Di Simone Pfliegel (*)
Introduzione editoriale
Il criterio che mancava
C’è una frase che ogni docente europeo ha sentito ripetere: l’intelligenza artificiale deve sostenere il lavoro dell’insegnante, non sostituirlo.
Sta nelle circolari ministeriali, nei corsi di aggiornamento, nei quadri etici scritti per contenere le aziende del settore — e, osserva Simone Pfliegel, anche nei materiali con cui quelle stesse aziende vendono i loro strumenti.
Che tutti siano d’accordo dovrebbe metterci in sospetto. Un principio che nessun fornitore trova scomodo e nessun ministero trova costoso difficilmente sta facendo molto lavoro.
Ma l’obiezione dell’autrice non è che la formula sia sbagliata. È che non si può seguire.
Immaginate una docente, la domenica sera, con trenta elaborati davanti e quattro ore prima di lunedì. Un sistema le genera un commento individuale su tutti e trenta in novanta secondi. È supporto o è sostituzione? La formula non lo dice — e non lo dice perché descrive un rapporto fra due astrazioni, «l’intelligenza artificiale» e «l’insegnante», mentre la decisione reale riguarda un compito solo: questo commento, su questo testo, per questo studente, stasera.
Da qui muove la proposta che dà valore a questo contributo. Il confine, sostiene Pfliegel, non va tracciato dove il dibattito lo traccia di solito — sulla capacità del sistema, criterio che si sposta a ogni nuovo modello e che avvantaggia proprio chi quei modelli li vende.
Va tracciato sulla natura di ciò che il sistema produce. La domanda è una sola: quel risultato costituisce un’affermazione su una persona identificabile?
Ne discendono tre categorie.
Il lavoro generativo — la bozza di un esercizio, tre versioni di un testo a livelli diversi, una batteria di domande — produce materiale che un essere umano valuterà prima che tocchi qualcuno: si può delegare senza perdita, e conviene farlo, perché è lì che il tempo di lavoro se ne va davvero.
Il giudizio formativo è delegabile a condizione che l’insegnante possa ricostruire perché una proposta è stata formulata.
L’affermazione con conseguenze — un voto, un giudizio di competenza, l’individuazione di uno studente in difficoltà, un’accusa di disonestà — non è delegabile mai. E, aggiunge l’autrice con una torsione che merita attenzione, lo è tanto meno quanto più il sistema è bravo: perché un sistema competente è quello le cui proposte sono più difficili da rifiutare.
Il criterio diventa operativo attraverso due domande che una docente può porsi in trenta secondi: che cosa direi allo studente su come è stata presa questa decisione, e che cosa direi a un genitore o a un ispettore? Un compito che non sopravvive a entrambe sta dalla parte umana della linea.
Va segnalata la sezione sui rilevatori di testo generato, perché contiene un dato che ogni dirigente scolastico dovrebbe conoscere prima di firmare un acquisto. Nello studio di riferimento in materia, sette rilevatori commerciali hanno classificato con accuratezza quasi perfetta i testi di studenti madrelingua inglesi, mentre segnalavano come generati da macchina oltre il sessanta per cento dei testi scritti da non madrelingua. Il meccanismo non è accidentale: i rilevatori penalizzano la varietà linguistica limitata — penalizzano, cioè, il trovarsi a uno stadio più arretrato dell’apprendimento. In un’aula europea questo significa che sbagliano più gravemente proprio con gli studenti che hanno meno protezione istituzionale.
Il contributo si àncora a Magnifica Humanitas, l’enciclica di Leone XIV del maggio 2026, e ne trae — senza chiedere al lettore di condividerne i presupposti teologici — la domanda che il dibattito educativo aveva smesso di porre: non che cosa la tecnologia sappia fare, ma che cosa accade alla persona mentre lo fa. In una scuola, quella persona è uno studente su cui si dicono cose, che vengono registrate e che lo seguiranno.
Un’ultima nota, di quelle che si fanno di rado. L’articolo dedica una sezione alle obiezioni contro sé stesso, e la prima è la più scomoda: che questo sia un discorso da privilegiati, che una docente con ventisei ore di lezione in una scuola senza risorse non possa permettersi la valutazione di processo. L’autrice non la aggira. Risponde che il criterio è esattamente ciò che autorizza a delegare senza scrupoli tutto il lavoro di preparazione, perché il lavoro di giudizio possa essere fatto come si deve — e che la condizione peggiore fra quelle disponibili, oggi il default in molti sistemi scolastici, è quella che non concede né l’una né l’altra cosa.
Pfliegel chiude riprendendo i tre verbi che intitolano il nostro anno: coltivare, educare, umanizzare. Non sono tre compiti distinti, scrive. Descrivono un unico rifiuto — il rifiuto di lasciare che la questione di che cosa sia una persona venga risolta da ciò che risulta di volta in volta più comodo.
Carmine Marinucci
Abstract
Educational policy across Europe has converged on a reassuring formula: artificial intelligence should support teachers, not replace them. This article argues that the formula, however well intentioned, gives practitioners no usable guidance. It states a value without supplying a criterion, and it therefore collapses at precisely the moment a teacher has to decide whether a particular task may be handed to a machine. Drawing on the encyclical Magnifica Humanitas (2026), which frames artificial intelligence as an anthropological rather than a merely technical question, and on professional development work with more than five hundred teachers, the article proposes a criterion that can be applied at the level of the individual task. The boundary is not drawn by what a system is capable of producing, but by whether its output constitutes a claim about an identifiable learner. On that basis the article distinguishes three categories of work — generative labour, formative judgement, and consequential claim — and argues that only the first can be delegated without loss. It closes with implications for classroom practice, for school-level agreements, and for the design of teacher education.
Keywords
artificial intelligence in education; educational responsibility; assessment and academic integrity; teacher professional development; human oversight; Magnifica Humanitas
Abstract (italiano)
Le politiche educative europee convergono su una formula rassicurante: l’intelligenza artificiale deve sostenere gli insegnanti, non sostituirli. Questo articolo sostiene che tale formula, per quanto ben intenzionata, non offre ai docenti alcun orientamento utilizzabile. Enuncia un valore senza fornire un criterio e viene meno proprio nel momento in cui l’insegnante deve decidere se un determinato compito possa essere affidato a una macchina. A partire dall’enciclica Magnifica Humanitas (2026), che presenta l’intelligenza artificiale come questione antropologica e non meramente tecnica, e da un’esperienza di formazione rivolta a oltre cinquecento docenti, l’articolo propone un criterio applicabile al singolo compito. Il confine non è tracciato da ciò che un sistema è in grado di produrre, ma dal fatto che il suo esito costituisca o meno un’affermazione su una persona che apprende. Su questa base l’articolo distingue tre categorie di lavoro — lavoro generativo, giudizio formativo e affermazione con conseguenze — e sostiene che solo la prima possa essere delegata senza perdita. Si conclude con le implicazioni per la pratica didattica, per gli accordi a livello di scuola e per la formazione degli insegnanti.
Parole chiave
intelligenza artificiale nell’educazione; responsabilità educativa; valutazione e integrità accademica; formazione degli insegnanti; supervisione umana; Magnifica Humanitas
1. A formula that cannot be followed
Ask a teacher in almost any European school system what the official line on artificial intelligence is, and some version of the same sentence will come back: artificial intelligence should support the work of teachers, not replace it. The sentence appears in ministerial guidance, in professional development programmes, in the marketing materials of the companies selling the tools, and in the ethical frameworks written to restrain those companies. Its ubiquity is remarkable. So is the fact that almost nobody disagrees with it.
That unanimity should make us suspicious. A principle that everyone accepts, that no vendor finds inconvenient and no ministry finds costly, is unlikely to be doing much work.
The difficulty is not that the sentence is wrong. It is that it cannot be followed. Consider a teacher on a Sunday evening with thirty pieces of written work in front of her and four hours before Monday. A system will generate individual feedback on all thirty in ninety seconds. Is using it support, or is it replacement? The formula does not say. It does not say because it describes a relationship between two abstractions — “artificial intelligence” and “the teacher” — while the decision she actually faces is about one task: this feedback, on this text, for this student, on this evening.
The same gap opens everywhere. May a system draft the exercises? Suggest which aspects of a text to comment on? Propose the grade? Write the report that goes home to parents? Identify which students are at risk of failing? Each of these can be described as support by someone who wants to use it and as replacement by someone who does not. The formula adjudicates nothing.
What follows is an attempt to supply what the formula lacks: a criterion that operates at the level of the task, that a teacher can apply on a Sunday evening, and that does not require her to predict which tools will exist in three years.
2. The encyclical’s question
In May 2026 Leo XIV published his first encyclical, Magnifica Humanitas, on the safeguarding of the human person in the time of artificial intelligence. Signed on the hundred and thirty-fifth anniversary of Rerum Novarum, the document runs to two hundred and forty-five paragraphs and its concerns range from labour to warfare. For the argument developed here, three of its moves matter.
The first is a refusal of the two easy positions. The encyclical does not treat technology as an adversary of the person — it is, on the contrary, “rooted in our history from the beginning” — nor does it treat it as neutral. Technology “can heal, connect, educate, care for the common home; but it can also divide, discard, generate new injustices”. In the abstract it is neither a solution to humanity’s problems nor an evil in itself; concretely, however, “it is not neutral, because it takes on the face of those who conceive it, finance it, regulate it, use it”.1 This is a more demanding position than either enthusiasm or refusal, because it makes the design and governance of a system a moral question rather than a technical one, and it does so before any question about the intentions of individual users.
The second move is the one that concerns schools most directly, and it is sharper than the summaries of the encyclical have suggested. Describing what it calls the syndrome of Babel, the text names among its temptations “the claim of a single language — digital too — capable of translating everything, even the mystery of the person, into data and performance”.2 The warning is repeated in the chapter on technology and domination, where the encyclical takes up Paul VI’s formulation that one may come to “have more” without “being more”, with the consequence that the person “risks being evaluated above all on the basis of the performance she guarantees”.3
Read in an educational setting, this is not a devotional sentiment. It is a constraint on assessment. A school is an institution that exists, among other things, to translate persons into data and performance: that is what a grade is. The encyclical states elsewhere that the dignity of a person “does not depend on the capacities she possesses”,4 and identifies as particularly insidious the ideology suggesting that each person must earn or justify her own worth, so that greater value is attributed to those who are more efficient and higher-performing — with the result that the person “ends up being reduced to a means for obtaining results, a resource to be used and exploited, and is no longer recognised as an end in herself”.5 If that is right, then the making of claims about a learner’s performance carries a weight that cannot simply be handed to whatever process is cheapest.
The third move follows from the second. Facing the concentration of power in the digital world, the encyclical proposes that the principles of Catholic social teaching — the inalienable dignity of the person, the common good, the universal destination of goods, subsidiarity, solidarity and social justice — become criteria for judging and discerning the new landscape.6 The ethical question is thereby moved upstream, from use to design and governance: it is not enough to ask whether a system is being used well.
One need not share the encyclical’s theological commitments to take its question seriously, and this article does not argue from them. What the document supplies is a formulation the educational debate has lacked: the question is not what artificial intelligence can do, but what happens to the human person when it does it. That is an anthropological question, and it is answerable in the specific setting of a classroom.
3. Three ways the formula fails in practice
Between 2023 and 2026 I have worked with more than five hundred teachers in professional development on generative artificial intelligence — in secondary schools, in adult and vocational education, and in the German-as-a-foreign-language networks of several European countries. The three patterns below are drawn from that work. They are not survey findings; they are recurring observations, and I report them as such, including the cases my own sessions produced.
3.1 Tool demonstration without task analysis
The most common shape of an in-service session on artificial intelligence is a demonstration. A trainer shows a tool, the tool produces something impressive, the participants are shown how to reproduce it. The session ends with enthusiasm and a list of applications.
What it does not end with is any analysis of the tasks the participants actually set. The demonstration operates on a generic exercise — a text nobody in the room has to teach, a class nobody in the room has to grade. Transfer is left as an exercise for the participant, on a Monday, alone.
Two observations recur. The first is a sentence I have heard at the close of sessions in almost this form: it was all very impressive, but with thirty-two learners in my class it is not workable. The judgement is not about the tool. It is about the distance between the demonstration and the room the participant has to walk back into.
The second is the request that most often ends a session: participants ask for the list of prompts to copy. They rarely ask for the procedure by which one arrives at a prompt oneself. The request is entirely reasonable from someone under time pressure, and it is also the clearest sign that the session has taught operation rather than judgement. A copied prompt works until the task changes; the procedure survives the task changing, and survives the tool changing too.
The failure here is not pedagogical laziness on anyone’s part. It is a category error about where the difficulty lies. The difficulty is not operating the tool. The difficulty is deciding which part of one’s own teaching the tool should touch — and that decision requires exactly the criterion this article is trying to supply.
3.2 Delegating judgement while believing one is delegating effort
The second pattern is more consequential and much less visible. A teacher sets out to save time on the labour of writing feedback and ends by handing over the judgement that the feedback expresses.
The slide is gradual and each step is defensible. First the system drafts the wording of a comment the teacher has already formed. Then it suggests which aspects of a text to comment on. Then it proposes a rating against the criteria. Then the teacher, under time pressure and finding the proposals reasonable, begins to accept them without independently forming the judgement they encode. Nothing at any single step looks like abdication. The end state is that a claim about a student’s competence has been produced by a process the teacher could not reconstruct if asked.
I have heard this described in a participant’s own words: she has the grade proposed and then corrects it, and she offered this as an example of time saved. She was not being careless. On her account the proposals were usually right, and the correction step preserved, as she saw it, her authority over the outcome. The difficulty is that a correction step applied to a judgement one did not independently form is not oversight; it is assent with an extra click.
An exercise in one session made the mechanism visible more sharply than any argument could. Participants were asked to write their own feedback on a student text, then to generate feedback with a system, then to compare the two. Several discovered in the course of the comparison that they had already agreed with the machine’s version — without having gone back to the student’s text. The agreement had been with a plausible piece of prose about a text, not with the text itself. That is the moment the delegation happens, and it happens quietly enough that it took an exercise to surface it.
This is the pattern the encyclical’s warning addresses most directly. What has been delegated is not work but the act of saying something about a person.
3.3 The detection fallacy
The third pattern is the attempt to solve a design problem after the fact. Confronted with the possibility that student work was produced by a system, institutions reach for detection: software that claims to identify machine-generated text, or informal heuristics about style and vocabulary.
Detection fails on two counts. Technically, the tools produce false positives at rates that are unacceptable when the consequence is an accusation of dishonesty, and the failure is not evenly distributed. In the best-known study of the problem, seven widely used detectors classified essays by native-speaking American school students with near-perfect accuracy while misclassifying more than half of the essays written by non-native speakers as machine-generated — a mean false-positive rate above sixty per cent, with almost every non-native essay flagged by at least one detector.7 The mechanism the authors identify is not incidental: the detectors penalise limited linguistic variety. In other words, they penalise being at an earlier stage of language learning. In most European classrooms that means they fail hardest for the students with the least institutional protection.
Two things I have observed follow from this. One teacher reported a student text she knew to be the student’s own work, written under her supervision, which a detector marked as machine-generated. What is striking is not the error, which the research predicts, but the position it puts the teacher in: she is asked to weigh her own knowledge of the student against a number produced by a system that offers no reasoning she can examine. Some teachers trust the number. That is the more dangerous case.
The second is a demand that surfaces regularly in sessions: that the school should finally purchase a reliable tool. The expectation behind it is that the problem is technical and awaits a technical solution — that somewhere there exists a detector accurate enough to restore the previous state of affairs. There is not, and there is unlikely to be, because the object being detected is by design a text that resembles human writing. But the demand deserves to be taken seriously rather than dismissed, because it expresses something real: teachers are being asked to carry a new professional risk without being given any new means to manage it. Detection is offered as the means. It does not work, and saying so is only half an answer. The other half is the redesign of assessment described below.
Pedagogically, detection also arrives too late by design. It operates on a finished product, at a point where the only available responses are punitive.
The three patterns have a common structure. In each case the teacher’s attention is on the tool, and the question of what is being given away goes unasked.
4. The criterion: what the output claims
The proposal of this article is to draw the line somewhere other than where the debate usually draws it.
The usual line is capability. It asks what the system can do well: if the system writes good feedback, let it write feedback; if it does not, do not let it. This line moves every time a new model is released, which means any guidance built on it expires. It is also the line the vendors are most eager for us to adopt, since every improvement in capability enlarges what may be delegated.
The alternative proposed here is accountability. The question is not what the system can produce, but what its output is. Specifically: does the output constitute a claim about an identifiable learner?
This yields three categories.
Generative labour. The system produces material that a human being will evaluate before it has any effect on anyone: a draft exercise, three variants of a text at different levels, a set of comprehension questions, a first version of a worksheet. The teacher reads it, judges it, and either uses it or does not. Errors are caught before they touch a student. This category can be delegated without loss, and the case for delegating it is strong: it is where the working time actually goes.
Formative judgement. The system informs a decision that remains open and revisable: a suggestion about which students appear to be struggling with a construction, a proposed focus for the next lesson, an indication that a text is above the level of a group. Here delegation is defensible, but conditionally. The condition is that the teacher can reconstruct why the suggestion was made, and that the decision remains reversible in practice and not merely in theory.
Consequential claim. The output asserts something about a named person that carries consequences for them: a grade, a report comment, a judgement of competence, a determination that a student needs support, an accusation of dishonesty. Here the criterion is categorical. The judgement is made by a human being who can be asked to justify it, or it is not made properly at all. This holds regardless of how good the system is — indeed it holds more urgently the better the system is, because a competent system is one whose proposals are hardest to refuse.
The strength of this criterion is that it does not move when the technology does. It does not require a teacher to know how a model works, or to track releases. It requires only that she ask, of a given task: at the end of this, will something have been said about one of my students?
Two sentences make the criterion operational. For any task a teacher is minded to hand over, she should be able to say what she would tell the student about how the decision was reached, and what she would tell a parent, an inspector, or a court. A task that cannot survive both sentences belongs on the human side of the line.
5. What changes in the classroom
The criterion has consequences that are more practical than a discussion of principles usually delivers.
Preparation becomes the site of delegation. If generative labour is where delegation is legitimate, then that is where professional development should concentrate — and it is, conveniently, where the workload actually is. The differentiation of a text into three levels, the production of exercise material in quantity, the generation of comprehension items from an authentic source: these are tasks whose outputs a teacher checks before use, and where the time saved is real.
The format I now use for this is deliberately unglamorous. Participants bring a text they actually have to teach. They produce three versions of it at different levels. Then they exchange versions with a colleague and check each other’s for errors. The errors they find are the point of the exercise — not an unfortunate by-product of it, but the content. Until a teacher has seen a plausible, fluent, wrong version of her own material, the instruction to verify remains an abstraction she will skip when tired.
What the exchange turns up is instructive. Factual slippage occurs, but the more insidious failure is register. In one session a participant’s factual text about a regional industry came back from the simplification step reading like advertising copy: the hedges had gone, the qualifications had gone, and what remained was enthusiastic, fluent and subtly untrue to the source. Nobody in the room had noticed on first reading. It reads well; that is precisely the problem. A simplification that is grammatically clean and tonally wrong will pass an inattentive check, and it teaches learners a register the original did not contain.
So the routine is not “generate”, it is “generate and verify”, and the verification is the part that requires the teacher.
Assessment design moves before the assessment. If consequential claims cannot be delegated, and if detection after the fact does not work, the remaining option is to design assessment so that evidence of a student’s own thinking is produced along the way rather than inferred at the end. Drafts and intermediate stages that are visible. A short conversation in which the student explains a choice she made. A revision protocol in which the student records what she changed and why. None of this is new — it is ordinary process-oriented pedagogy — but the arrival of generative systems has made it necessary rather than merely desirable.
The revision protocol has proved the most portable of these in my own teaching. Students submit not only the text but a short record of the changes they made between versions and their reasons. It costs the student ten minutes. It changes what the teacher is reading: no longer only a product whose provenance is uncertain, but a trace of decisions, which is much harder to outsource convincingly and much more interesting to assess.
The clearest evidence I have for the shift came from running the same task twice — once as a straightforward product submission, once with intermediate stages required. The products were of comparable quality. What differed was what became visible. In the second version it was apparent which students had struggled productively and revised, which had produced a good text in one pass, and which had a finished product with no history at all. The last group is not automatically dishonest; it is simply the group about whom the product alone permits no reliable claim. Under the criterion proposed here, that is decisive: the second design allows a defensible judgement, and the first does not.
The conversation becomes an assessment instrument. Where a piece of work has been produced partly with a system, the reliable evidence of understanding is not the artefact but what the student can do with it: explain a decision, defend a choice, answer a question the artefact does not contain. Oral examination is not a novel technique, but it acquires a new function.
Students are taught the criterion, not the rule. A rule — you may not use it for the essay — teaches compliance and expires with the next tool. The criterion is teachable and does not expire: you may use a system for work whose result someone will check before it affects anyone; you may not use it to produce something that will be taken as a statement about what you can do. Students of fourteen understand this distinction when it is put to them plainly. In my experience they understand it better than adults do, because they are the ones about whom the claims are made.
6. Objections
That this is a counsel of privilege. A teacher with twelve contact hours can afford process-oriented assessment; a teacher with twenty-six in an under-resourced school cannot, and telling her that consequential judgements must remain human is telling her to work more unpaid hours. The objection has force. The answer is not to relax the criterion but to be honest about where the load falls: the criterion is precisely what justifies delegating the preparation work aggressively, so that the judgement work can be done properly. A policy that permits neither — that leaves the teacher both preparing everything by hand and grading under pressure — is the worst of the available options, and it is the current default in many systems.
That the line between formative and consequential is blurry. It is. A formative comment repeated often enough becomes a reputation; a suggestion about who is struggling shapes what a teacher subsequently notices. The blurriness is real, and the response is not to pretend otherwise but to apply the criterion at the point where the consequence becomes traceable to a named person, and to assign boundary cases to the more protected category rather than the less.
That systems will eventually judge better than humans. Possibly, in some measurable respects, they already do; agreement between human markers is not flattering when it is examined closely. But the criterion proposed here is not a claim about accuracy. It is a claim about accountability: that a statement about a person carries an obligation to be answerable for it, and that this obligation cannot be discharged by a process that cannot answer. A system that judges more accurately than a teacher and cannot explain itself to the student has not solved the problem; it has made it harder to see.
7. Implications beyond the individual teacher
Left to individual teachers, the criterion will be applied unevenly, and the teachers most exposed — newest, most overworked, least confident — will apply it least. Three institutional consequences follow.
School agreements should be written at the level of tasks, not tools. Most school-level rules on artificial intelligence currently name applications, which guarantees that they are out of date within a year and silent about whatever appears next. An agreement built on the criterion is shorter, more durable, and can actually be explained to parents and students.
Teacher education has to work on participants’ own material. A session that demonstrates a tool on a generic example teaches operation. A session in which participants bring their own assessment tasks, sort them into the three categories, and rewrite the ones that do not survive the two sentences teaches judgement. This is a heavier format to run and it produces less immediate enthusiasm. It is also the only format from which anything survives the following Monday.
Systems should be required to expose what the criterion needs. If a teacher is to remain answerable for a judgement, she needs to be able to reconstruct what a system proposed and on what basis. Most educational products currently do not make this possible, and procurement rarely asks. The encyclical’s question about digital infrastructures and algorithms — whether they favour participation, responsibility, the protection of the vulnerable and the common good6 — is, in a school context, a procurement question before it is a philosophical one.
8. Conclusion
The formula that artificial intelligence should support rather than replace the teacher is not false. It is unfinished. It expresses a commitment without saying what the commitment requires, and in that gap the delegation of judgement proceeds quietly, task by task, with everyone’s approval and nobody’s decision.
Magnifica Humanitas asks a question the educational debate had largely stopped asking: not what the technology can do, but what becomes of the person while it does it. In a school, the person in question is a student about whom things are said, recorded and passed on. The proposal of this article is that this is where the line runs. Work whose output a human being will check before it affects anyone can be delegated, and should be, generously. Work whose output is a statement about an identifiable learner cannot be, however capable the system becomes.
To cultivate, to educate, to humanise: the three verbs framing this issue are not, on this reading, three separate tasks. They describe a single refusal — the refusal to let the question of what a person is be settled by whatever happens to be convenient.
Notes
All quotations from the encyclical are the author’s translations from the official Italian text (vatican.va); the original wording is given in the notes below.
1. Magnifica Humanitas, § 4: «radicata nella nostra storia fin dal principio»; § 9: «La tecnologia può curare, connettere, educare, custodire la Casa comune; ma può anche dividere, scartare, generare nuove ingiustizie»; «non è neutrale, perché assume il volto di chi la pensa, la finanzia, la regola, la usa».
2. Ibid., § 10: «la pretesa di un linguaggio unico – anche digitale – capace di tradurre tutto, persino il mistero della persona, in dati e prestazioni».
3. Ibid., § 94: «si “ha di più” ma non si “è di più”, e la persona rischia di essere valutata soprattutto in base alle prestazioni che garantisce» — taking up Paul VI, Populorum progressio (1967).
4. Ibid., § 50: «La sua dignità non dipende dalle capacità che possiede».
5. Ibid., § 51: «la persona finisce per essere ridotta a mezzo per ottenere risultati, a risorsa da usare e sfruttare, e non viene più riconosciuta come fine in sé».
6. Ibid., § 96: «i grandi principi della Dottrina sociale diventano criteri per giudicare e discernere il nuovo scenario: la dignità inalienabile della persona, il bene comune, la destinazione universale dei beni, la sussidiarietà, la solidarietà e la giustizia sociale»; «se il potere delle infrastrutture digitali e degli algoritmi favorisca davvero partecipazione e responsabilità, protegga i più fragili, assicuri un accesso equo alle opportunità e resti ordinato al bene di tutti».
7. Liang et al. (2023). The study evaluated seven detectors on 91 TOEFL essays from a Chinese educational forum and 88 essays by US eighth-grade students drawn from the Hewlett Foundation’s Automated Student Assessment Prize (ASAP) dataset; the mean false-positive rate on the non-native corpus was 61.22 per cent, with 97.8 per cent of those essays flagged by at least one detector.
References
European Commission, Directorate-General for Education, Youth, Sport and Culture (2022). Ethical guidelines on the use of artificial intelligence and data in teaching and learning for educators. Publications Office of the European Union.
Leo XIV (2026). Magnifica Humanitas. Encyclical letter on the safeguarding of the human person in the time of artificial intelligence. Signed 15 May 2026, published 25 May 2026. Available at vatican.va.
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. https://doi.org/10.1016/j.patter.2023.100779
Simone Pfliegel
E N G L I S H
Simone Pfliegel teaches German, English and Ethics at a secondary school in Nuremberg, Germany, and is the founder of Immersive Education Ideations. From 2018 to 2024 she led the PASCH portfolio for North-West Europe at the Goethe-Institut London, with responsibility for the region’s teacher development programmes across twenty-two schools in eight countries. Since 2023 she has trained more than five hundred teachers in generative artificial intelligence, and since 2026 she has served on the European Commission’s expert working group translating European ethical guidance on artificial intelligence into classroom practice. From September 2026 she is a selected member of the European Digital Education Hub squad on STEM Futures.
D E U T S C H
Simone Pfliegel unterrichtet Deutsch, Englisch und Ethik an einem Gymnasium in Nürnberg und ist Gründerin von Immersive Education Ideations. Von 2018 bis 2024 leitete sie am Goethe-Institut London das PASCH-Portfolio Nordwesteuropa und verantwortete die Fortbildungsprogramme der Region für 22 Schulen in acht Ländern. Seit 2023 hat sie über fünfhundert Lehrkräfte zu generativer Künstlicher Intelligenz fortgebildet; seit 2026 gehört sie der Arbeitsgruppe der Europäischen Kommission an, die die europäischen Ethikleitlinien für KI in Unterrichtspraxis übersetzt. Ab September 2026 ist sie ausgewähltes Mitglied der Squad „STEM Futures“ des European Digital Education Hub.
I T A L I A N O
Simone Pfliegel insegna tedesco, inglese ed etica in una scuola secondaria di Norimberga, in Germania, ed è fondatrice di Immersive Education Ideations. Dal 2018 al 2024 ha diretto il portfolio PASCH per l’Europa nord-occidentale presso il Goethe-Institut di Londra, con la responsabilità dei programmi di formazione docenti della regione in ventidue scuole di otto Paesi. Dal 2023 ha formato oltre cinquecento insegnanti sull’intelligenza artificiale generativa e dal 2026 fa parte del gruppo di lavoro della Commissione europea che traduce le linee guida etiche europee sull’intelligenza artificiale in pratica didattica. Da settembre 2026 è membro selezionato della squad „STEM Futures“ dello European Digital Education Hub.