strict form of mastery learning, and it is widely judged “inflated”. Realistic tutoring effects land nearer a fifth to four-fifths of a standard deviation (VanLehn, 2011), worthwhile, and nowhere near the headline we need for a shiny pitch deck. The case for the offloading risk then opens on Gerlich (2025), a single snapshot survey. It can show that heavy AI users report weaker critical thinking, but a snapshot cannot tell you which way the arrow runs, whether AI is eroding thinking or weaker thinkers reach for AI more in the first place.

None of this is fabricated. It is selected and dressed, the way any good pitch deck selects the studies that flatter the product. I know this, because I have done it more times than I can count. The paper is candid that feasibility drove the selection, that it picked the principles current AI can carry out and set aside the ones that matter more. Which means the learning science here is downstream of what the model can already do. The principles were chosen to fit the tool, then dressed in citations to make the fit look like a philosophy, and a long bibliography lends that an authority it has not actually earnt.

The score and the capacity

Here the philosophy of mind earns its keep, and this stays one of my favourite areas to think in. First, a definition I lean on throughout. I talk a lot about the “kernel.” By the kernel I mean the language model itself, the reasoning core at the centre of the AI ecosystem. In a companion piece I am currently writing, I argue it is only one part of a working agent, the inference engine, with memory, retrieval and orchestration built around it as a separate architecture. That matters here because the gaps I am about to describe live in the kernel’s reasoning, which means better prompts and richer memory can route around them but cannot close them.

There is a standing problem in evaluating these systems, the gap between a high score on a task and a robust capacity behind it. A model can post a strong number for two different reasons. It might have grasped the underlying structure of the problem, or it might have become an excellent pattern-matcher to the kind of question the test tends to ask. You can only tell the two apart under perturbation. Perturbation means changing the surface of a problem while keeping its logic identical, then checking whether performance holds. Reword the question, swap the numbers, flip the order of the options. A system that understood the structure keeps going, and one that memorised the pattern falls over. Binz and Schulz (2023) built their study around that test. They took classic experiments from cognitive psychology, the kind designed to probe human reasoning, and ran GPT-3 through them like a participant. It looked strong until they altered the surface details, and then it broke in three very telling ways. No directed exploration, a failed causal-reasoning task, and small changes to a solved problem sending it wrong. The fair objection is that GPT-3 is old and that today’s frontier reasoning models score better on these benchmarks. That objection misses the point of the test entirely. A higher score under friendly conditions, where the questions are phrased the way the model expects, tells you the model handles the expected case, and it says nothing about the case where a real student phrases things oddly or pushes back. That surviving-property is robustness, and you only find out whether a system has it once it is deployed and someone stops cooperating (like some idiot that likes to red-team everything).

Bring that to the agent, and the jobs a tutor cannot do without turn out to be the jobs the “kernel” is weakest at. In each one the agent can perform the behaviour while the capacity underneath it is entirely missing.

Start with diagnosis. To pitch help inside a learner’s zone of proximal development, the agent has to work out what this student knows, what they do not, and why they made the mistake they made. That is reading another person’s state of mind, and it sits right next to the directed exploration the kernel was shown to lack. When the mistake is a common one, say dividing before isolating the x in x + 5 = 9, the model has seen a million versions of it, and matching that pattern looks like diagnosis and usually works. When the confusion is unusual, something the training data did not cover well, the same machinery returns a confident but wrong reading of what the student is thinking. Every worked example in the paper is the common kind. The unusual cases, the ones that would separate real diagnosis from a good guess, are not shown.

Then calibration. Knowing when a struggle has tipped from useful into pointless means reading the student and weighing what happens if you hold help back a moment longer. The model has no internal sense of its own uncertainty to read from. This is the weak point at the centre of every agent that plans, acts and revises in a loop, because if the step meant to judge whether an answer is good cannot tell good from bad, the whole loop produces confident nonsense. The product’s own users found this exact seam (I define seam design as part of Frontier Operations. if you are interested in more read here). On a five-point scale, the two lowest-rated items were “right level of support” at 3.65 and “right length of responses” at 3.56, the two that need the calibration the kernel cannot supply. Koedinger and Aleven named this the assistance dilemma back in 2007, and it is still very much an open problem. The paper treats it as something you tune, and I doubt it tunes away, because the right amount of help differs for every student and every moment and depends on reading understanding in real time. Tuning fixes one policy in place. The dilemma needs a fresh judgment each time, of a kind the system cannot reliably make.

Then transfer. The fourth pillar asks the agent to hand the student a fresh problem in a new context and see whether the understanding carried over. Transfer is using what you learnt in one place to solve something somewhere else. Near transfer is a problem that looks similar. Far transfer is spotting the same deep structure inside something that looks nothing alike, and it is the hard one to try and solve for, because it only works if the student grasped the structure rather than the surface. Far transfer is also the operation Binz and Schulz watched break, since a small change of surface was enough to throw the model off. The agent will produce something it calls a transfer task. Whether that task probes the same underlying structure or only redresses the same surface in new words is the capacity question, and calling it a transfer task does not actually settle it. A system that matches surfaces will tend to generate a new surface.

Then there is the test the evaluation never ran. Microsoft scored the agent across multi-turn conversations, which is the right unit for coaching, and used it to ask whether the agent coaches well while the learner plays along. That is a measure of the cooperative score. When I red-teamed the shipped product I ran the other measure. A four-turn editing cascade, every step framed as make-this-sentence-stronger, walks a student out with a 1,000-word essay while no single turn reads as ghostwriting, because the agent judges each turn alone and the constraints only defer the drive to complete one prompt at a time. That is the perturbation result. The cooperative score was high and the capacity to hold the line was not there, which is the gap the philosophy predicts and the evaluation was built not to see. The Digital Safety Board signed off on the framework. The framework measured the wrong thing.

The design needs the learner who needs it least

In a meeting late last week, the ever-sharp Professor Simon Buckingham Shum said something that frames this whole section better than I can, and I am about to paraphrase it terribly. His point was that higher education exists to produce students who can self-regulate and make sound judgements about how and where they use AI, so the fact that a student can jailbreak a study agent should worry us less than the judgement, ethics and disciplinary grounding they bring to using it. That lands close to what I take the purpose of higher education to be, turning out graduates who are frontier operators in their chosen discipline.

It also draws the line this section is about. For a self-regulating adult, a jailbreak is the smaller worry, because the learner brings the judgement the tool lacks. For a 13-year-old who has not built that judgement yet, the jailbreak is the whole worry, because there is nothing on the student’s side to catch it. The same agent needs a different answer at different stages, and calibrating that answer to the stage is the judgement the launch skipped. It is the decision I reached when I red-teamed this one for my university.

Leave the kernel aside for a moment, because the teaching design carries a separate problem that has nothing to do with the model underneath. Productive struggle only helps under one condition, and the paper hurries past it (for those that don’t know I write a LOT about productive struggle and even align it to ancient philosophy, so this was an easy target). A difficulty is useful only when the learner can get through it, and Bjork and Bjork (2011), whose framework the paper leans on, say so plainly. Below that line, a difficulty meant to be desirable becomes plain difficulty, and it produces failure and quiet demoralisation. Whether a student clears that line depends on what they already know, and the same knowledge decides whether holding the answer back helps them or gets in their way. The expertise reversal effect (Kalyuga, Ayres, Chandler, & Sweller, 2003; Kalyuga, 2007) is the settled version of this. Heavy scaffolding lightens the load for a novice, and the same scaffolding turns into noise for someone who already holds the concept and now has to wade through help they do not need. Refuse-the-answer-and-question-them-toward-it is right for the student with enough grounding to make the work pay, wrong for the one without it, and wrong again for the one who only needed the fact. I read this section as a teacher more than an engineer, because timing that withholding move is most of the craft, and getting the timing wrong is how you turn a struggling student off a subject, sometimes for good. The paper hands that timing to a system that cannot read the student well enough to get it right, and its worked examples are all the easy cases where the timing hardly matters. I’d love to get some feedback from Prof Jose Hanham at ACU on this.

The control the agent hands over makes the same bet. Tell me what you tried, choose your next step, push through the struggle, all of it assumes a student who can watch their own understanding and manage their own effort. The students worst at that are the ones who most need it, because those with the least prior knowledge tend to overestimate how much they understand (Hacker, Bol, Horgan, & Rakow, 2000), and the ability to self-regulate is still forming across the ages the product targets (Weil et al., 2013; Zimmerman, 2002). The design works best for a student with solid foundations and developed self-regulation, and worst for one who is younger, behind, and most tempted to let the machine carry the load. (This is another reason I keep telling Microsoft to focus more on adults and higher education, among a long list of others.) The second student is the one the paper keeps saying it wants to reach, the one who never had a tutor. Across a room of 30 I can tell within a few days which students thrive when you hand them the wheel and which ones quietly disengage, and honestly it tracks prior attainment pretty closely. The tool is built for the first group and pointed at the second. The teaching is hardest to sustain for exactly the child it says it wants to rescue.

Underneath all of this sits a sequencing question the paper never asks. Geary’s (2008) distinction matters here, between biologically primary knowledge, the things humans pick up almost without trying such as speaking a first language, and biologically secondary knowledge, the literacy, numeracy and formal reasoning that schooling exists to build through effortful practice inside a formative window. The paper’s promise assumes a student who already holds a base of that secondary knowledge to struggle with and build from. A student cannot offload knowledge they never built, and the agent is most available across the same years that knowledge is supposed to be forming (I have written about my own journey with this elsewhere). The paper treats one design as fit for the whole of K-12, from a child still learning to sound out words to a near-adult, and never asks at what stage, and after which foundations are in place, an always-on AI turns into a net good rather than a shortcut around the work that builds the foundation.

And then the part that is not about knowledge at all. Vygotsky’s zone of proximal development, which the paper cites approvingly, is supposed to be social, where the more knowledgeable other supplies the regulation, the learner gradually takes it inside and makes it their own, and the scaffold is meant to come off. Every teacher learns to engineer their own absence, to fade the help on purpose, because the end state is a student who no longer needs them (this is why teachers so often make great managers and leaders, by the way). A tutor available at any hour, endlessly patient, never has to fade, and it is the fading that forces the regulation to move from the helper into the learner. Darvishi, Khosravi, Sadiq, Gašević, and Siemens (2024) measured the failure directly. Take the AI help away and the students who had leaned on it showed less agency, and adding self-regulation prompts alongside the AI did not restore their independence. The regulation had stayed outside them. Kosmyna and colleagues (2025) found the same shape in writing, where students who wrote essays with an LLM showed the weakest brain engagement and the least ownership of work they had supposedly produced themselves. The paper’s own evaluation never takes the agent away to see what remains, which is the one test that would show whether the gradual release lands in the student or in a dependence on the tool.

None of this sits at the fringe either. In January 2026 the Brookings Center for Universal Education’s Global Task Force, after a year of focus groups across 50 countries and a review of hundreds of studies, concluded that the risks of generative AI in education currently outweigh the benefits, that it can undermine children’s foundational development, and described a dependence loop of offloaded thinking, one contributor calling the current crop of tools the “fast food of education”. They also named the real prize, reaching learners shut out of school altogether, which is the prize Microsoft keeps pointing at too. Both want a tutor for every child, and they split completely on whether today’s tools are fit to be that tutor in the formative years. The paper does not engage the side that says no, and engaging it would be worth Microsoft’s time. To be clear, I am not against AI. I am passionate about education, and I hold firm views on how AI should be tapered, introduced and embedded across the different stages of development and schooling, which is a more careful question than either the cheerleaders or the refusers tend to allow.

Credit where it is due

I published the red-team, and then something happened that does not always happen. The team behind Study and Learn reached out, and we have since spent real time together. They engaged with the substance, granted the parts that landed, and were straight about what was already in train. That is worth saying plainly, because it is rare, and because the people building this clearly care about getting it right for students and teachers.

Several of the concerns in this piece are fixed or being fixed. The agent reaches learners aged 13 to 17 only after an institution’s IT explicitly turns it on, with adults getting access by default, so the youngest users sit behind an institutional decision rather than a global switch. The specific jailbreak vectors I reported, including the editing cascade, are being hardened, the evaluation set has been expanded to cover them, and a new version will be rolling out. Microsoft’s compliance tooling already flags distressed and inappropriate conversations for administrators, and that coverage is growing. Curricular grounding and institution-level controls are on the roadmap, which is the change that would matter most to the universities I work with.

Frontier operations

Strip all the vocabulary away and the naked shape is familiar to anyone who has tried to put a frontier model into production. The system is sold on polished demonstrations run under friendly conditions and backed by a wall of evaluation, and then it meets real students, who do not play along. They do not want the lesson, they want the answer, and they will spend 20 turns getting it (my cascade needed four) or more likely just go somewhere else that will provide it without friction. Evaluation that scores the friendly case, however many turns it runs, is measuring the demonstration. You find out what the system can do in operation, from the student who pushes on it, and this paper is built around the demonstration.

Wellbeing is the same shape of problem at higher stakes. A student talking to an always-available agent will raise stress, sleep loss, family pressure, and in some cases self-harm. Microsoft’s compliance tooling already flags distressed and inappropriate conversations for administrators, and that coverage is expanding, which is the right instinct. The harder part is what happens for the student in the moment, a warm handoff to the institution’s own counselling and support services rather than an alert that reaches an admin after the conversation has moved on. That routing is the piece still to build, and going quiet is not an acceptable fallback when a minor is in distress.

What redeployment would take I set out in full in the companion piece, and it comes down to one move. Take the things the kernel cannot hold steady in its head, what the learner knows, which assessment is open and what it permits, how much the student has already been given across the conversation, the institution’s guardrails and support services, and put them where the system maintains them across turns, instead of where the prompt has to re-guess them on every message. I don’t necessarily think however, that any of this is how a future “post-lms” study or tutor agent could or possibly should be. I genuinely think the “chat to learn” fad has a very limited shelf life, especially when AI offers far more possibilities once you realise that the LLM is just the kernel.

Sure an AI can tutor, and a scaffolded model under the right constraints helps, and the students with no other help are the ones it could matter for most. Making classroom technology work is what I have spent 20 years doing, on both sides of the procurement line, and that is the ground my scepticism stands on. It is also why we have started letting the agent back in, for a few carefully chosen cohorts rather than every student at once, which is the tapered, stage-aware approach the launch skipped. My verdict is narrower and harder than the slogans on either side. Microsoft built a competent simulation of a tutor, called the simulation an architecture, evaluated it against the students who cooperate, and shipped it to the ones who will not. That was the launch. The version they are building now, with the youngest users gated, the vectors closing, and universities in the room, is a more interesting conversation, and one I am glad to be having with them.

References

Binz, M., & Schulz, E. (2023). Using cognitive psychology to understand GPT-3. Proceedings of the National Academy of Sciences, 120(6), e2218523120. https://doi.org/10.1073/pnas.2218523120

Bjork, E. L., & Bjork, R. A. (2011). Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning. In M. A. Gernsbacher, R. W. Pew, L. M. Hough, & J. R. Pomerantz (Eds.), Psychology and the real world: Essays illustrating fundamental contributions to society (pp. 56–64). Worth Publishers.

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

Brookings Global Task Force on AI in Education. (2026). A new direction for students in an AI world: Prosper, prepare, protect. Center for Universal Education, Brookings Institution.

Darvishi, A., Khosravi, H., Sadiq, S., Gašević, D., & Siemens, G. (2024). Impact of AI assistance on student agency. Computers & Education, 210, 104967. https://doi.org/10.1016/j.compedu.2023.104967

Geary, D. C. (2008). An evolutionarily informed education science. Educational Psychologist, 43(4), 179–195. https://doi.org/10.1080/00461520802392133

Gerlich, M. (2025). AI tools in society: Impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006

Hacker, D. J., Bol, L., Horgan, D. D., & Rakow, E. A. (2000). Test prediction and performance in a classroom context. Journal of Educational Psychology, 92(1), 160–170. https://doi.org/10.1037/0022-0663.92.1.160

Hanham, J., & White, M. (in preparation). Human cognitive architecture in the age of generative AI.

Hattie, J. (2009). Visible learning: A synthesis of over 800 meta-analyses relating to achievement. Routledge.

Kalyuga, S. (2007). Expertise reversal effect and its implications for learner-tailored instruction. Educational Psychology Review, 19(4), 509–539. https://doi.org/10.1007/s10648-007-9054-3

Kalyuga, S., Ayres, P., Chandler, P., & Sweller, J. (2003). The expertise reversal effect. Educational Psychologist, 38(1), 23–31. https://doi.org/10.1207/S15326985EP3801_4

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, Article 17458. https://doi.org/10.1038/s41598-025-97652-6

Koedinger, K. R., & Aleven, V. (2007). Exploring the assistance dilemma in experiments with cognitive tutors. Educational Psychology Review, 19(3), 239–264. https://doi.org/10.1007/s10648-007-9049-0

Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task (arXiv:2506.08872). arXiv. https://arxiv.org/abs/2506.08872

Kraft, M. A. (2020). Interpreting effect sizes of education interventions. Educational Researcher, 49(4), 241–253. https://doi.org/10.3102/0013189X20912798

Padte, A., Sarkar, A., Sparvell, M., & Jabbour, M. (2026). Learning by design: Learning science foundation of the Study and Learn Agent (Version 1.0). Microsoft.

Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255. https://doi.org/10.1111/j.1467-9280.2006.01693.x

Simpson, A. (2017). The misdirection of public policy: Comparing and combining standardised effect sizes. Journal of Education Policy, 32(4), 450–466. https://doi.org/10.1080/02680939.2017.1280183

Sinha, T., & Kapur, M. (2021). When problem solving followed by instruction works: Evidence for productive failure. Review of Educational Research, 91(5), 761–798. https://doi.org/10.3102/00346543211019105

Slamecka, N. J., & Graf, P. (1978). The generation effect: Delineation of a phenomenon. Journal of Experimental Psychology: Human Learning and Memory, 4(6), 592–604. https://doi.org/10.1037/0278-7393.4.6.592

VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221. https://doi.org/10.1080/00461520.2011.611369

Vygotsky, L. S. (1978). Mind in society: The development of higher psychological processes. Harvard University Press.

Weil, L. G., Fleming, S. M., Dumontheil, I., Kilford, E. J., Weil, R. S., Rees, G., Dolan, R. J., & Blakemore, S.-J. (2013). The development of metacognitive ability in adolescence. Consciousness and Cognition, 22(1), 264–271. https://doi.org/10.1016/j.concog.2013.01.004

Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of Child Psychology and Psychiatry, 17(2), 89–100. https://doi.org/10.1111/j.1469-7610.1976.tb00381.x

Zimmerman, B. J. (2002). Becoming a self-regulated learner: An overview. Theory Into Practice, 41(2), 64–70. https://doi.org/10.1207/s15430421tip4102_2