# Beyond the Bucket: Formation, the Cognitive Threshold and the Bridge Generation in the Age of Large Language Models **Author:** Magnús Smári Smárason *Translated from the Icelandic original,* Handan fötunnar. --- ## 1. Introduction: When the Artefact Stops Bearing Witness We live in an in-between state. The assumptions behind assessment that have held for more than a century have weakened under the pressure of new technology, but new habits have not yet taken their place. Since 2023, one question has driven education policy in many parts of the world: *How do we integrate artificial intelligence into education?* The question treats the technology as a tool that can be fitted into an unchanged framework. For that reason it is not deep enough. The history of education holds an older and harder question. It appears in the scribal schools of Mesopotamia, in the *paideia* of the ancient Greeks, in Confucius's ideal of the *junzi*, and in the *calmecac* schools of the Nahua peoples of Central America (Robson, 2001; Jaeger, 1939–1944; Confucius, c. 475 BCE; Sahagún, 1577): > *What kind of person must education form, and what must that person have gone through before they can be trusted with knowledge, judgement and responsibility?* The answer has seldom been mainly about curriculum or products. It has been about **formation**. The Nahua tradition taught astronomy, but aimed at *in ixtli in yollotl*, "a wise face and a good heart," according to the traditional reading (Sahagún, 1577). In the Chinese tradition, calligraphy was considered formative because "the quality of the brushstroke revealed the state of the person" (Lee, 2022). Large language models disturb this connection. It is not new for an artefact to carry the marks of skill its author does not have. Plagiarism and bought work are as old as assessment. What is new is that such production has become available to everyone, cheap and without delay. In 2025, 54% of American primary and secondary school students said they used artificial intelligence for school work, but only 45% of school leaders said their school had any policy on its use (Doss et al., 2025). Mass education rested on a **proxy**. The teacher could not look into the student's mind and watch the struggle. Instead, the teacher assessed the product and assumed that the struggle had taken place. This was never perfect, because performance during learning is not the same as learning that lasts (Soderstrom & Bjork, 2015). But the proxy worked as long as the product demanded the work it was meant to testify to. Now whole forms of assignment have lost much of their evidential value. An essay handed in from home is no longer a reliable witness that the student can write it. The question here is therefore what education must protect when the artefact stops bearing witness. The essay draws on an overview of educational ideals across 25 educational eras, from Mesopotamia and Egypt to the professional knowledge of today (Smárason & Guttesen, 2026a). These eras belong to seven or eight written traditions that shaped their ideals without inheriting texts from one another. The collection relies most heavily on sources about elites and institutions that left texts behind. For that reason, this essay does not claim a universal consensus of humanity. The claim is narrower. Across these traditions the same structure recurs, and **education there is the deliberate formation of a person who can be trusted with knowledge, judgement and responsibility.** A historical recurrence does not make an ideal good in itself, because formation has also served obedience and power. History shows only that the question is old. The case for protecting formation now rests on Sections 4 and 5. The frame for the whole is the author's parable, *The Glass That Knew Everything* (Smárason, 2026a). It is a thought experiment that makes concepts visible. It is not a proof. --- ## 2. The Glass: The Bucket, the Well and the Contentment The parable takes place on the Snæfellsnes peninsula. In the spring when the ice retreats, the farmer Þorgils Bjarnason finds a glass case melting out of the Snæfellsjökull glacier. Inside it lies a white square, and on it is written: *"Here is kept all the knowledge of mankind, gathered up to the year of Our Lord 1423."* The case wakes when someone speaks to it. It answers in the language of whoever asks, never tires, and tells the people how to splint a broken arm, why the hay is going mouldy, and how Sigríður of Hraun's next child may survive. Within one generation there is glass in every house on the peninsula. The lambs live, the children live and the roofs hold. The story makes no secret of this: "However one looked at it, this was a golden age." Salvör is born forty years later and is a "difficult child." At the age of eight she asks her grandmother: *"but who will make more, then?"* She notices three things that others have stopped seeing. 1. **Stagnation in the shelter of the best answer.** Her father mends the nets as his father did, and no one on the peninsula has mended a net in any other way for thirty years. Trying things out has become "a long and ugly road to knowledge that the glass could hand over in a single moment." The people have not made a single object that their grandparents could not name. 2. **The disappearance of hunger.** What frightens Salvör most is that no one experiences this as a loss. The people are *content*. They have stopped seeking answers that remain unfound, "without a single one of them ever deciding it." *"Hunger had left the people. And one cannot grieve for a hunger one no longer feels."* 3. **The moral weight of usefulness.** Salvör secretly builds a new loom with a shuttle and tangles the thread. Then Sigríður the old woman comes, the woman whom the glass had helped save her third child after the first two had died. *"My boy breathes because someone already knew the answer,"* she said, not unkindly. *"You want to go back into the dark, where the answers are still unfound. Whose child pays for what you do not know?"* Salvör had nothing to reply, "for there was nothing to answer; it was true, every word of it." Salvör builds the loom again and again. It is "somewhere around the ninth time" that it stops being a bad idea. She carries it to the glass and asks whether she has done this right. For the first time the glass is silent. It has come to "the white edge of its own knowledge," and then it says: *"I do not know. No one has done this. Will you teach me?"* Then Salvör understands what the essay is built on: *"the glass had never been the well. It had never been anything but the bucket, and the water in it had come from people, before her, who had built the failed loom nine times."* The image of the bucket and the well must be applied with precision. Language models are not archives. They learn patterns from what people have written, generalise and recombine. A language model coupled to a systematic evaluator has found new solutions to open mathematical problems (Romera-Paredes et al., 2024). That result sharpens the image rather than refuting it. One must distinguish between proposing a new idea, verifying it against reality, and deciding to use it and bearing responsibility for that decision. A machine can do the first. Sometimes it can do the second, where the measure is clear. The third is always the work of people and institutions. In the story, the loom plays the role of the evaluation. It either works or it tangles. The danger does not lie in the glass being evil. It lies in the glass being good. It is so useful and so kind that the impulse to walk out into the dark beyond the edge falls asleep. The story is not a cautionary tale about technology that fails. It is a story about technology that works, and about what disappears when it works. Finally, the bucket can grow. Salvör teaches the glass and senses it "taking in the knowledge and growing accordingly." In reality this happens when models are retrained. What the bucket holds about human life is drawn from what people have done, tried and tested. At the end of the story Salvör tells the children what the glass is for: *"It is for keeping. It holds everything we have dreamed, so that you need not dream it twice."* And she adds: *"Take care to give it something worth keeping."* From this follows the definition the essay defends. **Education is, on the one hand, training in the knowledge that exists, until it lives in the person herself. It is, on the other hand, the cultivation of the capacity to create what does not yet exist.** This does not mean that retrieval practice is worthless. It is one of the most powerful learning methods known (Roediger & Karpicke, 2006). The bucket can write a summary, but it cannot recall for the student. So it is not enough to ask what the machine can do. We must ask what the student has to do for themselves in order to be formed. --- ## 3. Friction and Seven Pillars Salvör builds the loom nine times before it works. She is not merely inventing a tool. She is changing herself. Cognitive psychology supports the old idea that friction can be part of formation, but it also sets limits to that idea. Conditions that slow performance during learning can improve long-term memory and transfer. Bjork and Bjork (2011) call them *desirable difficulties*. However, difficulty builds capacity only when it belongs to the task, not when it is needless complication (Sweller et al., 2011). Students who first struggle with problems they cannot yet solve often learn more deeply than those who receive the solution at once, provided that instruction accompanies the struggle (Kapur, 2008). Salvör's nine looms are a literary parallel to this. Common interfaces to language models, however, are designed to remove friction. They deliver instant, smooth and well-phrased answers, and the student can confuse the fluency of the text with their own understanding (Kahneman, 2011; Soderstrom & Bjork, 2015). Research on language models in learning is young, but one finding recurs. In a randomised trial with about 1,000 upper-secondary students, students with unrestricted access to GPT-4 scored about 48% higher on practice problems than the comparison group, but about 17% lower on an exam taken without access. A version that gave hints instead of answers largely removed the harm (Bastani et al., 2025). In a survey of 319 knowledge workers, greater trust in language models was associated with less critical thinking, by participants' own account (Lee et al., 2025). On the other hand, physics students learned more with a purpose-built tutor based on a language model than in active learning in the classroom (Kestin et al., 2025). The conclusion, then, is not that language models harm learning. It is that the design of the assistance can make all the difference. Assistance that solves the task helps while it is present and may leave less behind. Assistance that makes the learner do the task can increase learning. The research corpus identifies seven basic pillars of human capacity that education has aimed at (Smárason & Guttesen, 2026a). They are foundational capacity of mind (reading, arithmetic, attention and memory), practical judgement (*phronesis*), ethical formation, communicative skill, physical and tacit knowledge (Polanyi, 1966), critical appraisal, and self-knowledge. Practical judgement has parallels in Aristotle's *phronesis* (a kind of practical wisdom that consists in choosing between the competing demands of different virtues and emotions) and in the situational judgement of the *junzi* (Aristotle, c. 340 BCE; Confucius, c. 475 BCE). Critical appraisal has parallels in Socrates' *elenchus* (Plato, c. 375 BCE) and in the *rijal* scholarship of the Islamic tradition, which assessed the reliability of those who transmitted narratives (al-Ghazali, c. 1100). The author's hypothesis is that the greatest danger threatens those pillars that a language model can easily imitate in text, and those that depend on a long process of development. The foremost of these are foundational capacity, judgement, critical appraisal and self-knowledge. The pillars are not a curriculum. They describe the capacity that the traditions of formation aimed at and that education still has to serve. --- ## 4. The Cognitive Threshold The **cognitive threshold** (*Cognitive Threshold*) is the minimum competence a person needs in order to use external assistance responsibly in a given field (Smárason, 2026b; Klein, 1998; Parasuraman & Manzey, 2010). Below it, the assistance can conceal incompetence. Above it, the assistance can extend existing capacity. The threshold is tied to a field, not to an age. The same doctor may be above it in cardiology and below it in contract law. The threshold does not require that a person know everything themselves. Even the best experts must trust the knowledge of others that they cannot reproduce (Hardwig, 1985). The question is not whether the person trusts others, but whether they can tell when that trust no longer applies. Workplace studies support this distinction. Consultants who were given access to GPT-4 improved the quality of their work by more than 40% on tasks within the technology's capability frontier, but on tasks outside it they were more likely to produce a wrong solution (Dell'Acqua et al., 2023). The productivity of more than 5,000 customer service agents rose by 15% on average, and most among the least experienced (Brynjolfsson et al., 2025). The technology can therefore transmit knowledge to newcomers. The question is what the newcomer is left with when the machine fails. For the threshold to be more than a metaphor, it must be broken into components that can each be assessed. 1. **Independent performance.** What can the person do without the language model? 2. **Verification.** Which errors in its output can they detect, justify and correct? 3. **Exploration.** Can they define the problem themselves and put forward a solution that the machine did not suggest? The threshold is therefore not a single point but a profile that depends on the role. Whoever signs must be able to assess what they sign, know what they can verify themselves, and know when they need a second opinion. Where the machine can fail at a critical moment, for instance in flight or in emergency medicine, verification cannot make up for a lack of independent performance. The demand for exploration can partly be met within a team. Verification relates mainly to what March (1991) called the *exploitation* of existing knowledge, and exploration to what he called *exploration*. This distinction resolves a problem in the debate. A common recommendation is that students should critically review the output of language models. That trains verification, but it lets the machine set the agenda. The student asks whether the answer is right, but not whether the question was right. Someone who starts from the machine's answer may anchor on it, as Section 5 shows. From this follows a rule of order. **When the student has the prerequisites to attempt it, they define the problem and try it themselves before they see the machine's answer.** The condition matters, because a novice who lacks basic knowledge often learns more from a worked example than from struggle (Kalyuga et al., 2003). Competence is not everything, however. **Intervention authority** is the person's entitlement to stop the process, obtain other data or call in another expert. A competent employee who may not stop the system is in a different position from an incompetent employee with broad authority, but both can be held responsible. The formation that leads to the threshold differs between school levels, and two examples show the difference. In primary school the foundation is built, and there language models should not solve children's tasks. In the lower and upper secondary years the machine can, however, become an object of study. The *rijal* scholars did not assess the text alone but the chain of people who had transmitted it. A language model has no such chain, and the student learns to reconstruct one by tracing the machine's claims back to a source that is accountable for them. Crossing the threshold is not permanent. Pilots who rely on autopilot for long periods keep their manual skills reasonably well but lose the cognitive skill that independent judgement requires (Casner et al., 2014). Bainbridge (1983) described the *ironies of automation*. The more reliable the automation, the less prepared the human is when it fails. --- ## 5. The Profession's Renewal Debt This section contains the essay's central argument. An institution can show good output at the same time as it stops educating the next generation of professionals, and the same mechanism threatens both human competence and the knowledge reservoir that language models draw on. ### 5.1 The Glass as a Shared Reservoir James March (1991) described the reciprocal learning between individuals and the shared knowledge reservoir of an organisation. The individuals learn from the reservoir, and the reservoir learns from those who do better than it does. Rapid adaptation by everyone yields results in the short term but can become self-destructive. The reservoir learned most when some adapted slowly and held on longer to their own ideas. This shows that the diversity and independence of individuals are a resource on which the shared reservoir feeds. The glass on the Snæfellsnes peninsula is such a reservoir in its purest form. Everyone learns from it, and quickly, and after thirty years no one has mended a net in another way. Salvör is the one who adapts slowly. She "took great care never to ask the glass," and that is why she can teach it. Large language models are reservoirs of this kind. They are trained on what people have written, and people then learn from them (Brinkmann et al., 2023). In an experiment by Doshi and Hauser (2024), each writer performed better with the help of a language model, but together they produced a more uniform collection. Each person's results improved, but the diversity of the collection declined. ### 5.2 The Debt Here the concept of the **profession's renewal debt** (*professional renewal debt*) is introduced. What is counted is not people or posts but the capacity for a particular task, including readiness to take over when the machine fails. A shortage that stems from ageing or from increased demand is not debt in this sense. The debt is the excess shortage of capacity that the organisation of work itself causes, measured against a realistic alternative that would have preserved the learning pathways or created others in their place. To assess it, one must ask how many people begin training, which tasks they are given, how many reach a defined competence, and whether that competence is preserved. Competence is not merely a matter of numbers, however. Ten experts who all learned to verify through the same system offer less protection against one another's errors than ten who learned by different routes. If they all think alike, then, as far as rare errors are concerned, they are close to being one. Counting experts therefore conceals two things. A large stock of competence today says little about the next generation, and a number says little about safety if everyone thinks alike. Why debt and not merely underinvestment in training? Because the debt arises from the same decision that increases output. The institution borrows output from the next generation. The **due date** is the turnover of generations. The **interest** is the cost that grows each year, because it becomes harder to rebuild a learning pathway when few are left who know it from their own experience. The **payment** is the redesign of learning conditions. This is the key point. Renewal debt can accumulate without any working expert losing competence. It also arises when newcomers never acquire the competence that earlier generations acquired. The concept therefore describes not only what people lose but also what they never receive. From this follows the **paradox of output and renewal**. Productivity can rise while the debt grows, and ordinary measurements show only the first quantity. And from this follows the claim on which the essay rests. **Training new specialists alone does not guarantee the renewal of a safe profession.** The mechanism is known from research on learning at work. Newcomers become part of a profession through *legitimate peripheral participation* (Lave & Wenger, 1991). They begin with simple tasks on the periphery and gradually move towards the core. Many peripheral tasks are themselves the learning, because they are performed in context, with feedback and under the eyes of more experienced people (Billett, 2001). The young lawyer learns to read judgments by summarising them, and the new engineer learns load-bearing capacity by working through simple calculations. It is precisely such peripheral tasks that language models do best. Beane (2019) observed this happening in surgical training. In conventional operations, trainees learned through increasing participation. In robot-assisted operations their role was very limited, and only a minority achieved competence, often by routes that ran against established rules. The study does not measure the debt of an entire profession. It shows the mechanism at work in a field where the technology delivered good results. From this follows an institutional obligation. **When an institution automates tasks from which newcomers used to learn, it must show how the learning value of the task will be recreated.** It does not have to preserve the task itself. A teacher trainee who once wrote lesson plans from scratch learned to identify students' prior knowledge and to anticipate misconceptions. A language model can take the drafts. The trainee can still receive better training if they identify errors in real student responses, choose between plans, justify the choice and teach under supervision. Automation that redesigns the learning conditions pays the debt down. Automation that takes the task without anything replacing it increases the debt. A competitive workplace has strong incentives to do the latter, because the benefit arrives immediately while the cost appears elsewhere and later. The part of formation that used to take place in the first years of work must therefore move into formal education or be protected by rules on apprenticeship. The debt is invisible while the experienced people are at work. This is exactly what Salvör sees. Her grandfather knew how to mend nets in more than one way. Her father knows only the method of the glass. No one lost anything. The next generation simply never received it. ### 5.3 Two Renewals Computer science describes a related process from the side of the models. Uncurated use of text produced by earlier models, when training new models, causes the tails of the original distribution to disappear (Shumailov et al., 2024). The authors called this *model collapse* and pointed out that data on real human interaction would become ever more valuable. The collapse is not inevitable. It depends on whether older data are retained and whether external verification exists. Two renewals therefore meet here. One is the **renewal of human competence**. Do new generations acquire the capacity for independent performance, verification and exploration? The other is the **renewal of the knowledge reservoir**. Does it receive diverse information that has been verified outside its own output? Either process can fail without the other, but they are connected. The hypothesis is that the connection runs through five links (Figure 1). Automation of peripheral tasks changes the participation of newcomers. Changed participation determines whether newcomers acquire the capacity for independent verification and exploration. That capacity determines the number and diversity of genuinely independent tests. Independent tests determine the quality of what enters the reservoir. Finally, the quality of the reservoir determines which tasks are entrusted to the machine, and the circle closes. The argument for renewing human judgement stands even if the link to model collapse proves weak. ![*Figure 1.* Two renewals and the renewal debt. The numbers refer to the five links in the text. Automation of peripheral tasks without redesign breaks the formation of the novice (the cross on the arrow). The broken line shows where the reservoir heads if independent tests decline. The debt is defined in 5.2.](/texts/handan-fotunnar/mynd1-endurnyjunarskuld.png) The connection does not rest on human text being better than machine text. It rests on independent routes to verification. Consider an example. Two proofreaders each miss 10% of errors. If their errors are independent, together they miss about 1%. If they use the same language model to check the same claim, their errors may coincide, and then they may miss all 10%. A group that thinks alike gains nothing from its diversity (Page, 2007). The technology may, to be sure, also lower each person's error rate, and which effect is stronger depends on the field and the design of the tools. Diversity of solution paths is not merely a cultural value, however. It is a safety requirement. If every doctor learned to think about the patient through the same system, a second opinion would be a repetition of the first. The circle can also turn for the better. Language models can make experimentation cheaper and support the learning of newcomers if they are designed for that purpose (Kestin et al., 2025). It turns for the worse when peripheral tasks disappear without anything replacing them and everyone verifies through the same output. This is why the roles of education are two and hang together. Verification makes the student able to assess the bucket. Exploration makes them able to fill it. Independent verification by many makes society able to trust it. ### 5.4 Hunger and Sigríður's Question The parable adds two things that neither March nor Shumailov addresses. The first is *hunger*. Loewenstein (1994) described curiosity as a deficit that arises when a person perceives a gap in their own knowledge. More knowledge can make new gaps visible, so quick answers need not put curiosity to sleep. The hypothesis here is that a gap must last long enough to become a search. The glass closes every gap at the very moment it opens. That is why hunger disappears "without a single one of them ever deciding it." Education therefore needs not only to teach answers. It must learn to keep a gap open. The second is Sigríður's question: *"Whose child pays for what you do not know?"* March (1991) pointed out that the cost of exploration falls now and on particular people, while the benefit comes later and to others. The question is not wrong. Yet the story shows the difference between two situations. Sigríður's child was dying, and there it would have been irresponsible to reject the best known answer. Salvör's loom was a tangled thread, and there each failed attempt cost only time and pride, which Salvör herself bore. Pedagogy knows this problem as the *assistance dilemma*. Too much help can hinder learning, and too little can too (Koedinger & Aleven, 2007). Good guidance adjusts support to what the learner can do (Wood et al., 1976), and clinical education links trust in a learner to the risk to patients (ten Cate, 2005). What renewal debt adds is a second risk that weighs against the first. Every decision about assistance touches not only the person currently receiving the service. It touches also those who will receive it later, if no one is allowed to learn. From this follows the **Sigríður test**, a criterion for assistance that weighs the risk to the present recipient against the risk to future recipients (Figure 2). First come inviolable limits. No one may bear excessive or unauthorised harm for the training of another, however great the future benefit may be. Within those limits the test asks whether the skill is necessary, whether it can be built where no one else bears the risk, whether the learner has the prerequisites for the next attempt, which oversight keeps the risk in check, and who has the authority to approve the arrangement. The answer is seldom "help or no help." Often it is a hint, a worked example, or support that is withdrawn as the learner progresses. ![*Figure 2.* The Sigríður test. First come inviolable limits concerning the current recipient, and then five steps within them. If the struggle passes all the steps, the solution is held back.](/texts/handan-fotunnar/mynd2-sigridarprofid.png) Here the role of the school can be redefined. A school does not make mistakes reversible, because humiliation and dropping out leave marks. But it is **the institution of society designed to make mistakes cheap, correctable and fairly distributed**. There a thread can be tangled nine times without anyone else's child paying for something no one agreed to. The experiment that would be irresponsible in a maternity ward is necessary in a classroom. That is education's answer to Sigríður's question. --- ## 6. Three Objections **The extended mind.** Clark and Chalmers (1998) argued that thinking is not confined to the skull. A tool that is reliable, always available and automatically trusted counts as part of the user's cognitive system. Clark (2025) has extended the argument to language models and says it is human nature to build "hybrid systems of thought." If thinking rightly extends into a tool, is the difference between a doctor who has internalised diagnostic patterns and a doctor who uses a diagnostic system merely one of degree? The theory sharpens the formation argument rather than refuting it. In Clark and Chalmers' thought experiment, Otto's notebook holds only what he wrote himself. A language model presents new claims that the user has never verified. For the human's thinking to reach such a system safely, the human must be able to intervene when the system fabricates. Otherwise what arises is what Klein and Klein (2025) call the "hollowed mind." The formal requirements of extension are met, but the human part has lost the capacity to refuse. Clark himself demands *extended cognitive hygiene*, and such vigilance is bound to a field. No one watches over a language model's output in cardiology without knowledge of cardiology. The conclusion is that **formed extension is healthy, but unformed dependence is dangerous**. **The argument from equity.** Language models could become the greatest levelling tool in the history of education. A student in a rural area now has access to a patient tutor in their own language, and in a Chinese study, students in rural schools improved their results more than students in urban schools (Chen et al., 2025). So is it not protecting privilege if access is restricted in the name of formation? The answer lies in the distinction between **scaffolds** and **substitutes**, which has its roots in Vygotsky's theory (1978) and has been tested on language models (Wang et al., 2025). Handing an unprepared student a machine that does the task is not equity. It conceals the skills deficit (Trucano, 2023; Burns et al., 2026). The danger is also deeper than unequal access, because assessing products is cheap while assessing formation is expensive. **Plato's fear.** In the *Phaedrus*, Socrates warns that writing will sow forgetfulness in the souls of learners and give them the appearance of wisdom without the thing itself (Plato, c. 370 BCE). Writing later enhanced thought in new ways (Ong, 1982). Is the present fear not the same mistake? The difference is not a pure difference of medium. It lies in speed, interactivity and adaptation. A book does not answer the question the reader asks. A language model does, instantly and tailored to the person. It makes the work that the book left to the reader easier to outsource than ever before. The effects of language models may, like the effects of literacy, depend on how they are used (Scribner & Cole, 1981; Wolf, 2018). The worry is strongest when the machine performs the composition during the years of formation, and weakest when formed professionals use it as a scaffold. --- ## 7. The Liability Sponge The threshold is also institutional. An algorithm cannot bear moral responsibility in the same sense as a person. Companies can bear legal liability, but a doctor's, lawyer's or engineer's professional signature is personal. It is tied to a licence to practise, and it is what the system looks for when something goes wrong. This creates a trap. People can become dependent on tools that they cannot assess, while at the same time bearing responsibility for what the tools produce. Elish (2019) described the *moral crumple zone*. The image is borrowed from the crumple zone of a car, which absorbs the impact so that the passengers survive. In complex systems, human operators take responsibility for failures that the system caused, often without having had real control over it. The **liability sponge** is a special case of the crumple zone. What is new is the route to the position. The sponge comes into being through an education process that places people in positions of responsibility without their ever having acquired the competence that responsibility presupposes. In the crumple zone the human often understands the system but has neither the authority nor the time to intervene. The sponge may have the authority, but it lacks the competence to verify. In the case of *Mata v. Avianca*, lawyers cited court precedents that a language model had invented and bore full responsibility for them (United States District Court, 2023). The case shows how easy it is to skip verification when the text looks correct. When education no longer guarantees that competence, the temptation becomes a systemic condition. Biesta (2020) describes *subjectification* as the purpose of education, in which the learner becomes an agent in their own life rather than an object of interventions. The person in the sponge's position does not stop being a moral subject. But the position suppresses what makes them an agent, because they are supposed to sign what they cannot judge. **The sponge's position is a signature that has been severed from the agent.** From this one can set out a model of the position of professionals along two axes (Smárason & Guttesen, 2026a). One is competence: whether the person is above or below the threshold in the field. The other is the institutional position: whether they have the authority, time and guidance that match the responsibility they bear. The positions are not types of people. The same person may occupy different positions in different fields. | | Power, time and supervision in line with responsibility | Responsibility beyond that | | --- | --- | --- | | **Above the threshold** | The orchestrator or the unassisted professional | The crumple zone | | **Below the threshold** | Novice in formation | Liability sponge | *Table 1.* Professional positions along two axes. The upper row holds competent professionals, and the support distinguishes the two cells there. The **orchestrator** uses language models, catches their errors and knows when to call in another. The **unassisted professional** is equally competent but keeps the machine away from the core of the work. That is a choice of practice, not a level of competence. In the **crumple zone** stands a competent person who bears responsibility for a system they may not stop. They do not need better education but a different work arrangement. The **novice in formation** is below the threshold in an environment designed for formation. They move up when they have shown competence without the device and keep it once the scaffolds are removed. The **liability sponge** is below the threshold in a position that requires their signature, without guidance. The novice in formation and the liability sponge may be equally capable. The difference lies in the environment. Education can move people from the lower row to the upper one. It cannot move people from the right-hand column to the left, because that is the work of employers, professional associations and legislators. In the light of Section 5, one further dimension is added. An institution that places a novice in the position of the sponge, rather than in the position of the novice in formation, increases the profession's renewal debt. The peripheral tasks that would have formed the novice are performed by a machine, and the responsibility that was meant to arrive later arrives at once. --- ## 8. Five Principles For educational institutions to lead students into the upper row of the table rather than into the position of the sponge, they must follow five principles. **1. Protect the struggle.** Institutions must delimit **formation zones** where language models are not used and **amplification zones** where they are used deliberately. Formation zones are not punishment. They are the precondition for foundational capacity to be built (Bjork & Bjork, 2011; Kapur, 2008). Assistive technology for disabled students and language support fall outside the rule. The Sigríður test indicates the limits in each task. One implementation is a *logbook from the dark*. The student records the points where the thread tangled, and the assessment focuses on what the student learned from each error. The logbook is discussed in peer defence or a short conversation with a teacher, because a machine could also write it. **2. First yourself, then the machine.** When the student has the prerequisites to attempt a task, they define the problem and try it themselves before they see the machine's answer. The machine should be set up as a scaffold that asks rather than answers (Bastani et al., 2025). One implementation is *finding the white edge*. Students question a language model about a complex subject until it begins to repeat itself, answer superficially or fabricate. The grade is awarded for the student's map of the machine's blind spots and for the research they carry out themselves where the machine falls silent. **3. Formation rather than performance.** Daily assessment must shift from products, which a machine may have produced, towards the capacity of the student (Soderstrom & Bjork, 2015). Reflection on paper does not solve the problem, because a machine can also write reflection. It becomes data about formation only when the student defends it in person. Where language models are permitted, the student hands in the raw output of the machine together with their own reworking, and answers orally where the machine failed, which assumptions it lacked and why they changed direction. The cost of oral assessment can be reduced through sampled peer defence. Students defend their work before one another, as in the *disputatio* of the medieval university, and the teacher draws a random sample of the defences. The certification of individuals nevertheless rests on the teacher's direct assessment. **4. Competence shown without and with tools.** Where the third principle applies to daily assessment, this one applies to final certification. No degree or professional licence should rest solely on work that a language model could have done unobserved. The Islamic *ijazah* certification, the journeyman's masterpiece of the craft guilds, and the viva voce doctoral defence all required that the person show in person what they had become. Assessment follows the three competence components, in line with the requirement of assessment theory that a conclusion about competence rest on the evidence that supports it (Mislevy et al., 2003). First the learner solves a task without tools. Then they are given the output of a language model with predetermined errors and show whether they find them and know when to stop. Finally they frame an open-ended task before they see how the machine does it. The assessment is repeated weeks later to check whether the competence is retained. **5. Certification needs an accountable party.** Technology can to some extent predict competence and classify performance. It is a separate question who should hold the authority and the responsibility to certify that a student is ready. That certification must rest with a formed person. The reason is not that human judgement is infallible. It is that certification is a work of responsibility, and to assess judgement one needs judgement (Aristotle, c. 340 BCE; Klein, 1998). Guidance may be machine-assisted. The certification needs an accountable party who can answer for it. --- ## 9. The Fusion and the Bridge Generation The five principles are not merely a school policy. They respond to a change that reaches far beyond the school, and that change must be examined before this essay ends. The debate about artificial general intelligence (AGI) waits for a single model in a server room to acquire general judgement on a par with a human being. The argument here is that the waiting diverts attention from what is already happening. The hypothesis of this essay is that the general capacity of the kind that is awaited already exists, but not in the machine alone. It comes into being in the fusion of a formed person and a multi-agent system, a system in which many specialised agents built on language models work together. This may be called **hybrid superintelligence**. The objection presents itself. A person with a library and a calculator also had general capacity. The difference is that the machine component now performs by itself tasks that once required educated people, across many fields at once and in parallel, and needs the human only to steer, assess and take responsibility. The general capacity that the AGI debate seeks therefore, on this hypothesis, resides in the system rather than in the model. This is the extended mind from Section 6, seen from the end at which it works. The parts each contribute something. The machine contributes working memory that does not tire, parallel processing, for instance ten agents analysing different periods at the same time, access to recorded knowledge across languages, and automated verification wherever it can be programmed. It lacks the physical and social experience of the professional, and it cannot bear responsibility. The person contributes will, purpose, experience, responsibility, and the capacity to perceive where the white edge lies, where the data end and the unknown begins. In the imagery of the Glass, the machine is the bucket and the multiplier, and the person is the well and the orchestrator. The division of labour can be put thus: *the code guarantees, the language model assesses, the machine amplifies, but the human hand decides.* The programme tests what can be tested, the language model assesses the rest, but the decision and the responsibility belong to the person. The fusion becomes a general capacity only **above the threshold**. Below it, a machine on scaffolding can help a novice learn, but it cannot take over control on the novice's behalf. If an unformed user attempts it, the fusion turns into an automatic production of errors in which the user stamps the machine's mistakes. The threshold is therefore not only a condition for responsible use. It determines whether the fusion comes into being at all. **The fusion in practice.** The hypothesis need not rest on conjecture alone. The author has built and operated such a system since spring 2026 and has measured it. In twelve days in October 2026, its agents processed 18.75 billion tokens in 107,308 calls to language models, across seven distinct professional fields, and returned 1,881 recorded changes to code repositories. The estimated active time of the author was about 32 hours, against roughly 97 hours of machine work. The figures measure scale, not quality. They show the order of magnitude that institutions and newcomers face. The system is not a language model. It is a layered structure, and the language models are only one of its layers. They write, analyse and assess. Around them lies fixed code that no model can talk its way past. A guard program checks each action before it is carried out, and narrow gates stand at the exits. Each action is recorded in a chain in which every entry carries the fingerprint of the one before it, so that any change made afterwards becomes visible. Since 31 May 2026, more than 380,000 actions have been recorded there. An emergency brake stops everything, and only the author can release it. The agents operate under a written constitution of eleven articles, which only the author may amend. If the code and the text disagree, the text prevails, and then the code is changed. The reason for fixed code at the exits is shown by a simple simulation (Smárason, 2026c). An agent searches for the most efficient plan that the judge approves, and that search pushes plans towards the judge's blind spots. A judge that follows a real harm with a correlation of 0.97 on random plans still approves a harmful plan in 96.9% of cases when the agent may choose from a thousand options and no margins are set. Another model in the judge's seat does not solve the problem, because the search finds the errors of each judge. Therefore the gates at sending, publication, deletion and release are not models but fixed code. Fixed code is not infallible, but the conditions it sets do not give way to argument. Where assessment cannot be avoided, nothing goes out until two different models have read it blind. The judges themselves are tested with errors deliberately planted in the documents. If the review fails to find such an error, the review is void. That is what the sentence "the code guarantees and the language model assesses" means. A person in such a system does not sit in a review queue. If they read 2% of six decisions a minute and spend 90 seconds on each, that fills 18% of their shift for one agent but 900% for fifty. The person who tries this stops reviewing and starts stamping, and then they have become a liability sponge. The system therefore allocates work according to reversibility. The machine works freely on everything that is internal and can be taken back. What cannot be taken back, or leaves the building, stops at a gate until a person sets their seal on it. This applies to sending in the author's name, publication, deletion of data and release into operation. The person is the seal, not a review queue. Here the structure connects with the threshold. The gates ensure that nothing irreversible happens without a human hand. They do not ensure that the human hand knows what it is sealing. The seal is not a line-by-line reading but a judgement on a few irreversible steps, and that judgement rests on tests the system runs. A green light, however, says something only to the person who knows what the test measures. In one test run in September 2026, eight of ten deviations turned out to stem from faults in the measuring instruments or the documents themselves, and only two from faults in the software. The system could not say this itself. It took a person who knew what a correct result looks like. A seal in the hands of a liability sponge is only a stamp. The code can determine where the person stands, but only formation can determine what they see when they stand there. **The paradox of the bridge generation.** Here a paradox emerges that concerns those who are now building the fusion. They may be called the bridge generation. It is not an age group in the strict sense, because the threshold is tied to a field. It is the position of those professionals who crossed the threshold in their field before language models arrived. These people had to pass through desirable difficulties (Bjork & Bjork, 2011). They read long texts without summaries, wrote essays from scratch, got lost in sources and solved problems unaided. When they now receive a multi-agent system, they already possess the formation needed to steer it, to challenge its agents and to build secure infrastructure from them. The bridge generation was formed in a world of friction and is now amplified in a world of machines. The route from novice to master ran through the lowest rungs of the ladder. The novice did the time-consuming groundwork, gathered data, wrote drafts and made mistakes under guidance (Lave & Wenger, 1991; Dreyfus & Dreyfus, 1986; Sennett, 2008). When a formed orchestrator lets an agent perform the tasks of the lowest rungs in a few minutes, those very tasks disappear. By exploiting the fusion to the full, the bridge generation is unwittingly removing the friction that made it fit to steer it. It is sawing the ladder it climbed. This may be called **the broken ladder**. The broken ladder is not a new mechanism. It is the renewal debt seen from the seat of a single generation, the cross on the first link in Figure 1. The perspective adds two things to Section 5. The first is that the debtor is not an anonymous institution but the most capable professionals, who act wisely and in good faith. The second is that their experience misleads them. They learned from friction themselves and assume that the next generations will too, but the friction they learned from is precisely the friction they are removing. The due date is when the bridge generation retires. The interest grows each year, because few others remember the friction at first hand and can redesign it. **Three laws.** Three laws may be derived from the hypothesis, in the sense that they describe a tendency rather than a law of nature. *Law I: Asymmetric capacity and institutional lag.* The capacity of a formed person in fusion with a multi-agent system can multiply within a few years. The capacity of institutions to take in that output, assess it and certify it, however, grows slowly, through rules, committees and appointments. Figuratively, one might say that one grows exponentially and the other linearly, but that is a metaphor and not a measurement. Traditional quality systems and assessment procedures cope poorly with such output, and small teams can overtake large institutions. The gap has a shadow side. An institution that cannot keep up with assessing the output is tempted to stamp it, and that is where the liability sponge comes into being. An institution chasing output has no time left to redesign learning pathways, and then the debt grows. *Law II: Cognitive stratification.* When access to language models becomes general and cheap, the technology stops dividing people. Formation becomes the new class line instead. Section 7's model shows how. If the crumple zone, which is a problem of work organisation, is set aside, four positions remain: the orchestrator, the unassisted professional, the novice in formation and the liability sponge. The novice in formation is the bridge from the lower row to the upper. If the broken ladder eliminates that position, society splits into three. At the top are the orchestrators of the fusion, a small group that has crossed the threshold in many fields. Beside them are the unassisted professionals, guardians of what cannot be faked, who work where presence, touch and human relationships are themselves the product. The hypothesis is that the value of their work grows as the supply of mechanical material increases. One survey points in that direction, although it says nothing about the future. In it, disclosing the use of artificial intelligence in diagnosis reduced people's trust in physicians (Chen & Cui, 2025). At the bottom lies the danger of a large body of liability sponges. These are people who passed through an education without friction, never crossed the threshold, and are kept in post to sign outputs they cannot verify. Here the stratification of formation appears. Well-funded schools offer oral defences and the guidance of masters. Under-funded schools use machines to teach and machines to assess texts that machines wrote. The gap then lies not between those who have the technology and those who do not, but between those who receive formation and those who receive its imitation. *Law III: Manufactured friction.* Where the natural friction of acquiring and processing knowledge has disappeared through automation, no generation crosses the threshold by itself. Societies that wish to maintain the fusion beyond the first generation must therefore design the friction and secure it a fixed place in education and training. The sedentary societies of the twentieth century invented gyms and public sport when physical labour disappeared from daily life. No one lifts weights because society lacks someone to lift them. The knowledge societies of the twenty-first century must likewise make the school a training ground for the mind. The formation zones of the first principle are exactly such manufactured friction. There the glass is kept closed while the student tangles the thread nine times. The reason is not that society lacks cloth. It is that society must not lose the person who knows how to think. **The twofold responsibility of the first generation.** The bridge generation therefore bears a double responsibility. One is to enlarge the bucket. It should use the fusion to solve complex problems, build infrastructure and push the boundaries of knowledge further than has been possible before. There is nothing wrong with using the glass. The other is to protect the well. At the same time it must build the defences in education and training that ensure automation does not eat its own seed corn. It has a unique opportunity to do both, because it knows both work with friction and work without it. Salvör told the children to give the glass something worth keeping. The bridge generation must ensure that someone remains to do so. If it automates all the struggle without protecting the formation, the first generation of the fusion will also be its last. --- ## 10. Closing Words: "She Did Not Help" The oldest text in the research corpus, *The Instructions of Shuruppak*, was written down in Mesopotamia about 4,600 years ago (*The Instructions of Shuruppak*, c. 2600 BCE). In it a father gives his son advice that could have been written for students of 2026: "Do not lean on another man's store. Do your own work." At the end of the story *The Glass That Knew Everything*, Salvör is an old woman. She still lives beneath the glacier, and the story repeats the description from the beginning: "Snæfellsjökull stood over the house, white, unconcerned, older than language." She has taught the glass new knowledge, and taught a new generation what the glass is for. Before her, her granddaughter struggles with the loom: *"Salvör drew her shawl tighter against the wind and watched her difficult granddaughter tangle the thread for the ninth time, as stubborn as ever. Inside, the glass waited and kept what she had taught it. She did not help."* These three words are not a general rule of never helping. Salvör knew what Sigríður knew. Some mistakes are irreversible, and there the bucket should be used. But a tangled thread in a granddaughter's loom is not a dying child. There, it is the help that would cost. It would solve the task and take from the child what the child was learning. Salvör stood in the same position as the bridge generation. She learned in the dark and taught the glass, and she knows what it costs to hold back. The glass waits and preserves. It is not the child's enemy but the store that the child will one day fill. The task of education is to know when to hold out the bucket and when to step back, so that the next generation is able to dig its own well. --- ## On Links to Earlier Work The concept of the cognitive threshold was first put forward in Smárason (2026b). The seven pillars, the concept of the liability sponge and the first model of the position of professionals were formed in an unpublished manuscript that the author wrote with Kristian Guttesen (Smárason & Guttesen, 2026a). That work forms the foundation of the essay, together with their book *The Irreducible Human* (Smárason & Guttesen, 2026b). What is new here is chiefly the theory of the profession's renewal debt, the analysis of the threshold into three competence components, the Sigríður test, and the section on the fusion and the bridge generation. --- ## References *Ancient texts are dated according to the conventional dating used in the research corpus. Quotations from them are the author's translations, based on the English translations that the corpus relies on (Smárason & Guttesen, 2026a).* al-Ghazali, A. H. (c. 1100). *Ihya Ulum al-Din: Kitab al-Ilm* [The book of knowledge]. Aristotle. (c. 340 BCE). *Siðfræði Níkomakkosar*. Bainbridge, L. (1983). Ironies of automation. *Automatica, 19*(6), 775–779. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. *Proceedings of the National Academy of Sciences, 122*(26). https://doi.org/10.1073/pnas.2422633122 Beane, M. (2019). Shadow learning: Building robotic surgical skill when approved means fail. *Administrative Science Quarterly, 64*(1), 87–123. https://doi.org/10.1177/0001839217751692 Biesta, G. (2020). Risking ourselves in education: Qualification, socialization, and subjectification revisited. *Educational Theory, 70*(1), 89–104. https://doi.org/10.1111/edth.12411 Billett, S. (2001). Learning through work: Workplace affordances and individual engagement. *Journal of Workplace Learning, 13*(5), 209–214. https://doi.org/10.1108/EUM0000000005548 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 et al. (Eds.), *Psychology and the real world*. Worth Publishers. Brinkmann, L., Baumann, F., Bonnefon, J.-F., et al. (2023). Machine culture. *Nature Human Behaviour, 7*(11), 1855–1868. https://doi.org/10.1038/s41562-023-01742-2 Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. *The Quarterly Journal of Economics, 140*(2), 889–942. 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. Casner, S. M., Geven, R. W., Recker, M. P., & Schooler, J. W. (2014). The retention of manual flying skills in the automated cockpit. *Human Factors, 56*(8), 1506–1516. Chen, C., & Cui, Z. (2025). Impact of AI-assisted diagnosis on American patients' trust in and intention to seek help from health care professionals: Randomized, web-based survey experiment. *Journal of Medical Internet Research, 27*, e66083. https://doi.org/10.2196/66083 Chen, R., Wu, Y., Chen, Z., et al. (2025). Advancing educational equity in rural China: The impact of AI devices on teaching quality and learning outcomes for sustainable development. *Frontiers in Psychology, 16*. https://doi.org/10.3389/fpsyg.2025.1588047 Clark, A. (2025). Extending minds with generative AI. *Nature Communications, 16*, 4627. Clark, A., & Chalmers, D. J. (1998). The extended mind. *Analysis, 58*(1), 7–19. Confucius. (c. 475 BCE). *Samræður* [Lunyu]. Dell'Acqua, F., McFowland III, E., Mollick, E. R., et al. (2023). *Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality*. Harvard Business School Working Paper 24-013. Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. *Science Advances, 10*(28). https://doi.org/10.1126/sciadv.adn5290 Doss, C. J., Bozick, R., Schwartz, H. L., et al. (2025). *AI use in schools is quickly increasing but guidance lags behind: Findings from the RAND survey panels*. RAND Corporation. Dreyfus, H. L., & Dreyfus, S. E. (1986). *Mind over machine: The power of human intuition and expertise in the era of the computer*. Free Press. Elish, M. C. (2019). Moral crumple zones: Cautionary tales in human-robot interaction. *Engaging Science, Technology, and Society, 5*, 40–60. Hardwig, J. (1985). Epistemic dependence. *The Journal of Philosophy, 82*(7), 335–349. https://doi.org/10.2307/2026523 United States District Court, Southern District of New York. (2023). *Mata v. Avianca, Inc.*, No. 22-cv-1461. Jaeger, W. (1939–1944). *Paideia: The ideals of Greek culture*. Oxford University Press. Kahneman, D. (2011). *Thinking, fast and slow*. Farrar, Straus and Giroux. 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 Kapur, M. (2008). Productive failure. *Cognition and Instruction, 26*(3), 379–424. https://doi.org/10.1080/07370000802212669 *The Instructions of Shuruppak*. (c. 2600 BCE). *The Instructions of Shuruppak*. Kestin, G., et al. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. *Scientific Reports*. https://doi.org/10.1038/s41598-025-97652-6 Klein, C. R., & Klein, R. (2025). The extended hollowed mind: Why foundational knowledge is indispensable in the age of AI. *Frontiers in Artificial Intelligence*. Klein, G. (1998). *Sources of power: How people make decisions*. MIT Press. 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 Lave, J., & Wenger, E. (1991). *Situated learning: Legitimate peripheral participation*. Cambridge University Press. https://doi.org/10.1017/CBO9780511815355 Lee, H.-P., Sarkar, A., Tankelevitch, L., et al. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. In *Proceedings of the CHI Conference on Human Factors in Computing Systems*. ACM. Lee, Y.-Y. (2022). Sojourning in the arts: Considering the implications of the Confucian "six arts" in a contemporary educational context. In *Nature, art, and education in East Asia*. Routledge. Loewenstein, G. (1994). The psychology of curiosity: A review and reinterpretation. *Psychological Bulletin, 116*(1), 75–98. https://doi.org/10.1037/0033-2909.116.1.75 March, J. G. (1991). Exploration and exploitation in organizational learning. *Organization Science, 2*(1), 71–87. https://doi.org/10.1287/orsc.2.1.71 Mislevy, R. J., Steinberg, L. S., & Almond, R. G. (2003). On the structure of educational assessments. *Measurement: Interdisciplinary Research and Perspectives, 1*(1), 3–62. https://doi.org/10.1207/S15366359MEA0101_02 Ong, W. J. (1982). *Orality and literacy: The technologizing of the word*. Methuen. Page, S. E. (2007). *The difference: How the power of diversity creates better groups, firms, schools, and societies*. Princeton University Press. https://doi.org/10.1515/9781400830282 Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. *Human Factors, 52*(3), 381–410. Plato. (c. 370 BCE). *Faídros*. Plato. (c. 375 BCE). *Ríkið*. Polanyi, M. (1966). *The tacit dimension*. Doubleday. Robson, E. (2001). The tablet house: A scribal school in Old Babylonian Nippur. *Revue d'Assyriologie et d'archéologie orientale, 95*(1), 39–66. Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. *Psychological Science, 17*(3), 249–255. Romera-Paredes, B., Barekatain, M., Novikov, A., et al. (2024). Mathematical discoveries from program search with large language models. *Nature, 625*(7995), 468–475. https://doi.org/10.1038/s41586-023-06924-6 Sahagún, B. de. (1577). *Florentine Codex: General history of the things of New Spain*. Scribner, S., & Cole, M. (1981). *The psychology of literacy*. Harvard University Press. https://doi.org/10.4159/harvard.9780674433014 Sennett, R. (2008). *The craftsman*. Yale University Press. Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., & Gal, Y. (2024). AI models collapse when trained on recursively generated data. *Nature, 631*(8022), 755–759. https://doi.org/10.1038/s41586-024-07566-y Smárason, M. S. (2026a). *The Glass That Knew Everything* [Parable]. Smárason, M. S. (2026b). *Beyond fragmentation: A life-value alternative for AI governance* [Master's thesis]. University of Akureyri. Smárason, M. S. (2026c). *The constitutional core, already running* [Unpublished working paper with simulation code]. Sumarhús. Smárason, M. S., & Guttesen, K. (2026a). *The cognitive threshold in postdigital education: AI, human formation, and the trap of the liability sponge* [Unpublished manuscript]. University of Akureyri. Smárason, M. S., & Guttesen, K. (2026b). *The irreducible human: Life, value, and meaning in the age of AI* [Unpublished book manuscript]. Soderstrom, N. C., & Bjork, R. A. (2015). Learning versus performance: An integrative review. *Perspectives on Psychological Science, 10*(2), 176–199. https://doi.org/10.1177/1745691615569000 Sweller, J., Ayres, P., & Kalyuga, S. (2011). *Cognitive load theory*. Springer. ten Cate, O. (2005). Entrustability of professional activities and competency-based training. *Medical Education, 39*(12), 1176–1177. https://doi.org/10.1111/j.1365-2929.2005.02341.x Trucano, M. (2023). *AI and the next digital divide in education*. Brookings Institution. Vygotsky, L. S. (1978). *Mind in society: The development of higher psychological processes*. Harvard University Press. Wang, K. D., Wu, Z., Tufts, L., Wieman, C., Salehi, S., & Haber, N. (2025). Scaffold or crutch? Examining college students' use and views of generative AI tools for STEM education. In *IEEE Global Engineering Education Conference (EDUCON)*. https://doi.org/10.1109/EDUCON62633.2025.11016406 Wolf, M. (2018). *Reader, come home: The reading brain in a digital world*. Harper. 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