Publish at March 18 2026 Updated March 18 2026
There's a pedagogical ambition that has spanned the centuries without ever losing its topicality: to adjust the challenge "to the learner's needs". Too easy, and the exercise bores and dissolves attention; too difficult, and it discourages and fractures confidence. Between these two pitfalls lies what Lev Vygotsky conceptualized as the zone of proximal development: that subtle space where the learner can progress provided he is supported, guided and accompanied. Adjustment is not simply a matter of measuring out the difficulty, but of discerning the opportune moment, the fruitful tension, the right gap between what has been acquired and what can be achieved.
For a long time, this adjustment was an art. An art nourished by observation, experience and professional intuition. The teacher observes a hesitation, perceives a tension, hears a silence too dense to be harmless. He reformulates, slows down, speeds up, suggests an analogy, changes register. His gesture is not pure improvisation: it draws on theoretical knowledge, models and conceptual frameworks. But it remains situated, embodied, relational. It presupposes a fine-tuned reading of the situation, the learner's history, group dynamics and emotional climate.
Today, this long-standing ambition meets a new technology. Adaptive platforms, digital learning environments and pedagogical analytics systems promise to automatically adjust difficulty to each learner. Response time, error rate, frequency of attempts, regularity of connections: these are all digital traces that feed predictive models capable of modulating the learning path in real time. The algorithm thus becomes the discrete architect of continuous personalization.
The promise is seductive. Where teachers have to deal with numbers, fatigue and time constraints, machines process massive volumes of data without respite. It identifies correlations invisible to the human eye, detects early dropout tendencies, instantly adjusts the difficulty of an item.
From the perspective of bounded rationality, as described by Herbert Simon, delegating certain decisions to computational systems appears to be an effective strategy: locally optimizing learning trajectories on the basis of objectivized indicators. At first glance, the algorithm thus seems to realize the pedagogical dream of permanent, individualized adjustment, free from subjective bias. It promises a new kind of fairness: everyone progresses at their own pace, according to their own needs, without being compared to the group average. It also promises greater efficiency: less time wasted on exercises that are too simple, less giving up when faced with prematurely complex tasks. Adaptation becomes dynamic, continuous and calculated.
But there's a question: does calibrating a challenge boil down to optimizing performance indicators? Can the adjustment "to its measure" be entirely deduced from behavioral data? In other words, is what the algorithm measures sufficient to describe what is transforming a learner?
Flow theory, developed by Mihaly Csikszentmihalyi, has shown that optimal engagement arises from a delicate balance between challenge and perceived competence. But perceived competence is not strictly reducible to observed performance. It involves confidence, feelings of self-efficacy and the meaning attributed to the task. Similarly, John Sweller's cognitive load theory reminds us that the objective difficulty of an activity does not always coincide with the load actually borne by the learner. An anxiety-provoking context, ambiguous instructions or evaluative pressure can increase the load beyond what the metrics suggest.
Thus, algorithmic adjustment operates mainly on observable variables: accuracy of responses, speed, repetition or sequential progression. It optimizes training trajectories. But learning is more than just passing a series of items; learning means transforming one's representations, integrating knowledge into a network of meanings and developing a professional identity.
Herein lies the tension at the heart of this article. Algorithms can undeniably support pedagogical adjustment. It can reveal regularities, objectify difficulties and warn of fragilities. But it does not perceive the trembling of a voice during a clinical simulation, nor the hesitant posture of a student facing his first responsibility, nor the silent dynamics of a group comparing and judging each other. He individualizes from data; the pedagogue contextualizes from a situation.
The philosophy of technology, notably that of Gilbert Simondon, invites us to look beyond the simplistic opposition between human and machine. Technology is not the enemy of the human being; it extends his or her capacities, but it also transforms the conditions of action. When calibration becomes automatable, the risk is not so much the disappearance of the pedagogue as the silent redefinition of his role. If adjustment is entrusted to algorithms, what remains of pedagogical discernment? Marginal supervision? Formal validation? Or, on the contrary, greater responsibility for interpretation and meaning-making?
The challenge, then, is not to pit human expertise against algorithmic adaptation, but to clarify their boundaries. Optimizing learning trajectories does not necessarily mean promoting meaningful progress. Measurable performance does not always coincide with intellectual or professional maturation. To confuse the two would be to reduce pedagogical complexity to a statistical engineering problem.
In the age of adaptive platforms, "tailoring" can no longer be thought of as an exclusively intuitive gesture; nor can it be abandoned to computational logic alone. We need to redefine the art of calibration: understanding what the algorithm does well - and sometimes better than us - recognizing what it cannot grasp, and assuming the irreducible share of judgment, interpretation and responsibility that remains human.
For behind the technical question lies a deeper one: what is learning, and what do we really want to adjust? A score? A success rate? Or the singular trajectory of a subject in the making? It's on this shifting, demanding frontier, between optimization and discernment, that the reflection proposed here takes place.
The rise of adaptive platforms and digital learning environments is no mere technological fad. They are part of a profound transformation: the now tangible possibility of continuously collecting, processing and exploiting traces of pedagogical activity. Every click, every response, every latency becomes data. Learning leaves an exploitable digital footprint. From these traces, the algorithm adjusts.
This first movement deserves to be analyzed without caricature. Because algorithmic adjustment is not just a marketing promise. It relies on robust statistical models and sophisticated engineering that effectively transform the conditions of personalization.
Adaptive systems operate according to an apparently simple logic: measure to modulate. Response time to an item, error rate on a given skill, regularity of connections, persistence after failure, sequence of attempts - these are all indicators that feed into predictive models. The aim is not just descriptive, it's prescriptive. From past data, the system infers the probability of future success and adjusts the difficulty accordingly.
From this perspective, learning becomes an optimizable trajectory. Each learner is positioned in a probability space: which notions have been mastered? Which are fragile? What level of challenge maximizes the probability of engagement and success? The algorithm seeks the point of statistical equilibrium, where error is sufficiently present to stimulate effort, but not so present as to discourage it.
This logic refers to instrumental rationality in the sense of Herbert Simon: faced with the complexity of reality, we build simplified models that enable us to make effective decisions in an uncertain environment. The algorithm does not claim to understand the learner in his or her entirety; it processes observable variables and optimizes under constraint. Its horizon is the measurable improvement of indicators defined upstream.
From this point of view, pedagogy becomes in part a science of flows: flows of responses, flows of progress, flows of remediation. Adjustments are made in real time, without delay, without fatigue, without mood swings. On a large scale, this capability is an undeniable asset.
One of the most powerful arguments in favor of algorithmic adaptation is the promise of objectivity. Where teachers can be influenced - consciously or unconsciously - by expectation bias, subjective impressions or implicit comparisons between learners, the machine applies identical rules to all. It doesn't "prefer" anyone. It calculates.
In mass or hybrid training contexts, this standardization can help reduce certain inequalities. The fast learner is no longer slowed down by the pace of the group; the learner in difficulty benefits from additional targeted exercises without being publicly exposed to his or her backwardness. The learning path becomes singular, almost intimate.
What's more, the algorithm makes it possible to make adjustments of a granularity inaccessible to human beings. A trainer may perceive that a student is "struggling" with a notion. The system, on the other hand, precisely identifies the micro-skills involved: conceptual confusion, incomplete automation, incomplete transfer. It modulates the level of difficulty according to the item, adjusts the frequency of revision, and activates optimized spaced repetition.
From this perspective, algorithmic adaptation appears as an extension of human pedagogical capacity. It offers simultaneous panoramic and micro-analytical vision. It transforms pedagogical management into a process based on fine-grained, continuously updated data.
Yet it is precisely in this efficiency that a conceptual shift is taking place. Optimization is primarily concerned with observable performance: passing an item, improving a score, reducing a response time, stabilizing an error rate. In other words, the algorithm adjusts what is measurable.
But short-term success is not always synonymous with sustainable learning. A learner can rapidly improve his performance by familiarizing himself with an exercise format without radically transforming his representations. They may optimize their response strategy without consolidating their conceptual understanding. The algorithm, based on available data, tends to reinforce what improves the indicators.
It is here that we observe a subtle tension: the platform optimizes trajectories of exercise, but pedagogy aims at trajectories of transformation. The difference may seem slight, but it is decisive. Optimizing a score is a matter of performance. Transforming a relationship with knowledge is a longer, more complex and less immediately measurable process.
The risk is not that the algorithm will make a mistake - it can be extremely efficient in its register - but that it will insensitively impose its register as the norm. When dashboards become central, when indicators structure decision-making, the temptation is great to confuse metric improvement with real progress. What is not visible in the data tends to lose its value in the eyes of the system.
Algorithmic fine-tuning operates according to a coherent rationality: measure, model, optimize. It excels in pattern detection, fine-tuning of difficulty and large-scale individualized management. It is a powerful tool for pedagogical engineering.
But precisely because it is powerful, it calls for critical vigilance. For what the algorithm sees and adjusts corresponds to the variables it can capture. But the learning experience goes beyond these variables.
Herein lies the decisive question: what happens in the unmeasured space? What happens to the fit when we move away from indicators and into relationships, affect, group dynamics and the individual history of the learner?
What the algorithm optimizes is measurable. What the pedagogue perceives often goes beyond measurement. It's towards this shadow zone - not obscure, but unquantified - that analysis must now turn.
While algorithms excel at processing observable variables, pedagogues work in a depth of signs that goes beyond the digital trace. Where the machine captures performance, the teacher perceives situations. Where the algorithm adjusts a level of difficulty, the teacher adjusts a relationship, a climate, a sense of trust.
It's not a question of opposing rationality and intuition, but of recognizing that pedagogical activity mobilizes registers that data only partially capture. Adjusting "to measure" is not just about calibrating a task, it's also about reading a subject in the making.
In a training room, whether physical or virtual, learning takes the form of micro-events. A glance that turns away when a question is asked. A hesitation that lasts too long before speaking. A body posture that closes. A nervous laugh in the face of a situation.
These weak signals are not anecdotal. They are interpretative clues. The experienced pedagogue observes them not as isolated data, but as configurations. He cross-references posture, context, learner history and group dynamics. He or she infers fragility, anxiety, and sometimes a lack of understanding masked by correct performance.
To date, no adaptive platform has fully grasped this embodied dimension of learning. Even when enriched with sophisticated behavioral analyses, data remains an abstraction: it describes interactions with a device, not the totality of the lived experience.
Pedagogical discernment rests precisely on this ability to articulate heterogeneous elements - verbal, non-verbal, contextual - in a situated interpretation. It's less a question of "intuition" in the vague sense of the word, than of an expertise derived from accumulated experience, comparable to that of a clinician who recognizes a configuration even before being able to make it formally explicit.
The flow theory developed by Mihaly Csikszentmihalyi reminds us that optimal commitment depends on the balance between challenge and perceived competence. But perceived competence is not always correlated with observed performance.
A student may pass an exercise while deeply doubting its legitimacy. Another may fail momentarily while retaining strong confidence in his ability to progress. The algorithm sees success or error; the teacher perceives the affective tone that accompanies them.
In vocational training, in particular, this dimension is crucial. Learning involves an identity under construction. Students don't just want to "pass an item"; they want to become competent, responsible and recognized.
Pedagogical adjustment must therefore take into account symbolic elements: the fear of doing badly, the fear of judgment, the relationship with authority, previous school history. These are all dimensions that influence how a challenge will be experienced. A task that is objectively accessible can be subjectively threatening. Conversely, a demanding challenge can become a driving force if confidence is sufficiently established.
John Sweller's cognitive load theory emphasizes that the difficulty of a task does not depend solely on its intrinsic structure. The emotional and motivational context modifies the load actually borne. But these parameters are largely beyond the reach of quantitative metrics.
So, adjusting difficulty is not just a matter of calibrating a level of conceptual complexity. It also means regulating a psychological dynamic. Sometimes, it means lightening the pressure before increasing the demand. Sometimes, on the contrary, it means provoking slight discomfort in order to raise awareness. This fine-tuning is based on a situational reading, impossible to deduce from an isolated score.
The algorithm individualizes. The teacher contextualizes.
In the classroom, learning never takes place exclusively at the individual level. It is permeated by collective dynamics: implicit comparison, solidarity, rivalry, emotional contagion. A student in difficulty may be supported by the group or, on the contrary, feel exposed to it. An individual success can strengthen or weaken cohesion.
Human calibration integrates this systemic dimension. Adjusting a challenge for one learner sometimes means considering the effect on others. Introducing a more complex activity may stimulate the whole group. Maintaining a slower pace can make a fragile climate more secure.
Gilbert Simondon's philosophy of technique reminds us that technical objects function according to their own operating regimes. Algorithms operate within a formalized framework, processing defined variables. Humans, on the other hand, operate in an open, evolving, unpredictable situation.
This difference is not a fault of the machine, but rather a mark of its nature. But it does mean that algorithmic calibration cannot claim to exhaust pedagogical complexity. Adjusting a training challenge involves taking into account ethical issues, future responsibilities and institutional expectations. These are qualitative, normative and situated dimensions.
The pedagogue is not content with optimizing an individual trajectory; he or she must ensure the coherence of a pathway, the maturation of a collective, and the overall significance of the experience. His discernment is exercised in a plurality of registers that data only partially captures.
Ultimately, what the pedagogue sees goes beyond what the data records. Not because the machine is incapable of progress, but because learning is an embodied, relational and symbolic phenomenon.
Algorithms individualize from traces. The pedagogue contextualizes from a living situation. This distinction opens up a fertile tension: how can we combine the efficiency of statistical individualization with the depth of human contextualization?
It is precisely this tension that leads us to examine more closely the difference between algorithmic individualization and pedagogical contextualization. Because "tailoring" does not simply mean adapting a level of difficulty. It also means understanding what it means to learn in a given context.
If the algorithm individualizes and the pedagogue contextualizes, then the question becomes more precise: what does this change in scale change? Individualizing consists in adjusting a course on the basis of data specific to a subject. Contextualizing, on the other hand, implies placing the subject in a situation, a history, a collective, a professional horizon.
This shift is decisive. Adjusting a challenge is not simply a matter of calibrating a cognitive difficulty; it involves understanding the meaning of learning in a given environment.
Adaptive systems are based on segmentation logic. From the traces collected, they build profiles: fast but imprecise learners, regular but slow learners, high-performance memorizers but fragile transferers, etc. These profiles can be used to orientate learning programs. These profiles can be used to guide differentiated learning paths.
In technical terms, this often involves clustering, Bayesian modeling and predictive neural networks. The aim is to identify recurring patterns so as to offer the most relevant resource at the right time.
In this logic, individualization is a local optimization. We adjust an item, a series of exercises, a revision rhythm. The environment becomes adaptive: it reacts to past performance to anticipate future performance.
This approach has undeniable benefits. It avoids uniformity in learning paths. It reduces the time spent on skills already mastered. It precisely targets identified gaps. It can even encourage commitment by maintaining a statistically optimal level of challenge.
But this individualization remains focused on the interaction between an individual and a system. It treats the learner as a relatively isolated analytical unit. It only takes into account what can be modeled within the framework of the device.
But learning is more than just interacting with a sequence of exercises.
Contextualizing means changing focus. It's not just about adjusting a level of difficulty; it's about understanding why and for what we're adjusting.
Particularly in professional training, an exercise is never neutral. It prepares you for real-life situations where responsibility, uncertainty and the relational dimension are central. Adjusting a challenge in this context implies integrating symbolic issues: the relationship with error, stress management, confrontation with the unpredictable.
An algorithm can reduce the difficulty of a simulated case if the answers are incorrect. But the teacher may decide, on the contrary, to maintain a certain complexity to confront the learner with the reality of the field. He or she may decide that what's at stake is not just immediate success, but learning to manage uncertainty.
Contextualization is an interpretative act. It is based on a situated understanding: who is this learner? Where is he or she in the process? What are the group dynamics? What are the medium-term institutional and professional objectives?
It is also an ethical act. Because adjusting a challenge means guiding a trajectory. It means deciding on the level of demand to which a trainee is exposed. Systematically reducing difficulty to the slightest sign of failure can undermine the ability to cope with complex situations. Conversely, maintaining too high a level of challenge can lead to discouragement and withdrawal.
Pedagogical discernment consists precisely in navigating this tension, taking into account dimensions that algorithmic modeling only partially captures: professional maturity, reflective posture, ability to cooperate, relationship to authority and responsibility.
John Sweller's theory of cognitive load warns against overload, which impedes learning. Adaptive systems can usefully reduce this overload by calibrating task complexity.
However, in some training programs, the aim is not simply to master content in a stable environment, but to develop the ability to act in uncertain contexts. Training a professional is not simply about optimizing performance on graded exercises; it's about gradually exposing him or her to the complexity of the real world.
Over-protective algorithmic adaptation could unintentionally smooth out the learning experience. By constantly seeking the optimal balance between difficulty and immediate success, it risks minimizing exposure to the unexpected, ambiguity and constructive frustration.
Yet professional competence includes the ability to tolerate uncertainty, persevere in the face of failure, and mobilize resources in non-routine situations. These dimensions are not easily modeled by performance indicators.
Pedagogical contextualization makes it possible to introduce this complexity progressively and intentionally. It doesn't just react to mistakes. It anticipates future situations. It articulates the present of the exercise and the horizon of the profession.
In this way, the shift from individualization to contextualization is not just a terminological nuance. It marks a difference in nature: one optimizes local interactions, the other guides a trajectory within a symbolic, collective and professional framework.
Ultimately, the algorithm personalizes based on profiles, while the pedagogue contextualizes based on a project. One adjusts parameters, the other constructs meaning.
This is not to deny the relevance of algorithmic individualization, but to recognize that it operates in a specific register: that of measurable optimization. Contextualization, on the other hand, involves an interpretative and normative judgment that goes beyond observable performance alone.
From then on, the question was no longer to choose between individualization and contextualization, but to organize their articulation. How can we harness the analytical power of adaptive systems without reducing pedagogical calibration to a statistical function? How can we preserve human discernment in increasingly instrumented environments?
It is this search for critical complementarity that we must now explore.
If the algorithm optimizes and the pedagogue contextualizes, the decisive question is no longer which one should prevail, but how to distribute the roles without confusion. The danger lies not in the existence of adaptive systems, but in their naturalization. When statistical optimization becomes the implicit norm of "good fit", human discernment risks being reduced to peripheral validation.
Thinking in terms of critical complementarity means clarifying the boundaries: what algorithms do better, what educators must continue to assume, and what is a matter of demanding co-elaboration.
In its most relevant function, the algorithm acts as a revealer. It reveals trends that are difficult to detect with the naked eye: stagnation in a micro-skill, irregularity in commitment, gradual decline in performance over a given sequence.
In large cohorts or hybrid systems, this capability is a major asset. Learning analytics offer a dynamic mapping of learning. They enable early identification of dropout risks, target remedial needs and objectify certain pedagogical intuitions.
From this perspective, the machine does not replace judgment, but rather enhances it. It broadens the trainer's perceptive field. Where human observation is punctual and situated, analytics is continuous and exhaustive in its register.
This complementarity can enhance the quality of pedagogical management. A trainer can, for example, cross-reference his or her qualitative impressions with data from the platform. A hesitation perceived during a session may be reflected in a measurable drop in performance. Conversely, a stable performance may challenge an overly pessimistic intuition.
The algorithm thus becomes a diagnostic tool. It supports, but does not replace, a reflective approach.
The boundary becomes problematic when the tool prescribes rather than enlightens. If the system automatically decides on courses without human mediation, there is a risk of confusing statistical recommendations with pedagogical decisions.
But interpreting data does not mean following it mechanically. A drop in performance can signal a conceptual difficulty, but also temporary fatigue, emotional overload or disengagement linked to factors outside the system.
The trainer, as interpreter, recontextualizes the indicators. He places them in a story, an exchange, a real-life situation. He may decide not to follow an algorithmic recommendation if it conflicts with his overall analysis.
In this redefinition of roles, the trainer becomes less the exclusive dispenser of content and more the architect of scenarios. He or she designs hybrid systems in which the algorithm manages certain technical dimensions - spaced repetition, difficulty adjustment, automated training - while synchronous time is devoted to analysis, debate or the creation of complex situations.
Human added value shifts towards the clarification of meaning, relational regulation and reflective support. The art of calibration does not disappear. It changes level. It no longer focuses solely on the difficulty of an exercise, but on the orchestration of a course.
This complementarity is not spontaneous, however. It requires acculturation. Educational data are not neutral; they are constructed according to models, hypotheses and technical choices.
Understanding what an indicator actually measures, what it ignores, what assumptions it incorporates, becomes a professional skill in its own right. Reading a dashboard requires specific digital literacy: knowing how to distinguish correlation and causality, immediate performance and lasting consolidation, statistical average and singular trajectory.
Algorithmic rationality is based on a necessarily simplified formalization of reality. It selects certain variables to the detriment of others. This selection implicitly structures the definition of success.
Training trainers in this critical reading means preserving their capacity for discernment. It's not a question of rejecting data, but of questioning it. It's about considering data as part of a broader diagnosis, not as a definitive verdict.
From this perspective, Gilbert Simondon's philosophy of technology offers valuable insights: technology is not external to the human being; it is part of a process of individuation. But this integration presupposes an understanding of the operating regimes of the technical object. The tool must be understood if it is to be integrated without alienating.
At the heart of this critical complementarity lies a question of responsibility. Adjusting a challenge is not a neutral act. It engages a learner's trajectory, sometimes his confidence, sometimes his professional orientation.
To delegate this calibration entirely to a statistical system would be to shift responsibility to a system whose criteria are set upstream. But education cannot offload this responsibility entirely onto a calculation.
Preserving human responsibility does not mean rejecting useful automation. It does mean maintaining a space for interpretative decision-making. A space where recommendations can be suspended, discussed and contextualized.
Thus, critical complementarity does not consist in juxtaposing human and machine, but in articulating optimization and judgment. The algorithm can optimize sequences; the pedagogue guarantees the coherence, meaning and purpose of the course.
As we redefine the boundaries of calibration, it becomes clear that what's at stake goes beyond the simple organization of roles. What's at stake behind algorithmic adjustment is not just a question of pedagogical efficiency, but a very conception of learning and success.
Is optimizing measurable performance the same as training a subject capable of discernment? Is statistical improvement enough to qualify meaningful progress?
This is the more fundamental, almost philosophical question that we must now confront.
As adaptive devices are refined and predictive models perfected, a deeper question emerges. Calibration is not just a technical operation. It involves a certain conception of learning, of the subject and of success.
Optimizing an exercise trajectory does not necessarily mean training a subject capable of judgment. Adjusting a difficulty according to performance indicators says nothing, in itself, about the inner transformation that constitutes the act of learning.
Behind the technical question of the "right adjustment" lies a philosophical question: what are we really trying to adjust?
Adaptive platforms function according to a logic of optimization: reduce error, increase success, stabilize observable skills. This logic is coherent and legitimate. It enables measurable gains. It makes progress visible.
But measurable performance does not exhaust meaningful progress. Learning is not just about improving a score; it's about transforming one's relationship to knowledge, to oneself and to the world.
A learner can achieve high scores without having integrated knowledge into a stable conceptual framework. They may succeed at familiar tasks without being able to transfer their knowledge to a new context. Conversely, a phase of apparent stagnation may correspond to a profound restructuring of representations.
The zone of proximal development described by Lev Vygotsky cannot be reduced to a statistically optimal range of difficulty. It presupposes mediation, meaningful interaction and social dynamics.
When optimization becomes the dominant metric, the risk is to privilege what is immediately quantifiable to the detriment of what transforms more slowly: critical thinking, reflective posture, the ability to problematize.
Algorithmic calibration can thus produce an illusion of mastery. Because the curve is rising, we assume that learning is progressing. But not all numerical progress is necessarily intellectual maturation.
The efficiency of adaptive systems can give rise to a subtle temptation: that of gradually delegating the art of calibration. If the machine adjusts faster, more finely and more objectively, why maintain systematic human mediation?
This temptation is part of a broader rationalization drive. In environments constrained by time, numbers and institutional requirements, automation appears to be a pragmatic solution. It promises consistency, traceability and continuous optimization.
But to delegate grading entirely is to shift the pedagogical standard to the underlying statistical model. Success criteria become those that have been programmed. The implicit aims of the system, often oriented towards observable performance, silently structure the learning experience.
The bounded rationality described by Herbert Simon finds a new expression here: faced with complexity, we simplify. But any simplification selects certain parameters and excludes others. When calibration is based primarily on quantifiable indicators, what is not quantified risks being marginalized.
The challenge is not to reject automation, but to prevent it from becoming an exclusive principle. Learning involves ethical, symbolic and relational dimensions that are not easily formalized.
The art of the pedagogue lies not only in the transmission of content, but also in discernment. Discernment means interpreting a singular situation in the light of a broader horizon. It means deciding when to support, when to confront, when to slow down, when to speed up.
This discernment is as much an ethical act as a technical one. It involves a responsibility: that of exposing a learner to a level of demand that fosters his or her growth without unnecessarily weakening it.
The technical object must not be opposed to the human being, but integrated into a process of individuation. This presupposes a conscious, critical relationship with technology. The tool becomes fruitful when it is understood and mastered, not when it implicitly dictates the norm.
Preserving the art of discernment doesn't mean rejecting learning analytics; it means refusing to let them become our sole compass. It's about maintaining a space for judgment, where the pedagogical decision is not reduced to an algorithmic recommendation.
In an increasingly instrumented environment, the trainer's key skill could become precisely this ability to articulate data and interpretation, optimization and meaning, efficiency and responsibility.
The "fit" cannot be entirely calculated, because it concerns a subject in the making, inserted into a living context. Data describes part of this reality; it does not exhaust it.
At the end of this exploration, one conviction emerges: calibration is not just an engineering problem. It is a question of purpose. Do we want to optimize performance or support transformation? Are we looking for the most efficient or the most meaningful trajectory?
These questions don't disqualify algorithms; rather, they invite us to place them within a broader architecture in which human expertise remains the guarantor of meaning. Only then can technological adjustment become a shared resource, rather than a silent delegation of the art of teaching.
The algorithm adjusts. The teacher interprets. This is not a duel, but a redefinition of the boundaries of pedagogical calibration.
Adaptive platforms have undeniably transformed our ability to personalize learning paths. They measure with unprecedented precision, detect invisible regularities, and optimize training sequences on a scale impossible to sustain by human observation alone. In massive, hybrid or constrained contexts, this analytical power is a real lever for improvement. But optimizing is not the same as discerning.
Algorithmic optimization operates in a formalized space: it adjusts according to defined, selected and weighted indicators. It maximizes the probability of success in the short or medium term. It excels at managing calculable complexity.
Pedagogical discernment, on the other hand, is exercised in the context of lived complexity. It combines performance and confidence, difficulty and meaning, measurable progress and inner maturation. It integrates relational, symbolic and collective dimensions. It takes into account the context, the moment and the singular history of a learner.
Pedagogical data describes. It does not understand. It signals. It does not determine meaning.
Herein lies the central challenge: not to confuse measurable performance with meaningful progress. A rising curve is no guarantee of profound transformation. A controlled error rate alone says nothing about the ability to transfer, to problematize, to act in the face of uncertainty. Conversely, a phase of apparent fragility may correspond to a fruitful restructuring.
In the age of digital environments, the temptation is great to delegate the art of calibration to statistical systems. After all, they are fast, consistent and traceable. They seem to offer a reassuring objectivity. But to delegate adjustment entirely would be to shift the pedagogical norm to what can be calculated.
But not everything that counts in training can be counted.
Adjustment "to one's own measure" cannot be fully calculated, because it concerns a subject in the making. It presupposes a situated reading, a mediation, a responsibility. It engages an ethic of challenge: to expose without crushing, to support without over-protecting, to demand without discouraging.
So the real question is not: human or machine? It's: how do we articulate optimization and discernment without one absorbing the other?
A demanding complementarity is possible and even desirable. The algorithm can become a diagnostic tool, revealing trends and supporting differentiation. The trainer, on the other hand, remains an interpreter, an architect of scenarios, a guarantor of meaning and purpose. One broadens the perceptive field; the other guides the trajectory.
From this point of view, the key skill of the years to come will perhaps be neither purely technical nor purely pedagogical, but hermeneutic: knowing how to read data without submitting to it, knowing how to integrate indicators without reducing the complexity of learning to metrics.
As devices become more intelligent, human responsibility does not diminish; it shifts. It consists less in producing each adjustment than in deciding what deserves to be adjusted and why.
The optimal fit can be calculated. The right challenge, on the other hand, must be discerned.
This is perhaps where the art of the pedagogue lies in the face of the algorithm: not to compete with the machine on its own ground, but to keep the question of meaning open, so that technology remains a means and never the ultimate measure of what learning means.
Illustration: The art of the pedagogue
Generated by AI (Canva) - Flavien Albarras
References
Alhazbi, S., Al-ali, A., Tabassum, A., Al-Ali, A., Al-Emadi, A., Khattab, T., & Hasan, M. A. (2024). Using learning analytics to measure self-regulated learning: A systematic review of empirical studies in higher education. Journal of Computer Assisted Learning, 40(4), 1658-1674. https://doi.org/10.1111/jcal.12982
Anderson, J., & Taner, G. (2023). Building the expert teacher prototype: A metasummary of teacher expertise studies in primary and secondary education. Educational Research Review, 38, 100485. https://doi.org/10.1016/j.edurev.2022.100485
Artino, A. R. (2012). Academic self-efficacy: From educational theory to instructional practice. Perspectives on Medical Education, 1(2), 76-85. https://doi.org/10.1007/s40037-012-0012-5
Azevedo, R., Bouchet, F., Duffy, M., Harley, J., Taub, M., Trevors, G., Cloude, E., Dever, D., Wiedbusch, M., Wortha, F., & Cerezo, R. (2022). Lessons Learned and Future Directions of MetaTutor: Leveraging Multichannel Data to Scaffold Self-Regulated Learning With an Intelligent Tutoring System. Frontiers in Psychology, 13. https://doi.org/10.3389/fpsyg.2022.813632
Bandura, A. (1978). Self-efficacy: Toward a unifying theory of behavioral change. Advances in Behaviour Research and Therapy, Perceived Self-Efficacy: Analyses of Bandura's Theory of Behavioural Change, 1(4), 139-161 . https://doi.org/10.1016/0146-6402(78)90002-4
Biggs, J. B. (1987). Student Approaches to Learning and Studying. Research Monograph. Australian Council for Educational Research Ltd. https://eric.ed.gov/?id=ED308201
de Vries, M. J., Feenberg, A., De Boever, A., & Hoel, A. S. (2015). Book Symposium on The Philosophy of Simondon: Between Technology and Individuation. Philosophy & Technology, 28(2), 297-322. https://doi.org/10.1007/s13347-013-0144-5
Kosel, C., Bauer, E., & Seidel, T. (2024). Where experience makes a difference: Teachers' judgment accuracy and diagnostic reasoning regarding student learning characteristics. Frontiers in Psychology, 15, 1278472. https://doi.org/10.3389/fpsyg.2024.1278472
Lan, M., & Zhou, X. (2025). A qualitative systematic review on AI empowered self-regulated learning in higher education. Npj Science of Learning, 10(1), 21. https://doi.org/10.1038/s41539-025-00319-0
Learning Analytics as a Metacognitive Tool: (2014). Proceedings of the 6th International Conference on Computer Supported Education, 380-384. https://doi.org/10.5220/0004933203800384
Lodge, J. M., & Corrin, L. (2017). What data and analytics can and do say about effective learning. Npj Science of Learning, 2(1), 5. https://doi.org/10.1038/s41539-017-0006-5
Mai, L., Köchling, A., & Wehner, M. C. (2022). "This Student Needs to Stay Back": To What Degree Would Instructors Rely on the Recommendation of Learning Analytics? SN Computer Science, 3(4), 259. https://doi.org/10.1007/s42979-022-01137-6
Marton, F., & Säljö, R. (1976). On Qualitative Differences in Learning: I-Outcome and Process. British Journal of Educational Psychology, 46(1), 4-11. https://doi.org/10.1111/j.2044-8279.1976.tb02980.x
Perone, C. S. (2025, January 9). Notes on Gilbert Simondon's "On the Mode of Existence of Technical Objects" and Artificial Intelligence | Terra Incognita. https://blog.christianperone.com/2025/01/notes-on-gilbert-simondons-on-the-mode-of-existence-of-technical-objects-and-artificial-intelligence/
Perrenoud, P., & Meirieu, P. P. (2026). La pédagogie différenciée: Des intentions à l'action. ESF éditeur.
Porayska-Pomsta, K., Holmes, W., & Nemorin, S. (2022). The Ethics of AI in Education. https://doi.org/10.4324/9780429329067
Raffaghelli, J. E., Manca, S., Stewart, B., Prinsloo, P., & Sangrà, A. (2020). Supporting the development of critical data literacies in higher education: Building blocks for fair data cultures in society. International Journal of Educational Technology in Higher Education, 17(1), 58. https://doi.org/10.1186/s41239-020-00235-w
Schön, D. A. (2017). The Reflective Practitioner: How Professionals Think in Action. Routledge. https://doi.org/10.4324/9781315237473
Schwarz, G., Christensen, T., & Zhu, X. (2022). Bounded Rationality, Satisficing, Artificial Intelligence, and Decision-Making in Public Organizations: The Contributions of Herbert Simon. Public Administration Review, 82(5), 902-904. https://doi.org/10.1111/puar.13540
Selwyn, N., & Gašević, D. (2020). The datafication of higher education: Discussing the promises and problems. Teaching in Higher Education, 25(4), 527-540. https://doi.org/10.1080/13562517.2019.1689388
Simon, H. A. (1955). A Behavioral Model of Rational Choice. The Quarterly Journal of Economics, 69(1), 99-118. https://doi.org/10.2307/1884852
Simon, H. A. (1981). The sciences of the artificial.
Simon, H. A. (1991). Bounded Rationality and Organizational Learning. Organization Science, 2(1), 125-134.
Simondon, G., Mellamphy, N., Mellamphy, D., & Mellamphy, N. B. (2011a). On the Mode of Existence of Technical Objects. Deleuze Studies, 5(3), 407-424.
Simondon, G., Mellamphy, N., Mellamphy, D., & Mellamphy, N. B. (2011b). On the Mode of Existence of Technical Objects. Deleuze Studies, 5(3), 407-424.
Xue, H., & Lin, L. (2026). The ethics of artificial intelligence in education: Practices, challenges, and debates: by Wayne Holmes, Kaśka Porayska-Pomsta (Eds.), New York, Routledge, 2023, XXIII+288 pp., £139.21 (hardcover), ISBN: 978-0-367-34971-4. Journal of Education for Teaching, 0(0), 1-4. https://doi.org/10.1080/02607476.2026.2634972
You, X., & Hands, D. (2019). A Reflection upon Herbert Simon's Vision of Design in The Sciences of the Artificial. The Design Journal, 22(sup1), 1345-1356. https://doi.org/10.1080/14606925.2019.1594961
You might also like:
From grade to value: assessment at the service of the learner
If we shift some of the responsibility for assessment onto the learner, we achieve a dual objective: making assessment a learning process and developing the learner's critical judgment.
Music at school, more than just culture
Music in schools is not just about culture. It is now complemented by other, more cross-disciplinary contributions. Memory, listening and concentration are all enhanced by learning and mastering music. From kindergarten to university, music offers a technical dimension, mastery of solfeggio and singing, and a second dimension that opens doors to literature or elements of personal development, complementing purely academic knowledge.
From master trainer to involved facilitator
4 key activities to promote the evolution of the trainer's profession towards the facilitator's posture, in order to move towards an approach to training that makes more room for the autonomous power to learn.
The notion of "frame of reference": the expansion of a practical idea
From a material approach to a leading concept in the humanities and education, explore the origins and expansion of the notion of framework.
Virtual reality and vocational training: between promise and vertigo
What does the shift from situation-based to simulation-based training entail? Virtual reality simulations are revolutionizing training, enabling risk-free practice despite the complexity of the situations involved. But their widespread use raises questions: what effect will they have on knowledge, professions and professional identities? The challenge is to integrate them with discernment, so as to enrich experience without losing the meaning of work.
