The future does not become human because a machine speaks naturally, anticipates a need, or accompanies every hour of the day. It becomes human when an adolescent can say stop, ask for a person, erase a conversation, change channels, or continue without the system while receiving equal treatment. A genuine exit is neither a hidden link nor an exceptional concession: it is evidence that the relationship preserves freedom. Where leaving artificial intelligence means losing a grade, an appointment, a group, or an opportunity, apparent choice becomes designed obedience.

Adolescence intensifies this question because it combines exploration, institutional dependence, and a powerful need to belong. An assistant can be useful for studying, creating, or expressing a difficult concern. It can also quickly occupy the place of tutoring, friendship, or a doorway to services. The problem is not limited to what a model says. It matters who decided that it should be present, what happens when it fails, and whether another form of help is genuinely reachable. The quality of technology is also measured by the dignity with which people can leave it.

Leaving is more than closing a window#

An effective exit has several dimensions. A person must be able to interrupt a particular interaction, disable a function, withdraw data where applicable, and choose an equivalent non-automated service. They must understand consequences before deciding and be able to return without retaliation. Closing a screen is insufficient if a school continues grading through the profile produced, if a service retains sensitive inferences, or if the only alternative requires a young person to explain a private situation in public.

An effective exit has several dimensions. A person must be able to interrupt a particular interaction, disable a function, withdraw data where applicable, and choose an equivalent non-automated service.

There is also a cognitive exit. A conversational system may present itself as helper, adviser, or companion until suggestion, instruction, and judgement become difficult to distinguish. Its interface should recall its limits, signal uncertainty, and avoid language engineered to create emotional dependence. A pause should remain available, especially after long or sensitive exchanges. Keeping a door open to silence, paper, human encounters, and independent deliberation prevents automated convenience from becoming a permanent occupation of attention.

A social exit requires refusal to feel ordinary. If an entire class uses an automated tutor, someone who asks for a teacher’s explanation may appear slow or difficult. If a wellbeing app awards badges for daily conversations, interruption becomes visible. Institutions should normalise alternatives, announce them alongside the digital option, and avoid rewards that indirectly punish refusal. Freedom does not depend only on a button. It depends on a culture that does not interpret a boundary as failure to cooperate.

A form accepted by a family does not make voluntary something required for study, support, or membership of a group. This is even clearer when terms are long, change without explanation, or combine essential functions with secondary uses. Authorisation may form one part of protection, but it does not replace data minimisation, purpose limits, adult oversight, or an equivalent route. The institution that selects a system retains responsibility even when an adolescent taps accept.

Decisions commonly bundled together should be separated. Someone may accept help summarising a text while refusing to let drafts train a model. They may want academic recommendations without an emotional profile. They may use an automated channel to locate a service and prefer a person when the conversation becomes intimate. Granular controls reflect this diversity. An all-or-nothing alternative forces a disproportionate surrender: give up data and agency, or lose the principal benefit.

Understanding must be tested in real situations. A notice may be legally accurate yet useless to someone who is tired, distressed, or working in another language. Before deployment, diverse adolescents can try to find how to stop the tool, correct an inference, download content, and contact a person. If they cannot succeed without the design team’s help, the exit does not yet exist. Testing should include old phones, screen readers, unstable connections, and pressured moments, not merely an ideal demonstration.

Designing a human door from the beginning#

A human alternative cannot be improvised when a crisis arrives. It requires named people, hours, training, maximum response times, and protocols that distinguish an ordinary question from immediate risk. The system should explain precisely when someone becomes involved, which part of a conversation that person can see, and what will happen next. A message that advises seeking help without enabling the next step abandons responsibility at the moment when responsibility matters most.

The transition should preserve context without becoming intrusive. It can let an adolescent choose which passage to share, draft a summary for review, or begin again. Automatic transfer of everything may make a service faster, but it can reveal unnecessary details. In dangerous situations, a safeguarding obligation may restrict choice; even then, the institution should communicate in understandable language, limit disclosure, and avoid false promises of absolute secrecy.

Designing human capacity costs money. If a project funds licences, integration, and promotion but reserves nothing for tutoring, support, review of decisions, and attention outside the automated flow, its exit will be decorative. A budget exposes ethical priorities. Cover for absences, accessible channels, and routes for escalating complaints are also necessary. Technology should not create a silent queue of complex cases that no person has time to address.

Three cases that test whether an exit is real#

Case 1: the tutor that conditions an assignment#

A secondary school adds a generative assistant to its homework platform. To submit an essay, students must go through a conversation that proposes an outline and flags alleged weaknesses. One student wants to write without assistance because she is developing a bilingual voice that the system corrects towards a uniform register. A skip button exists, but using it reduces her process score and sends an alert to the teacher. Formally she can leave; academically she pays for doing so.

The school revises the design and creates two equivalent routes. Both retain drafts, reflection on decisions, and access to tutoring; neither gains an advantage from using AI. Automated suggestions are visually distinct, are never incorporated without explicit action, and can be deleted. Teachers assess the text and demonstrated process, not compliance with the tool. The relevant measure stops being daily use and begins to include learning, resolved requests for human help, and differences between students choosing each route.

Case 2: a wellbeing conversation that needs to stop#

A municipality offers a night-time chatbot to guide young people towards wellbeing resources. A boy begins by asking about sleeplessness and eventually describes fear inside his home. The system repeats questions, declares that it is there for him, and displays a generic phone number at the end. There is no clear way to request a person, control what is shared, or know whether anyone will receive the message. The conversation sounds close, but its exit leads into an operational void.

The service then confines the system to information and initial navigation. At every point it allows human contact, explains availability, and offers calls, text, or an in-person appointment. For urgent risks, it displays specific actions and activates a supervised protocol, with advance notices about confidentiality. The young person decides which summary to send unless a clear safeguarding duty applies. Professionals and an independent lead review incidents; conversations are not reused as training material or promotional stories.

Case 3: the filter that decides who belongs#

A network of youth activities uses a recommender to assign places according to interests, location, and past behaviour. One adolescent continually receives distant suggestions and never sees the robotics workshop she requests. She can close the recommendations, but the manual form is hidden and places are filled before an employee responds. The exit exists in the interface, not in the timetable. Automated delay becomes a barrier to access.

The organisation reserves places for manual applications, publishes understandable criteria, and establishes human review before closing. Every recommendation allows a person to say that it does not fit without turning the objection into a negative signal. The adolescent can correct data and request an explanation. An audit compares distribution by neighbourhood, gender, disability, and application route. If unjustified differences persist, the system pauses. The alternative has the same deadline, information, and opportunities, so it ceases to be an inferior option.

Governing the right to stop, correct, and return#

An exit policy should assign responsibilities. Someone maintains controls, another person handles transfers, and a body with authority decides pauses or withdrawal. Contracts should guarantee portability, deletion, access to relevant records, and continuity when suppliers change. If an institution cannot recover content or explain what happens when an account is cancelled, it has surrendered an essential part of its capacity to care.

Review requires measures other than growth and retention. Institutions should count how many people find the exit, how long they wait for an alternative, how many return without difficulty, which complaints appear, and whether leaving produces worse outcomes. Data should be disaggregated carefully to identify barriers without exposing identities. A high number of exits does not prove failure by itself: it may reveal a population able to choose, or a tool requiring change. Interpretation requires listening to reasons.

Adolescents should participate in this governance without being burdened with it. They can identify pressure, test messages, review pathways, and propose pause thresholds. Leaders, professionals, and suppliers remain responsible for safety, legality, and resources. A minor should never have to decide whether serious harm is acceptable in exchange for innovation. Their experience informs a decision; it must not transfer risk to them or legitimise a conclusion reached in advance.

Four institutional references that turn exit into an obligation#

The United Nations Convention on the Rights of the Child places best interests, participation, privacy, and access to appropriate information within a rights framework. UNICEF’s policy guidance on AI for children translates this framework into requirements concerning inclusion, data protection, safety, transparency, and accountability. Together, they prevent an exit from being treated as a mere interface preference.

UNESCO’s guidance for generative AI in education and research calls for a human-centred, age-appropriate approach subject to data protection and pedagogical validation. The European Union General Data Protection Regulation establishes principles including lawfulness, transparency, purpose limitation, minimisation, and rights relating to specified forms of processing. Precise application depends on function and jurisdiction, but no educational purchase should hide these questions behind novelty.

These references should become observable decisions: a non-automated alternative, accessible information, competent human intervention, complaint channels, retention limits, and the capacity to suspend. Citing a standard in a policy does not compensate for a blocked door in practice. The team should document tests, incidents, changes, and reasons. It should also review whether the exit still works after an update, a budget reduction, or a change of supplier.

An open future accepts refusal#

Before approving a system, an institution can ask simple, demanding questions. What does a non-user lose? How long does it take to reach a person? Can someone withdraw one part without surrendering everything? Who corrects an outcome, and by when? What remains after departure? Can a person return without being marked? Which event would stop the service? If answers depend on favours, technical knowledge, or family persistence, the right exists only for people who already hold power.

An exit is not nostalgia for a world without technology. It makes bounded experimentation possible, allows learning from failure, and preserves alternatives when promises are not fulfilled. A reversible pilot is more responsible than a deployment presented as inevitable. An available person does not compete with innovation; that person contributes judgement, context, and a duty to respond. Paper, face-to-face conversation, and manual procedures may look less impressive, but they provide redundancy when a network fails or a model is wrong.

A generation that talks with machines does not need every conversation to last forever. It needs to know that silence will be respected, a doubt can reach someone responsible, and no opportunity depends on performing trust in a system. The future retains humanity when people are not enclosed within the solution chosen for them. Designing the exit means recognising that changing one’s mind, asking for help, stopping, and returning are not user failures. They are essential forms of freedom.