When a company presents a school with a platform that can mark exercises, detect difficulties or produce materials in seconds, the decisive question is not whether the demonstration looks impressive. It is this: what educational problem are we trying to solve, and what kind of relationship among students, teachers and families do we want to protect while solving it? A provider can describe features, prices and accuracy rates. It cannot decide how much effort a learner should make, which mistakes belong to learning, when a human explanation is indispensable or which inequality is unacceptable. That responsibility belongs to the educational community.
Imagine a realistic meeting. The leadership team proposes an artificial-intelligence application to personalise tasks. A mother asks whether her daughter can complete the work without a phone; a teacher wants to know who will review the recommendations; a student worries that every draft will be recorded; the salesperson displays a colourful dashboard and promises to save time. These people are not asking the same question. If the school reduces the decision to buying or not buying, it will miss what matters: defining the purpose, recognising hidden costs and agreeing on safeguards that keep education in command of the tool.
The educational problem first, the product second#
Technology deserves a place in the classroom when it improves an identifiable practice, not merely because it is available. Write the problem in a sentence that can be tested: “first-year students wait two weeks for comments on their essays” is measurable; “we want to be an innovative school” is not. Then ask which non-technological response has been tried, who will benefit and who could be left behind. The answer might be coordination time, a shared rubric or smaller groups. A licence cannot by itself repair an organisation that does not reserve time for teaching, feedback and thoughtful support.
Technology deserves a place in the classroom when it improves an identifiable practice, not merely because it is available.
For a family, this becomes a set of practical questions. Does the tool replace an activity or add another screen? Will the student receive a clue that helps them think, or an answer that removes the need to think? Will a grade depend on an automated prediction? Is there an equivalent option when a home lacks a device, a reliable connection or an available adult? Requesting answers is not resistance to progress. It is a demand that progress have an educational purpose and that its cost not fall silently on people with fewer resources, less support or weaker digital skills.
A school can use a one-page record before approving any pilot. It should state: the problem being addressed; the class and subject; duration; the current practice used for comparison; data to be entered; the accountable person; an alternative without the tool; success indicators; foreseeable risks; and the closing or review date. If these fields cannot fit on one clear page, the proposal is probably not mature enough. The record also prevents a temporary trial from becoming, through inertia, a permanent obligation that nobody remembers choosing.
What the school must decide and the company must demonstrate#
The school determines aims, limits and evaluation. The company must demonstrate how its service behaves within those limits. Before contracting, asking whether the product is legally compliant is not enough. The school should request where data are hosted, how long they are retained, who can access them, whether they are used to train models, how they are erased, which subcontractors are involved and what happens when the contract ends. The Spanish Data Protection Agency’s Guide for Educational Centres reminds schools that they must understand applications, privacy policies and conditions before using them.
A pedagogical demonstration, not just a technical one, is also needed. The provider should work with examples prepared by teachers: an excellent but unexpected answer, writing with spelling mistakes that contains a valuable idea, an explanation in another language, work by a learner who needs additional support and an ambiguous case. The aim is to observe when the tool is wrong, oversimplifies or responds with unwarranted confidence. An average accuracy rate does not reveal who is harmed by errors. The school needs to state which decisions will never be automated and who reviews decisions with academic or personal consequences.
One boundary is especially important: a machine may help organise evidence, suggest activities or return preliminary feedback; it must not become the only voice deciding a grade, an accommodation, a sanction or a suspicion of cheating. Those decisions require context, dialogue and an opportunity to challenge the conclusion. If a detector labels an assignment as AI-generated, that signal is not proof. The teacher should examine the process, speak with the student and let them explain sources, drafts and choices. The burden of proving innocence must not automatically move to a child because of an opaque score.
The European DigCompEdu framework is helpful because it does not reduce teachers’ digital competence to operating buttons. It covers selecting resources, teaching, assessment, inclusion and developing students’ digital competence. That view changes professional learning: instead of a one-hour sales presentation, teachers need time to design an activity, anticipate barriers, test it, observe results and share adjustments. The question is not “can we use the application?” but “can we teach better, identify harm and set the application aside whenever it adds no value?”
A small, reversible and observed pilot#
A sound pilot starts small: one purpose, one or two groups, a short period and a named person able to stop it. The starting position is recorded beforehand. If the aim is faster or better feedback, measure current waiting times, the quality perceived by students and teachers, and the number of revisions students make. During the trial, record not only minutes saved but errors, queries, dropouts, access incidents and additional work. Then compare. Saving a teacher ten minutes is not an improvement if every family must spend half an hour dealing with accounts, faults and passwords.
Reversibility must be genuine. Every student should know an alternative route that does not stigmatise them or remove opportunities. Teachers need to continue a lesson when the internet fails. Essential materials must be recoverable in a usable format, and the school must know how to export or erase information at the end. It is also useful to rehearse failure before deployment: simulate an outage, an incorrect recommendation and a family complaint. A protocol that works only when everything goes according to plan is not a protocol; it is a written hope.
Success is evaluated through several voices. A school can use a short student interview, samples of work from before and after, a teacher log and an accessible family consultation. Novelty must be separated from learning: a tool may initially increase participation because it is different without improving long-term understanding. Examine distributions, not only averages. If most learners advance slightly while students with dyslexia, recent arrivals or poor connectivity lose access, the result requires correction. Inclusion is a design criterion, not a repair added after deployment.
The European Commission’s Digital Education Action Plan sets out a vision of high-quality, inclusive and accessible digital education while recognising challenges involving infrastructure, capacity and training. This helps avoid two opposite errors: assuming that every device improves teaching, or concluding that prohibition is the only protection. A mature school chooses according to purpose, age, context and evidence; combines analogue and digital resources; and preserves spaces for sustained reading, conversation, play, independent writing and cooperation, where efficiency is not the principal value.
Generative AI: teach the process instead of chasing shortcuts#
With generative AI, banning without teaching usually pushes use out of sight. Allowing it without rules, however, makes help, substitution and deception impossible to distinguish. A school needs a policy that is understandable across ages and tasks. A traffic-light model can help: green for ideas, vocabulary or practice when authorised by the teacher; amber for editing, translation or feedback that must be disclosed; red for submitting generated text as one’s own, entering personal data or using the tool in a closed assessment. Subjects can add examples, while shared language prevents contradictory messages.
Assessment must change as well. If an assignment can be completed by copying one instruction into a platform, it may not reveal what educators intend to teach. Staged submissions, justified decisions, annotated sources, a short conversation about the work or applying knowledge to a classroom case are more informative. The point is not to turn every task into an interrogation but to make the process visible. Keeping outlines and versions helps students recognise progress and allows teachers to assess reasoning, revision and authorship without treating unreliable detectors as judges.
UNESCO’s Guidance for generative AI in education and research calls for a human-centred and age-appropriate approach. In practice, that means not opening an individual account simply because a tool is appealing, checking age conditions, minimising data and providing supervision. It also means teaching limitations: a fluent answer may contain errors, bias or invented references. Verification is not a warning added at the end. It requires checking claims, finding the primary source and explaining which evidence caused a conclusion to be retained or changed.
The conversation families are entitled to request#
Families do not need to master every system to participate. They can ask for five answers: what specific objective is being pursued; which data are used; which activity is replaced or changed; who reviews results; and how someone can complain or choose an alternative. The school can explain its decision before requesting consent or installing applications, using plain language and examples. A brief meeting that shows a real task and its limits works better than a catalogue of brands and technical terms. When uncertainty exists, “we are testing until this date and will evaluate these indicators” builds more trust than a promise of absolute safety.
Students should participate in ways appropriate to their age. They can identify when a recommendation feels unfair, which explanations they understand, what barriers they encounter and when a tool helps or takes away initiative. Listening does not mean making every preference a rule; it means adding experience that no contract contains. A student council, a feedback box with guaranteed responses and pilot-review sessions reveal problems early. They also teach digital citizenship: technology is not an inevitable destination but a collection of decisions that people can question, debate, limit and improve.
If harm occurs—data exposure, harassment enabled by a platform, an assessment error or exclusion—the priority is protection and repair, not defending the investment. There must be an identifiable responsible person, a simple channel and a response deadline. Necessary evidence is preserved without wider circulation, an immediate academic alternative is offered and the school explains what will change. It then asks whether the problem was individual or designed into the process. Blaming the person who pressed the wrong button prevents learning when the interface, training or an unclear rule made that error predictable.
A responsible decision in ten steps#
Before approving a technology, a community can follow this sequence: define the problem; listen to those who will live with the change; compare technological and non-technological options; review privacy, accessibility and age; establish prohibited uses; prepare an equivalent alternative; train through real cases; test on a small scale; measure learning, workload and inequality; and decide whether to continue, change or withdraw. Every step needs an owner and a date. “Not yet” is a valid decision when safeguards are missing. So is withdrawing a product that no longer serves its purpose, even when implementation required substantial effort.
The question a company cannot answer alone is which education is worth building. It can contribute technical expertise and must be accountable for its product, but it does not by itself represent a child’s interests, teachers’ experience, families’ circumstances or the school’s shared project. Sound judgement is not measured by how many applications a school adopts. It is visible in the school’s ability to explain why each is used, what it protects, which results are observed and what will happen if it is wrong. Where those answers are public, reviewable and human, technology occupies its proper place: a tool, never the purpose.
