If you teach and you read the research on generative AI, you have probably noticed that it gives you two sets of instructions. One set of studies reports that language models personalize instruction and improve performance. Another reports cognitive offloading, weakened memory traces, eroded critical thinking, and what Fan and colleagues called “metacognitive laziness.”
A teacher reading current research in good faith is told both to adopt the technology and to restrict it. I have started calling this the Practitioner’s Paradox, and I think it is the most consequential unresolved problem in educational technology today. That paradox is not a result of flawed research. Instead, it points to a mistake in how the research question is being asked in the first place.
Kestin and colleagues ran a crossover trial in undergraduate physics and found that a purpose-built AI tutor produced larger immediate learning gains than an in-class active-learning lesson. Similarly, Bastani and colleagues ran a field experiment in high school mathematics and found that access to a general GPT interface improved practice performance and then lowered performance on the unaided test that followed. A second tutor that withheld answers and gave teacher-designed hints largely erased that loss. And Fan’s writing study found that ChatGPT improved essay revision scores while producing no advantage in transfer.
All of these results are academically sound. They just do not add up to a single consolidated answer about “the effect of AI on learning.”
I do not think such an answer exists. To address this properly, I have been developing a theory I call Synthetic Constructivism, which grew out of my earlier work on game-based learning. Its main claim is that the contradiction dissolves once you stop asking what the tool does to the learner and start asking what the learner and the tool do together.
An academic paper on the theory has just been accepted for ICERI2026, the 19th International Conference of Education, Research and Innovation, in Seville this November, and I have been invited to give a keynote about it at LLM2026, the International Conference on Large Language Models, in Las Vegas in December. Both of those are targeted at academic researchers and are fairly technical. So in today’s post, I want to lay out the core of that theory in plain terms and then state what it asks of teachers.
What the games debate taught me
Game-based learning spent about twenty years stuck in the same kind of argument. The General Aggression Model (Anderson & Bushman) held that violent games train players in aggressive patterns of thought. The Catalyst Model (Ferguson et al.), on the other hand, held that the medium is a negligible trigger for aggression driven by biology and environment. Neither could move the other, because both assumed a linear media-effects model in which a stimulus produces an effect in a passive recipient.
In a keynote given at the 2011 Clash of Realities conference in Cologne, Germany, I argued that the way out was to drop that model completely and treat a digital game as a closed feedback loop that the player helps constitute. With that framing, learning is a stabilization of action inside the loop rather than an effect transmitted by a medium. And once learning belongs to the loop, contradictory findings stop being anomalies. They are what you should expect from a system that is sensitive to who the learner was before they started to play.
Generative AI belongs to the same class of media, but while a game constrains you with fixed rules, a language model constrains you with probabilities over an open-ended space of text. Every prompt and correction reshapes what comes next. Consequently, expressive power goes up, and so does path dependence. That is why I think the cybernetic frame fits AI-assisted learning even better than it fit games.
You are the one watching the loop
Radical constructivism, in the version developed by Ernst von Glasersfeld, holds that what we call knowledge is a construction, and that the test of a construction is whether it works. Von Glasersfeld’s image for this is a key in a lock. A key that opens a lock tells you nothing about what the lock looks like inside. It only tells you that it fits. Knowledge, on this view, is a key that fits, and there is no way to check it against the lock directly. The constructing itself is unified and internal, the mind watching and revising its own operations mostly below awareness.
When you work with a language model as a learner, that process splits in two. The model executes the symbolic operations of the task. It drafts, summarizes, restructures, and explains. You stand outside that work, directing and monitoring it, and you decide whether to accept what it produced. In the vocabulary of Heinz von Foerster, the model is a first-order cybernetic system and you are the second-order observer regulating it.
Together, the two positions form a feedback loop. Projection is the feed-forward half: you push a mental model or a set of constraints into the system as a prompt. Back-Projection is the feedback half: the output the system returns to you for evaluation. Because the loop is externalized, the thinking that is normally silent becomes visible. You can read your own half-formed idea back off the screen and revise it. I call this an Epistemological Laboratory.
But this is also the trap. The first-order system keeps producing fluent output whether or not anyone is judging it, and a fast run of completion requests just externalizes action while bypassing monitoring entirely. Which raises the question of what makes a good second-order observer, and here a sixty-year-old theorem in cybernetics can provide the answer. Conant and Ashby showed that a regulator can only be as good as its understanding of what it regulates.
In other words, a learner who does not understand how the model fails, or what counts as evidence in the discipline, cannot regulate the loop. Prior knowledge and metacognitive skill are the controls. This means that machine capability is not the binding constraint on educational outcomes. Learner-side model quality is.
Why identical tools produce opposite results
It is that last point that causes the practitioner’s paradox to fall apart. Mathematically speaking, the loop runs in a high-dimensional space of possible text. Systems like that are highly sensitive to the initial conditions the learner supplies. Prior knowledge decides what a student can ask and which errors they can see. Metacognitive habit decides whether they check or accept.
The result is that small differences get amplified across iterations. This is effectively a butterfly effect, applied to learning. And so, two students with the same tool and the same assignment end up on completely different trajectories. The scaffolding literature and the erosion literature are therefore both accurate. They are just describing different states of the same system.
I call the most critical consequence of this the Non-Prompter Disadvantage. An experienced prompt writer projects viable prompts and reads the output critically, so the recursion settles fast onto something useful. A novice struggles to get the loop going at all. Their sessions drift, or they converge early on a fluent output that encodes no durable understanding. And that disadvantage compounds, quickly, because every failed loop deprives the novice of the practice that would have built the competence.
I need to acknowledge that this is easy to misread as a claim about student deficits. It is not. Clerc and colleagues showed that a brief classroom intervention improved middle schoolers’ query reformulation and accuracy judgments in an LLM-supported science task. The gap is a design and instruction problem.
What is skilled prompting, then? A model offers more continuations than anyone can inspect, so regulation depends on constraint. Specifying an audience, an evidence standard, a required counterexample, or the instruction to ask rather than answer removes irrelevant possibilities and steers the output toward what is useful. Prompting is construction by subtraction. But subtraction only narrows the space. It does not guarantee that what is left is true. A badly chosen constraint can amplify a misconception. Constraints organize possibility. Evidence tests viability.
One more distinction has to be mentioned. Salomon, Perkins, and Globerson separated effects with technology from effects of technology. Generative AI produces large effects with the tool, such as faster completion and better prose, while leaving almost no effect of the tool once assistance is removed. Learning Gain, as I define it, is the effect of the tool. It is the improvement in the learner’s own knowledge, assessed under conditions where the model cannot substitute for the target competence.
Learning gain is therefore not the same as the gap between what a student produces with the model and what they can do without it some time later. The gain tells you whether the learner’s own knowledge improved. Unassisted performance shows what they can do alone, while the gap shows how much of the assisted performance remains dependent on the model. A large gap is fine when the goal is augmentation. But it is a serious problem when the stated goal is independent understanding.
Two pathways run on the same loop
It is important to recognize that cognitive offloading is not necessarily a vice. Using a calculator moves work out of the head so that attention can go to the part of the problem that requires thought. The risk only appears when the offloaded operation is the one thing the student was supposed to learn.
The division of cognitive work between learner and model follows one of two pathways. On the agency-enhancing pathway the learner keeps the central operations: defining the problem, selecting evidence, generating and testing explanations, monitoring uncertainty, and deciding what to revise. Peripheral or accessibility-related subtasks go to the model.
On the agency-eroding pathway the core operations are outsourced wholesale, and the learner checks the result for surface acceptability. Monitoring goes shallow because a polished output looks finished, whether it is or not. The risk peaks when model confidence exceeds model reliability and learner confidence exceeds learner expertise.
But nothing in the technology selects between these pathways. It is the second-order observer who does, which is why the same prompt can mean two different things depending on who sends it. Asking for a summary is a convenience for an expert who has already read and judged a text. It is a substitution for comprehension when a novice has not read it at all.
Poquet and colleagues propose how this can be measured. They profile offloading by volume of prompting and by content, meaning who is doing the cognitive work and at what level of comprehension. They then show that volume alone tells you very little. A student can make many short, active queries while refining a concept, or make a single request that delegates an entire assignment. In their study of AI-mediated synthesis writing, passive delegation of whole-text comprehension was the pattern most associated with diminished authorship. This means that a teacher looking at a prompt log should read the prompts for function rather than count them.
Here is where “less AI is better” would be the wrong lesson to draw, and I can speak to that from experience. The game-based theory this whole post builds on is one I had abandoned after that keynote in Cologne, fifteen years earlier. It came back to publishable form only through a structured collaboration with several language models, which I described in a separate paper.
I am neurodivergent, and the mismatch between a nervous system built for ideation and the executive-function-heavy labor of linear documentation had left the work stranded in keynote transcripts and slide decks. The models synthesized that material. I validated every concept and kept every judgment. That was heavy, global-level offloading. But it restored my intellectual agency rather than eroding it, because conceptual authority stayed with me, on the human side of the loop. Treating all such help as substitution confuses access with avoidance.
Fluency stays behind the glass
Johan Huizinga coined the Magic Circle for the bounded space in which play runs under its own rules. In the context of game-based learning, the circle acts as a double agent. It lets a player try risky moves without real consequence, but it also traps the resulting learning inside the game. Players learn to beat the game rather than master the domain, which is why far transfer, the application of competence to a new and unassisted task, has been the persistent failure of game-based learning.
The conversation with a language model forms the basis of the modern Magic Circle. I call that circle the Synthetic Context. Inside it learners generate and recombine text at almost no cost. That is valuable, because they can make mistakes without consequence. But a language model lives entirely in the domain of symbols. It never meets the resistance of material consequences. It smooths over missing knowledge before the learner notices, and the fluency that results is deceptive.
Subramonyam and colleagues call the gap a “gulf of envisioning” between what a person can co-produce with the model and what they can independently conceive. Performing with the model is one thing, learning from it is another, and assessment has to tell them apart.
Engineering the exit
Far transfer does not happen on its own. The teacher has to design for it. The first of two heuristics I draw from the theory is Strategic Translation: the deliberate conversion of AI-mediated activity into representations the learner owns. It is a designed rhythm that keeps moving the learner between working with the model and working without it.
It is important to point out that a simple quiz attached to the end of a passive task is not sufficient to establish that rhythm. The rhythm is built from repeated translation cycles, in which each full cycle has six steps.
The learner makes an initial attempt or prediction before any model assistance.
The interaction with the model is constrained so that it exposes reasoning instead of bypassing it.
The output is evaluated against evidence or criteria that live outside the model.
The learner compresses the result into their own explanation, whether as a hand-drawn concept map or an oral defense of the AI’s feedback.
After a delay, they retrieve the material without assistance.
They apply it to a problem that works the same way but looks different.
None of this is really new, though nobody has yet tested all six together. Requiring a judgment before seeing system advice (step 1) is the cognitive-forcing action that reduced overreliance in Buçinca and colleagues’ work. Constraining the interaction (step 2) is the operative difference in the Bastani and Kestin studies, and teach-back (step 4) is Gordon Pask’s original certification of learning. Each element on its own is a design principle. Composed, they form a repeating instructional design rhythm.
In this context, “unassisted” does not mean stripped of support in an ableist sense. Accessibility tools and accommodations stay. The test is whether the generative system is substituting for the target competence. A student who uses the model to get past executive-function barriers can keep that support and still be tested on the intended outcome.
A classroom full of loops
The second heuristic, Process Governance, is about the teacher’s role in relationship to the loop. The teacher’s job shifts from delivering content to regulating a classroom full of individual loops, standing to each loop as the learner stands to the model. The learner watches the model’s output and decides what to accept. The teacher watches how the learner runs that loop and decides when to step in. It is the same second-order position, one level up.
In practice that means shared prompt protocols, checkpoints, brief process reflections, and delayed transfer tasks. It demands intervening when a loop shows a recognizable failure mode: goal drift, repeated completion requests, unverified acceptance, circular prompting, or a widening gap between assisted and unassisted work.
Assessment follows the same shift. A finished artifact is something the first-order system can now produce alone. It is therefore the visible process of regulation that becomes the object of interest: prompt logs and revision histories, and the ability to correct an error the teacher planted. Oral defenses and process portfolios work well, especially when they are short and sampled, but they should never turn into a hunt for AI use. Their purpose is to observe learning.
It should be noted that Process Governance also carries a hazard. The trace data that make second-order regulation possible also make surveillance possible. The log that shows a stalled loop can also expose a disability or an immature idea, which is precisely the exploratory material the Synthetic Context exists to protect. Process Governance therefore should never require analysis of every keystroke. It should require proportionate process evidence and full disclosure of what is recorded and who sees it. Accountability always has to stay human.
Whatever the loop constructs has to survive outside the loop, outside that Magic Circle. That single principle sits behind both heuristics, seen from the two sides of the loop. So the question for teachers is no longer whether AI helps learning. It is which prompting profiles and which classroom configurations lead the loop to converge on understanding that is still there when the tool is turned off.
No side to pick
I opened this post with a teacher reading two sets of studies and receiving two conflicting instructions. That teacher was right to be confused. They were being handed measurements from two ends of one system and asked to pick a side. But there is no side to pick.
The student who leaves the loop knowing more than they brought to it and the student who leaves with a polished artifact and nothing else sat in front of the same tool. The tool did not decide which of the two a given student turned out to be. The loop did, and the loop is something a teacher can design. That is a heavier responsibility than “adopt” or “restrict” ever was.
It is also, I think, closer to what teaching actually should be.
The images in this article were generated with Nano Banana 2.
If you’d like to go further, the following NotebookLM-generated audio deep dive goes beyond the post into my foundational academic paper that defines Synthetic Constructivism. This is meant as a companion to the argument, offered as an optional extra rather than a summary of it.
P.S. I believe transparency builds the trust that AI detection systems fail to enforce. That’s why I’ve published an ethics and AI disclosure statement, which outlines how I integrate AI tools into my intellectual work.