A number acquires unusual authority when it appears on a screen: it looks exact, comparable, and beyond argument. A school sees an 82 per cent risk of dropout; a family receives a report showing six hours of use; a platform announces a 91 per cent probability that a text was generated by artificial intelligence. Each number may provide a signal, but none contains the story it claims to summarise. During adolescence, when identity, relationships, learning habits, and the need for privacy are changing quickly, turning a signal into a verdict can create more harm than knowledge. Measuring better does not mean collecting limitless data or building a more colourful dashboard. It means deciding which question matters, combining different sources, seeing who disappears inside the average, and keeping a door open for explanation, challenge, and correction. Artificial intelligence can help find patterns, but responsibility for interpreting its results remains human and institutional.
A number answers only the question that produced it#
Every indicator begins with a choice. Someone decides what counts as participation, well-being, originality, progress, or excessive use; selects a source, a period, and a threshold. The decision may be reasonable, but it is never neutral. Counting messages sent does not measure the quality of a friendship. Recording time inside an application does not distinguish making a film, talking to a grandmother, and drifting through recommendations. An average grade does not show whether students understood a concept, memorised an answer, or received help that others lacked. Before interpreting a number, decision-makers should reconstruct the question, the method, and what could not be observed.
Every indicator begins with a choice. Someone decides what counts as participation, well-being, originality, progress, or excessive use; selects a source, a period, and a threshold.
Artificial-intelligence systems add another layer. A score is often the output of a model trained on earlier cases. It may express statistical similarity, not certainty about this person. If historical data contain inequalities, the pattern can learn them and present them as prediction. When the context changes, published accuracy no longer describes current use. Even a high overall rate can conceal errors distributed unequally. An institution therefore needs to ask precisely what the value means, which reference was used for validation, how many false positives it produces, and for which groups it performs worse.
The effect of measurement matters too. When an adolescent knows that every pause, word, or connection feeds a score, they may adapt their behaviour to look acceptable instead of seeking help. When a teacher is evaluated through the percentage of completed assignments, easier activities become tempting. When a company celebrates minutes of interaction, design learns to retain users even when well-being calls for ending the conversation. A sound indicator guides a question; an indicator turned into a target can distort the reality it was meant to improve.
Measure for care, learning, and accountability#
The first discipline is to declare the purpose. The same data should not be collected to adapt an explanation, identify an emergency, evaluate public policy, and conduct research. Each purpose requires different proportionality and a different accountable person. If the aim is to support study, samples of work and a conversation may be enough. If there is a clinical concern, a qualified professional and health procedures are needed. A general-purpose application must not transform a conversation into a diagnosis, and a school should not accumulate intimate information because it might become useful one day.
Measurement for care reduces uncertainty without reducing the person. It combines signals: sustained changes, context, the adolescent's own words, relationships, barriers, and available resources. Measurement for learning compares a process with explicit objectives and returns feedback that makes action possible. Measurement for accountability also examines the organisation: how many errors it corrected, how long it took to respond, which groups experienced more incidents, and which decisions it changed. If every indicator describes the adolescent while none evaluates the platform, responsibility has been distributed unfairly.
A useful framework can bring together four dimensions. The first is outcome: what happened. The second is process: how it happened and what assistance was involved. The third is experience: how affected people understood and lived through it. The fourth is equity: who benefited, who lost, and who could not participate. These dimensions should not be collapsed into another total score that erases differences again. They are read together and accompanied by examples, limitations, and a review date. The purpose is not to manufacture certainty but to make a prudent decision with sufficient evidence.
Three cases in which the average misleads#
Case 1: the authorship detector. A teacher receives a 94 per cent probability that Leo used AI in an essay. The figure looks conclusive, but it does not identify which passage generated the signal or prove who wrote it. Leo is bilingual, used a spelling assistant, and rewrote several sentences after tutoring. The school does not impose an automatic sanction. It compares drafts, sources, and earlier work, speaks with him, and asks him to explain two choices. It finds his own process and permitted assistance. The score is recorded as unreliable evidence, not as part of the student's disciplinary history. The school also revises its policy: no detector will be sole proof, the burden will not fall on a minor, and every suspicion will have a route for challenge.
Case 2: hours of screen time. A family report shows that Nur spent seven hours connected on Saturday. The first reaction is to remove her phone for a week. After asking, the family learns that two hours were a video call with cousins living abroad, two were spent editing a school project, one involved music while drawing, and the remainder went to short videos that Nur herself describes as exhausting. The intervention changes: valuable activities remain, night notifications are disabled, and a specific pause is agreed for content that displaces sleep. A later review considers rest, mood, autonomy, and relationships, not minutes alone. The same total contained experiences with opposite meanings.
Case 3: an academic-risk alert. A school model classifies Salma as high risk because she missed several days and completed fewer exercises. A counsellor might assign a simplified pathway. Before doing so, she consults the student, tutor, and family. The absences followed a temporary move and caring for a younger brother; Salma's assessments show understanding, and she wants expectations to remain high. The school offers flexible deadlines, transport, and support without narrowing the curriculum. It then audits whether the tool confuses poverty or family responsibilities with lack of ability. The alert is useful because it opens a conversation; it would have been unjust if it had closed an opportunity.
Building an assessment from several forms of evidence#
The process begins with a concrete question and a possible decision. “Is this student learning?” is too broad; “can the student justify the choice of sources in this project and improve it after feedback?” can be observed. Independent evidence is then selected: final product, drafts, oral explanation, self-assessment, and teacher observation. Independence matters because five metrics derived from the same log are not five perspectives. The institution also states which information will not be collected, how long records will be retained, and who may see them.
A baseline is established before a model is introduced. What happens under the current method? How many cases does it identify, what mistakes does it make, and how much work does it require? Without comparison, any new dashboard can look like progress. A pilot should be small, reversible, and time-limited. Correct findings, false positives, false negatives, resources, and consequences are recorded. An accurate alert that arrives too late may not help; a correct recommendation that stigmatises may fail as well. Quality includes usefulness, timing, understanding, and the possibility of repair.
Results are disaggregated carefully by age, gender, disability, language, socioeconomic circumstances, or other relevant factors, while avoiding groups so small that individuals can be identified. The question is not only whether the average improves, but whether a gap emerges. Adolescent participation must be real: young people should know what is measured, contribute interpretations, and identify effects adults miss. Participation does not equal consent to every form of processing or transfer legal responsibility. The institution retains its duty of protection and offers a dignified alternative when a tool is not essential.
Every consequential decision requires competent human review. This does not mean pressing “accept” on a screen. The reviewer needs context, time, authority to depart from the system, and an intelligible explanation. They should document why the proposal was confirmed or changed and communicate without turning correlation into identity. “The system observed these signals” is different from “you are a high-risk student.” Language protects the possibility of change and prevents a prediction from following an adolescent as a permanent label.
Four institutional sources and the questions they enable#
UNICEF's policy guidance on AI for children places rights, inclusion, privacy, safety, and well-being at the centre of design. It helps institutions ask whether benefits have been demonstrated for diverse children and whether accessible redress exists. UNESCO's guidance for generative AI in education and research supports human agency, data protection, and age-appropriate use; it encourages educators to examine pedagogical purpose before adopting a tool.
The World Health Organization's guidance on ethics and governance of artificial intelligence for health explains why transparency, accountability, and protection of autonomy matter when a recommendation affects well-being. The European Commission's account of the European AI regulatory framework sets out a risk-based approach with stronger obligations for higher-impact uses. These four sources do not certify products or replace applicable law. They provide criteria that help schools, families, and public authorities document purpose, necessity, oversight, inequality, and routes of appeal.
A responsible institution records the version and consultation date of each document, checks local rules, and publishes its own decisions in accessible language. Citing guidance does not demonstrate compliance. The organisation needs to show how practice changed: which data ceased to be collected, which threshold was revised, which group participated, which contract was amended, or which tool was withdrawn. Institutional evidence matters when it creates accountability, not when it merely decorates a policy.
An agreement to look beyond the number#
Families and schools can agree on a short protocol. When an important score appears, they ask what it measures, its error margin, which context is missing, what other evidence exists, who can challenge it, and when it will be erased. If it affects health, safety, discipline, educational access, or reputation, they do not act on the automatic result alone. They hear the adolescent before reaching a conclusion and offer immediate support when delay might worsen harm. Technical speed does not require human haste.
Platforms must present limitations as clearly as benefits. They need to disclose validation populations, errors across groups, version changes, data use, and complaint channels. They should allow users to turn off personalisation, export information, and erase it. A business success metric such as retention or conversation frequency cannot be mistaken for well-being. Designing for young people requires measuring voluntary exits, healthy pauses, answered requests for help, and harms prevented as well.
Considering several kinds of evidence does not mean treating everything as relative. Some signals demand firm action: a specific threat, data exposure, or persistent deterioration must not disappear inside averages. That is precisely why context and identifiable responsibility are needed. Good measurement distinguishes exploration from danger, a setback from a trend, and a need from a label. The best decision is not the one displaying the most decimal places, but the one able to explain its reasons, acknowledge its limits, and correct course without abandoning the person affected.
A generation that talks with machines will also learn from the way adults use its numbers. If young people see a score command action without questions, they will understand measurement as surveillance. If they see data open conversations, distribute resources, and compel institutions to correct themselves, they will understand that measurement can be a form of democratic care. The goal is not to distrust every number, but to restore its proper size: a partial representation serving human decisions. Better measurement begins when one number no longer has the final word.
