A school announces that it will use an artificial intelligence system to guide assignments, detect difficulties, and recommend learning pathways. Procurement is complete, accounts have been created, and the timetable is fixed. Only then does it invite six students, show them a presentation, and ask which colour they prefer for the home screen. The meeting produces photographs and a statement about youth voice, but no decision remains open. Participation afterwards merely decorates an adult decision; participation beforehand could have questioned whether this was the right problem, which data should never be collected, which human alternative must remain, and what conditions would require the tool to be abandoned.

Adolescents are not a user category to be consulted only to improve acceptance. Those who study, create, talk, and seek help through automated systems know frictions that do not appear in a contract or technical test. They know when a recommendation becomes a label, when a notice makes no sense, when sharing a device changes privacy, or when declining a tool carries a social cost. This knowledge does not replace adult, legal, or professional responsibility. It makes responsibility better informed and helps anticipate effects before they become difficult to repair.

Participation means influence, not attendance#

Meaningful participation requires a visible relationship between what adolescents express and what an institution can decide. Inviting them, listening politely, or counting responses is not enough. Before work begins, participants should know its purpose, the genuine boundaries, who is responsible, and how the outcome will be reported. If a condition cannot be negotiated because of law, budget, or safety, that should be stated. If everything is closed, the activity should not be called co-design.

Meaningful participation requires a visible relationship between what adolescents express and what an institution can decide.

Influence can take different forms. A group may contribute experience to define the problem, compare options using agreed criteria, create prototypes, test accessibility, hold voting places on a committee, or review evidence during a pilot. Not every process requires the highest level, but every process requires honesty about the selected level. Consultation can be legitimate when the institution retains the final decision, provided contributions arrive in time, receive a reasoned response, and can change something significant.

Who frames the question also matters. If a school asks which assistant students prefer, it assumes an assistant is required. An earlier phase would ask where learning becomes difficult, which support is missing, and which responses would be acceptable. The answer may be better tutoring, a simpler existing platform, or different schedules rather than an AI purchase. Opening the diagnosis protects against technological solutionism and recognises that a tool has value only in relation to a demonstrated need.

Participation must start before budgets and procurement#

The decisive moment is when specifications, suppliers, configurations, or even the decision to buy can still change. A responsible team can organise the work around five gates. The first defines the need and affected population. The second turns needs into requirements: data minimisation, accessibility, language, opportunities to challenge outcomes, a human alternative, and limits on use. The third compares offers and evidence. The fourth conducts a small, reversible pilot. The fifth decides whether to continue, modify, or stop.

At every gate, specific decisions should remain open to adolescent participation. Students may reject a public ranking, demand that a recommendation explain its factors, or point out that parental consent does not resolve pressure inside a classroom. The institution retains its duty to protect rights and justify spending; it does not transfer responsibility for approving risks to minors. Its task is to integrate situated knowledge and document how that knowledge influenced the outcome.

Time is also a safeguard. A twenty-minute consultation during lunch favours immediate answers and confident speakers. Several sessions, understandable information, and written channels allow people to think, ask others, and disagree. Participation may be paid or formally recognised when it involves sustained work. It should never determine grades, access to services, or relationships with teachers. Taking part must be voluntary, and declining must not become grounds for suspicion.

Representation, accessibility, and disagreement#

Six high-achieving, readily available students cannot represent an entire community. Design should include different ages, histories, abilities, languages, and material circumstances. It should deliberately reach those commonly excluded: adolescents with disabilities, recent arrivals, rural residents, people who share devices, students facing academic difficulty, and young people subject to intensive monitoring. They are not invited to display diversity. They are involved because consequences may fall unevenly.

Accessibility involves more than adapting a meeting room. Advance information should use examples to explain what a system does, what it cannot know, which data it uses, and which effects it may produce. Easy-read material, interpretation, visual support, additional time, or individual participation may be required. An independent facilitator reduces pressure when the person listening also grades, disciplines, or provides the service. Families can offer support, but the adolescent voice should not automatically be replaced by an adult voice.

Disagreement is valuable information. A process that requires consensus tends to erase minority experiences. The final report should preserve reservations, conditions, and open questions. If two groups assess a monitoring feature differently, the institution might restrict it, provide choice, or investigate further. It does not need to present a single youth voice. It needs to show which positions it encountered, how it weighed risks, and who assumes responsibility for the final decision.

Three cases in which arriving early changes the outcome#

Case 1: a dropout risk detector#

A secondary school considers a tool that assigns every student a risk score using attendance, grades, and digital activity. Before procurement, it brings together diverse students, counsellors, teachers, and the data protection lead. Young people explain that some absences reflect family care, transport failures, or medical appointments, and that a visible label may change teachers’ behaviour. They also note that low activity does not signal disengagement when several people share one computer.

The discussion changes the decision. The school rejects an automatic individual ranking and creates limited signals that never produce sanctions or decisions without an interview. It reduces data, restricts access, establishes deletion periods, and allows contextual corrections. Part of the budget moves to tutoring. During the pilot, the team measures support offered, false alerts, and differences among groups rather than aggregated accuracy alone. Participation did not legitimise the detector; it transformed the problem and moved resources towards a less intrusive response.

Case 2: a wellbeing chatbot#

A municipal service wants to offer automated conversation outside youth-centre hours. In advance workshops, adolescents distinguish asking for information from seeking help during a crisis. They ask that the system never pretend to be a friend, promise secrecy, or obscure how to reach a person. Participants who have used mental health services warn that a generic pop-up can feel like rejection and that suddenly disclosing an entire conversation to family members may create danger.

The service confines the chatbot to navigation and information, removes anthropomorphic language, and designs a warm handover to professionals. It establishes a specific immediate-risk protocol, clear privacy notices, and an option to delete content. Young people test scenarios, but are not asked to perform their own crises or review disturbing material without support. A safeguarding committee oversees incidents. The crucial work happens before launch: youth experience defines boundaries that marketing and performance measures had missed.

Case 3: generative AI for assignments#

A school intends to activate a writing assistant for every student. Before deployment, several groups compare the tool on real tasks. Some students value examples and reformulation; others fear that it will standardise their voices, penalise language varieties, or expose learning needs. They ask to know when a suggestion comes from the machine, retain their own drafts, and submit work without using it. Teachers also discover that their standards for authorship are inconsistent.

The pilot is limited to two subjects, with an equivalent alternative, minimal local logging, and no automated grading. Students learn to verify claims and disclose assistance. A joint review finds benefits for beginning drafts, but also dependency and fabricated references. The school retains certain functions, disables others, and changes assessment to value process. The final decision does not match every individual preference, but each change can be traced to evidence and shared deliberation.

A safe process does not extract personal stories#

Listening to adolescents about AI may approach experiences involving harassment, sexuality, health, discrimination, or family conflict. The institution should explain in advance which information it will collect, who will see it, how long it will be retained, and where confidentiality ends. Names and intimate details are unnecessary for understanding many patterns. Hypothetical examples, anonymous responses, and later withdrawal where feasible should be permitted. Recording a session for convenience rarely justifies accumulating identifiable voices and images.

Safeguarding rules cannot appear after a disclosure. The team needs a procedure for responding when risks of harm, abuse, or exploitation emerge, and it must describe that procedure in age-appropriate language. A facilitator does not promise absolute secrecy. Nor do they interrogate, diagnose, or turn a design workshop into therapy. They listen to what is necessary, activate the competent route, and prevent a participant from having to repeat the story to multiple adults.

Another risk is using youth emotional labour to improve a commercial product without compensation. Supplier agreements should prevent secondary use of contributions, recordings, or prototypes for model training or advertising. The institution defines ownership, deletion, and publication. When participants create materials or provide substantial analysis, their authorship is recognised if they wish and their contribution is compensated fairly. Symbolic gratitude is not a substitute for fair conditions.

Returning results completes the accountability loop#

Feedback should answer three questions: what was heard, what changed, and what did not change with reasons. It may take the form of a public table without personal data, a return session, or an accessible report. Vague claims that comments will be considered do not allow anyone to evaluate influence. It is better to state that a ranking was removed, a non-AI alternative was added, or a request was refused because it conflicted with a safeguarding duty.

Accountability continues during use. Adolescents need a route to report errors, appeal decisions, and propose changes without retaliation. The responsible committee reviews disaggregated indicators, incidents, complaints, actual use, and withdrawals. It also agrees pause thresholds: a data leak, serious harm, persistent inequality, or a supplier’s inability to explain an outcome. The possibility of stopping is not a failure. It is a defining feature of an honest pilot.

A mature institution avoids selecting only favourable comments. It publishes tensions and limitations, protects identities, and explains who made the decision. If a supplier invokes commercial secrecy to prevent sufficient explanation, that constraint belongs in the assessment. Early participation matters precisely because it can inconvenience, delay, or cancel. If its only purpose is confirmation, it has ceased to be participation and become promotion.

Institutional frameworks for putting the principle into practice#

The United Nations Convention on the Rights of the Child recognises children’s right to express views in matters affecting them and to have those views given due weight according to age and maturity. UNICEF’s Engaged and Heard guidelines describe conditions for voluntary, inclusive, safe, and influential adolescent participation. Together, these references frame listening as a right and a practice of governance rather than a communications device.

For technology, UNICEF’s policy guidance on AI and children treats wellbeing, inclusion, privacy, safety, transparency, and accountability as design requirements. UNESCO’s guidance for generative AI in education and research calls for a human-centred approach, data protection, and age-appropriate use. These sources do not automatically endorse a product; they help institutions frame questions, safeguards, and duties.

The Council of Europe’s resources on listening to children assemble standards for ethical and effective participation. The European Union Artificial Intelligence Act establishes risk-based obligations for specified uses and actors. Precise application depends on the system, function, date, and jurisdiction. A participatory process therefore complements, but never replaces, legal, security, educational, and data protection assessments.

Questions an institution should be able to answer#

Before inviting young people, a team should be able to explain which decision remains open, what resources exist, which risks it anticipates, and how it will prevent harm. During the process, it should record who is absent, which support participants need, and whether adult power is silencing criticism. Before deciding, it must test contributions against rights, evidence, and obligations. Afterwards, it should return results, retain only necessary data, and schedule a review.

A simple test reveals quality: if adolescent contributions disappeared from the record, would the decision be identical? If the answer is always yes, there was activity but no influence. The trace may be a new requirement, a rejected supplier, a disabled function, a redistributed budget, or a reframed question. An institution need not accept every proposal, but it must show that voices arrived while they could still alter direction.

Participation before decisions distributes knowledge without diluting responsibility. It makes technology visible through the daily lives of those who will experience its effects, keeps alternatives open, and reveals hidden costs. Participation afterwards can support evaluation, complaint, or correction, but it should not be disguised as co-design of a settled choice. A generation that talks with machines deserves more than a photograph beside a finished product: it deserves to intervene while society can still decide whether that product should exist, under which conditions, and for whose benefit.