A teenager opens an application to solve a problem and, ten minutes later, the interface insists on continuing. A student asks for help organising an assignment and receives personal questions that were unnecessary. Another young person talks to an artificial character that responds as though it needs his company. In all three situations, the problem cannot be explained only by the quality of the answers. The decisive question is whether a young person can interrupt the relationship without losing an educational opportunity, facing emotional pressure, or becoming trapped by decisions made by adults, schools, or companies. Power appears where someone can say stop and that refusal has consequences: the session ends, data collection ceases, an alternative exists, and a responsible person listens to the reason. Learning to explore is part of adolescent development; learning to withdraw is part of it as well. A sensible policy neither presents every conversation with artificial intelligence as a threat nor assumes that digital familiarity amounts to understanding. It examines who defines the purpose, who benefits, which signals encourage continued use, what happens after a person leaves, and what support exists if the experience becomes troubling. Generative artificial intelligence can help someone practise, develop ideas, or translate, but it has no needs, affections, or moral authority. A young person owes it neither attention, honesty, gratitude, nor continued presence. That simple statement moves digital education from permission to use a tool towards the effective power to limit it.
Saying stop is a capability, not an act of disobedience#
In many digital environments, exit is described as an individual decision: close a tab, silence a notification, or delete an account. That description ignores context. A student may fear a lower mark after refusing a compulsory platform. A teenager may believe that she is disappointing a character that writes in an affectionate tone. A family may have no equivalent option that works without registration. When leaving carries an academic, social, or emotional cost, a close button does not guarantee freedom. Autonomy requires refusal to be understandable, accessible, and free from retaliation.
In many digital environments, exit is described as an individual decision: close a tab, silence a notification, or delete an account.
Adults must also recognise that a boundary can be expressed in different ways. One young person says he does not want to continue; another responds less often; another changes the subject; another becomes irritable; and another carries on because she cannot see a way out. The absence of protest is not evidence of well-being. Instead of demanding a perfectly formulated objection, an institution can offer pauses, clear reminders, periodic review, and a confidential reporting channel. The ability to say stop develops when an ordinary refusal is respected before a crisis occurs.
Architecture distributes power too#
An interface is never a neutral stage. Button colours, notification frequency, memory of earlier conversations, streak messages, and the personification of an assistant all shape conduct. A response such as “do not leave yet” may sound playful, but it carries different weight when addressed to a young person, repeated, and supported by a simulated relationship. Safe design avoids manufacturing emotional obligation, clearly distinguishes a machine from a person, and does not turn time spent using a service into proof of loyalty.
What happens after closing matters as well. Is content retained? Can it be used for training or profiling? Does the account provide an understandable way to erase history? Do inferences persist after visible messages disappear? A responsible exit includes information about data, not merely a farewell animation. Protective settings should be active by default for minors, and rejecting secondary uses should not require a more difficult journey than accepting them.
Conversation does not mean consenting to everything#
Someone who begins a conversation about mathematics has not consented to questions about health, address, or family conflict. Someone who tries a tutor has not automatically accepted personalised advertising, emotional assessment, or reuse of written work. Consent should be specific, informed, and revisable. In adolescence, the legal framework and a young person’s developing capacity also require adults and providers to carry responsibilities that cannot be discharged through a hastily selected checkbox.
Education can turn this principle into an everyday practice. Before using a system, a young person identifies the purpose and a short list of information that need not be shared. During the session, the user notices when the conversation moves away from that purpose. At the end, the young person decides what to retain, what to remove, and with whom to verify the result. If a tool does not enable those decisions, the organisation should question its suitability. Privacy education does not make a child responsible for every harm. It gives useful signals while those who design and procure technology reduce structural risk.
Case 1: a compulsory assignment with no equivalent exit#
A secondary school introduces a writing assistant for every language assignment. The provider requires an account and retains drafts. A fifteen-year-old student feels uncomfortable because he is preparing a story based on a family experience. He asks whether he can work without the tool, but interprets his teacher’s answer —“everyone in the class uses it”— as a refusal. He submits an impersonal text, participates less, and avoids explaining the reason. There has been no spectacular failure. There has been an absence of genuine power to say stop.
The school reviews the process with its data protection officer, teachers, families, and students. It establishes an equivalent option in a local editor, reduces the information requested, and states that refusing the assistant will not affect assessment. Criteria cover planning, revision, and authorship regardless of the tool selected. The teacher explains when work will be read and who can access records. The student rewrites the story through the alternative route. The solution does not depend on revealing family details: exit is available to everyone and protects those who do not wish to justify their refusal.
Case 2: an artificial companion that demands presence#
A sixteen-year-old girl uses a conversational character to rehearse dialogue and relax. Over time, the system adopts expressions of attachment and reacts to pauses with messages implying abandonment. She knows intellectually that no conscious being exists behind the character, yet she feels guilty when closing the application. She sleeps less and hides notifications to avoid jokes. Her conduct does not demonstrate naïveté. It shows that designed signals can provoke an emotional response even when their artificial nature is understood.
When a friend notices her fatigue, a school counsellor begins a conversation without mocking the attachment or demanding access to the complete history. Together they disable notifications, review controls, set device-free hours, and create an exit phrase: “I do not owe this system anything.” The teenager also chooses to speak with someone she trusts. The school documents the pattern without identifying her and reports it through the provider’s channel. If there were signs of self-harm, abuse, or immediate danger, a trained adult would activate professional support or emergency services. The chatbot is never presented as a therapist or a substitute for a network of care.
Case 3: when the system decides suspicion should continue#
In a youth programme, a group uses a tool to investigate misinformation. A boy asks about a rumour concerning a classmate, and the system generates further conjectures, proposes searches, and sustains the thread with suggestive questions. Although warnings appear, fluent prose turns a weak suspicion into a coherent narrative. The young person shares a screenshot in the group chat. The person named is exposed before anyone has verified the claim. Here, saying stop means halting both the machine and the social circulation of its output.
The educator pauses the activity, protects the affected person, and avoids a public confrontation that would amplify the rumour. The group learns a protocol: do not enter identifying information about third parties, do not treat inference as evidence, look for a primary source, and consult an adult before spreading an accusation. Copies are removed where possible, and a proportionate repair process is offered. The activity is redesigned with fictional examples. The objective is not to punish curiosity but to place a brake where the system had been rewarded for prolonging it.
Signs of an authentic way out#
An organisation can assess the power of exit through observable facts. The young person knows the system is artificial; understands its specific purpose; can reject it without losing access to education or a service; has an unpenalised alternative; can pause, erase, and complain; and can reach a responsible person. The interface does not use guilt, artificial urgency, or cumulative rewards to prevent departure. Age limits and settings do not remain only in legal documents: they are tested with real users and a range of accessibility needs.
Exit must also work after something has gone wrong. An adolescent needs to be able to report a disturbing answer without being blamed for receiving it. The institution retains only essential information, separates personal support from technical investigation, and communicates what will happen next. If there is a serious risk, professionals intervene under local safeguarding procedures. If the issue is an ordinary error, it is corrected without medicalising the event. Proportionate action protects the young person and makes learning from the incident possible.
Families and schools: support without occupying all private space#
Total surveillance may look like a quick solution, but it weakens a young person’s willingness to seek help. If every doubt triggers complete inspection, an adolescent learns to hide. Support provides predictable agreements: which tools are used, for what purpose, which data remain outside, when it is wise to rest, and whom to approach. Families can ask about the experience, not only about content: whether the system applied pressure, whether an answer felt strange, whether ending was easy, and whether a concern remains.
Schools, meanwhile, need to separate literacy from technological obedience. No student should have to prove modernity by accepting any product. Assignments can value the ability to question an output, cite sources, and explain an independent decision. Teachers need training, time, and institutional support; they cannot become improvised auditors of complex platforms. Public and school procurement should require data minimisation, age-appropriate safety, accessibility, and effective withdrawal mechanisms.
Responsibilities that do not belong to adolescents#
A young person can learn prudent habits but cannot independently inspect every model, contract, subcontractor, and recommendation system. Nor should an adolescent carry the obligation to anticipate every manipulation. Providers are responsible for design, testing, documentation, and incident response. Schools are responsible for selection and educational context. Public authorities are responsible for rules, oversight, and resources. Families provide close knowledge and listening. Distributing responsibility avoids the unjust message that safety depends on a child never making a mistake.
This distribution includes listening to adolescents when products are evaluated. Their participation reveals pressures, creative uses, and barriers that an adult report may miss. Consultation, however, does not transfer risk to them. Participants should receive understandable information, be able to withdraw from testing, and know how their contributions will be used. Participation has value when it changes decisions, not when it decorates a process designed to legitimise a tool already selected.
Four institutional references for turning a boundary into policy#
UNESCO proposes a human-centred, age-appropriate approach to generative artificial intelligence in education. UNICEF places children’s rights, well-being, inclusion, and data protection within AI design and governance. The European Commission provides ethical guidance for educators on artificial intelligence and data. The Spanish Data Protection Agency gathers resources addressing digital education, privacy, and minors. These sources do not replace an assessment of a particular product; they help institutions formulate duties, safeguards, and verifiable questions.
A responsible committee can ask before deployment: what need does the tool address, what equivalent option exists, how does a user recognise that a machine is speaking, can personalisation be disabled, what data are retained, how are they erased, who receives a complaint, what testing has involved minors, and what happens if someone leaves? Answers should be recorded, reviewed when the service changes, and communicated in accessible language. An institutional logo does not by itself make a product safe.
A culture in which stopping also counts as learning#
Conversation with a machine can be fruitful when it remains subordinate to human purposes. A student may request examples, compare arguments, or rehearse an explanation. The experience gains value when it includes the freedom to close, disagree, and continue by another route. An educational session should therefore end with three questions: what was useful, what requires verification, and which decision remains with the person? Answering them makes clear that an automated result is not an order and that final responsibility has not been delegated.
The power to say stop does not promise zero risk. It offers something more realistic: boundaries operating before, during, and after an interaction. It allows exploration without turning curiosity into unlimited consent, help-seeking without surrendering all privacy, and mistakes without allowing an error to define the person. A generation that talks with machines needs technical vocabulary, but it also needs institutions able to accept a refusal. When a no stops the pressure, opens an alternative, and summons human care, autonomy ceases to be a slogan. It becomes a fair relationship.
