The Student’s Algorithmic Future

Does a risk score predict success — or produce it?

The real legal question in learning analytics is not how accurately the algorithm reads today, but how that reading changes tomorrow’s educational opportunities.

Consider an educational institution that uses an early warning system to calculate the probability that students will fail or leave the programme at the end of the term. The system produces a risk score for each student on the basis of attendance, frequency of LMS logins, incomplete content, examination results and late submissions.

Of two students at the same academic level, one is marked “high risk” and the other “low risk”. The first student is monitored more closely, directed towards basic-level content and not recommended for an advanced project. The second gains access to enriched content, the teacher’s high expectations and more academic options. A few weeks later the first student’s participation does indeed decline; the second’s performance rises. At the end of the term the system reports that its predictions were accurate.

Yet the future being measured here is not a future independent of the algorithm. After seeing the score, the institution related to students differently; the support given, content offered, teacher expectations and opportunities available all changed. In other words the system did not remain merely an observer of the outcome it predicted; it became one of the causes of that outcome.

This possibility is learning analytics’ comparatively under-discussed legal problem. The issue is no longer only whether the student is profiled, on what legal basis data are processed or whether a human is in the decision. The further question is: Can an educational institution present the student’s future — shaped by its own intervention — as evidence of the algorithm’s accuracy?

Where does this article diverge from the earlier discussion?

The point at which success predictions produced by learning management systems turn into profiling, inferential personal data and automated decision-making problems was examined in detail earlier.[1] That article investigated when scattered LMS traces evolve into an institutional profile that produces consequences for the student.

This article does not re-explain that threshold. The analysis begins not where the profile is produced but where the profile begins to be used. Because even a score created for a lawful purpose, technically accurate and placed before a human as decision support can systematically change the student’s options.

This distinction matters. Profiling law mostly looks at the source of data, conditions of processing, transparency and the effect of automated decisions. Education law must also take account of the student’s opportunities for development, equality of opportunity, the quality of pedagogical guidance and the institution’s duty to provide support. The algorithmic future problem arises at the intersection of these two fields.

A prediction may lose its neutrality the moment it enters the institution

In machine learning literature the concept of “performative prediction” describes a prediction changing the outcome it tries to predict through decisions taken on the basis of that prediction. In the framework set out by Juan C. Perdomo and colleagues, the model’s output shapes decisions; decisions in turn shape the world the model will later see as data.[2]

Education is one of the fields where this interaction is strong. Because student success is not only the product of the student’s past behaviour. The feedback given, teacher expectations, counselling capacity, content level, peer group, material resources and institutional allocation of resources also affect the outcome. When the algorithm steers a decision that changes one of these elements, it no longer merely measures the student; it also rearranges the educational environment in which the student will act.

Not every performative effect is negative. A high-risk alert that provides timely and appropriate support can reduce the student’s probability of failure. In that case the prediction becomes a beneficial intervention that falsifies its own outcome. By contrast, keeping a student away from an advanced course, a scholarship candidacy or a qualified project because of a risk label can contribute to the prediction fulfilling itself.

Therefore the question “did the model’s prediction materialise?” is insufficient on its own. The institution must at the same time ask:

• What was done to the student after the prediction?

• Did the opportunities the student could or could not access change?

• How much of the change in outcome stemmed from the initial risk and how much from institutional intervention?

• What might have happened if no intervention had been made or a different support had been given?

When these questions go unanswered, a high accuracy rate may be an indicator not of educational success but of the system re-measuring its own effect.

Four feedback loops

Student profiles often gain force not through a single decision but through small operations that feed one another over time. Institutions must make four loops especially visible.

In the intervention loop, the risk score triggers support or restriction; the intervention changes the student’s behaviour; changed behaviour enters the new score. The core risk is that the model cannot separate the initial risk from the effect of institutional intervention. In the expectation loop, the label affects the teacher’s or counsellor’s expectation of the student; expectation in turn affects tasks given, feedback and the relationship. Here a probabilistic output becomes institutional opinion. In the visibility loop, more notes and records are produced about the marked student; unmarked students are observed less. Data density may be mistaken for risk itself; students who need help but do not appear to the model may be left out. In the institutional memory loop, an old score becomes data for a new term or new model; a previous inference feeds the next. A temporary difficulty becomes institutionalised as if it were a permanent trait of the student.

Among these loops the expectation effect must be handled carefully. There is a broad research field on teacher expectations affecting student success; but exaggerating the effect is also wrong. Jussim and Harber’s review of thirty-five years of research states that self-fulfilling expectations exist but are generally small and often temporary; by contrast stronger effects may appear for stigmatised groups.[3] The cautious conclusion for algorithmic systems is this: a risk score does not necessarily determine the student’s fate; but it can scale low expectations, standardise them and embed previously scattered opinion in institutional workflow.

Algorithmic agenda-setting: Even if a human decides, who sets the order?

Educational institutions often state that artificial intelligence provides only “decision support” and that the final assessment is made by the teacher or administrator. This assurance is important but incomplete. Because even if the system does not make the final decision, it can determine which student comes before the human, in what order and with what problem definition.

If a counsellor with hundreds of students sees on screen only the twenty who received a red alert, the algorithm has distributed the institution’s attention before the outcome of the decision. The counsellor may examine every file carefully; yet the student invisible to the model may never enter the institutional agenda. Similarly a label such as “low participation”, “attrition risk” or “learning difficulty” predetermines which question the meeting will start with.

This is the issue called algorithmic agenda-setting power in this article. The expression is not an independent concept defined in current law. It is a governance concept used to explain that a system can shape institutional visibility, priority and problem definition without producing a binding decision.

Therefore human oversight cannot be measured only by the question “who signed the final decision?”. Oversight must also cover:

• Which students never appear in the system at all,

• Which risks receive higher priority,

• In what language and colour code the alert is presented to the teacher,

• Whether staff can access information beyond the score,

• Whether there is an alternative channel for the student to express their need without an algorithmic marker.

Keeping a human in the process does not constitute adequate safeguard when the algorithm is allowed to set the agenda alone.

Predictive accuracy is not educational legitimacy

Learning analytics tools are mostly marketed with technical indicators such as accuracy, precision, sensitivity or false alarm rate. These measures are necessary; but they do not prove that the system is legitimate and beneficial for the educational institution. At least four distinct levels of success must be separated.

Technical validity asks to what extent the model correctly predicts the targeted outcome in the local student group; adequate evidence is false positive/negative rates, group-based performance and calibration. Intervention effectiveness asks whether the support applied after the alert genuinely improved the student’s situation; appropriate comparison, pilot results and effect by type of intervention are sought. Distributional impact asks to whom the system directed more support, fewer opportunities or more intensive monitoring; evidence is group-based analysis of resource and opportunity allocation. Rights compliance asks whether the process preserved the student’s opportunity for development, participation, appeal and re-assessment; appeal, correction, independent second review and non-closure of options are evidence at this level.

The randomised controlled EWIMS study published by the Institute of Education Sciences of the U.S. Department of Education shows why this distinction matters. In research conducted in seventy-three high schools with 37,671 students, the early warning and intervention system reduced chronic absenteeism from 14 per cent to 10 per cent and failure in at least one course from 26 per cent to 21 per cent. By contrast no measurable effect was found on low grade point average, suspension or the student’s credit progress towards graduation.[4]

The researchers’ warning is even more important: because the previous relationship between graduation risk indicators and timely graduation was based on correlation, it cannot yet be said that improving the indicator will necessarily improve graduation rates. In other words the system may successfully change a particular intermediate indicator; yet it may not produce the ultimate educational outcome the institution seeks.

This example is not a categorical objection to artificial intelligence. On the contrary, it shows that some early warning applications can be beneficial. But it also shows that “whom did we predict correctly?” and “for whom did which intervention work?” must be answered separately. The educational institution must validate not only model accuracy but the prediction–intervention–outcome chain.

Why the issue in Turkish law is not only KVKK

The Ministry of National Education’s 2025–2029 Policy Document on Artificial Intelligence in Education has included learning analytics, development tracking, early warning and decision support applications in the education system’s future plan.[5] Systems that produce forecasts about the student are therefore not a hypothetical future debate. As institutional use increases, legal assessment must expand from processing data to how the educational relationship is shaped on the basis of that data.

In applications within MEB scope, YAZEK ties analysis of student data with artificial intelligence and decision-support uses such as determining the student’s level, grouping or producing feedback about them to ethical declaration. The MEB Ethics Guide also requires that personalised recommendations should not limit the student’s potential; that processes with critical effect on academic future should not be left solely to artificial intelligence output; and that appeal mechanisms be established for the student or parent.[6]

These principles offer an applicable starting point for the performative prediction problem. The duty not to limit the student’s potential requires not only that the model not use discriminatory variables but also that model output not turn into an opportunity-reducing institutional response. Similarly human oversight should mean not only having someone to approve the score but being able to question the agenda the algorithm creates and the consequences of intervention.

Law No. 1739 on Basic National Education envisages directing individuals towards various programmes or schools in line with their interests, aptitudes and abilities; it also counts generality and equality and equality of opportunity and means among the fundamental principles of national education.[7] Algorithmic guidance, when used correctly, can serve these aims by recognising the student’s need early. Yet when a probabilistic score replaces pedagogical reasoning that assesses the student’s current development, “direction” ceases to be support that opens paths before the student and becomes classification that closes paths in advance. The right to education and learning protected in Article 42 of the Constitution must therefore be thought of not only as formal access to school but as fair access to important opportunities within education.[8]

On the KVKK side, the principles of accuracy and currency, purpose limitation, proportionality and storage period are decisive. Article 11/1-g of the Law also grants the right to object to an adverse result arising solely from automated analysis when the conditions of the provision are met.[9] Yet performative prediction adds a layer that data protection law alone cannot answer: a score may be technically accurate on the day it is produced; nevertheless the institutional response to the score may disproportionately narrow the student’s educational opportunities. For this reason accuracy on the record and the fairness of the decision architecture must be examined separately.

Scope distinction must also be preserved here. YAZEK and the MEB Ethics Guide cannot be treated as automatically applying to all university and EdTech uses outside their regulatory domain. Yet read together with children’s rights, personal data protection and comparative artificial intelligence governance, they offer a strong policy standard for high-impact learning analytics. The higher education, private education, labour or consumer law regime to which each institution is subject must also be assessed separately.

The child’s development does not fit into past data

The Convention on the Rights of the Child requires that the child’s best interests be a primary consideration, that the child’s views be heard in matters affecting them and that education be directed towards the fullest possible development of the child’s personality and talents.[10]

The United Nations Committee on the Rights of the Child’s General Comment No. 25, in carrying these principles into the digital environment, makes two important warnings that look directly to the future. The Committee requires that automated systems not be used to influence the child’s behaviour or emotions or to limit their opportunities and development. It also states that negative consequences of automated processing and profiling applications for the child may carry into later periods of life. It emphasises evidence-based policy and standards for education technologies; that use be justified by educational purpose; and that it not violate the child’s rights.[11]

UNICEF’s 2025 Guidance on Children and Artificial Intelligence also recommends monitoring the short- and long-term direct and indirect effects of artificial intelligence on the child throughout the lifecycle, validating benefit with evidence in school-wide applications and taking the child’s developmental stages into account.[12]

A more cautious threshold is also seen in comparative documents. The Council of Europe Committee of Ministers’ 2018 recommendation and the 2021 guide on protecting children’s data in the educational environment recommend that profiling of children be prohibited as a rule; but that an exception may be considered where the child’s best interests or overriding public interest exist and legal safeguards are provided.[13] These documents cannot be presented as if they were direct legal provisions for Turkey. Nevertheless, treating child profiling not as ordinary personalisation but as an intervention requiring exceptional justification matters for policy building.

The EU Artificial Intelligence Act’s treating some uses — access to education, evaluation of learning outcomes and determination of the person’s appropriate educational level — as high-risk is the regulatory counterpart of the same direction.[14] This classification does not apply directly to every application in Turkey; but it shows that a prediction in education may be not only software output but decision infrastructure that distributes fundamental opportunities.

A policy proposal: Open future safeguard

In this framework an open future safeguard may be proposed for educational institutions. This expression is not an independent fundamental right regulated by that name in Turkish law or an official obligation defined by MEB. It is a governance standard developed for applying the principles of the child’s best interests, development and right to education, equality of opportunity, data accuracy, storage limitation, participation and appeal to high-impact student profiles.

The basic idea of the open future safeguard is this: The institution must not give a probabilistic explanation produced from the student’s past behaviour the power permanently to close their future options. The student’s capacity to change, to refute the score, to be re-assessed in a new context and to access opportunities independent of past inference must be made genuinely possible by the system.

This safeguard can be concretised in six rules.

1. It may add support; it must not automatically reduce opportunity

The default function of an early warning score should be to offer the student an additional support option. If the score reduces access to an advanced course, project group, scholarship, programme or similar opportunity, that requires separate and enhanced review. Automatic reasoning in the form “not deemed suitable because high risk” is insufficient. A restrictive decision must be based on current, independent and personalised pedagogical grounds.

2. A firewall must be built between support data and sanction data

Transfer of a risk profile created to help the student silently into discipline, scholarship withdrawal, admission, level reduction or another sanction process must be prevented. A new purpose, legal basis, necessity and proportionality analysis is required for second use; where necessary a new ethical declaration and impact assessment must be run. Technical access rights must also support this distinction.

3. A new term must not automatically inherit the old score

A risk inference is not a permanent personality trait of the student but a time-bound claim established under specific data and conditions. Therefore every score must be accompanied by an “expiry architecture”: validity date, event that voids the score, re-review threshold and list of systems to which the outcome is carried. The default rule at a new term, course or meaningful developmental change should be a clean slate; continuation of the score should be possible only with new and documented grounds.

4. The student must be able to contest not only the data but the interpretation of the future

The appeal mechanism must not remain at the level of “my attendance record is wrong”. The student or parent must be able to say “success cannot be inferred from this data”, “the recommended intervention does not fit my need” or “this profile must not close an option before me”. For high-impact outcomes review must include a second assessment sufficiently independent of the system that produced the first score and the person who routinely uses it.

5. The intervention chain, not the model, must be audited

The institution must track score accuracy, support applied and outcome produced in separate records. Assessment must not look only at how many at-risk students actually failed; it must also examine which intervention was offered to whom, whether the student participated in it, which opportunities opened or closed and whether similar students received different treatment. Otherwise the model may write the change it caused to its own success account.

6. An algorithm-free path for help and assessment must be kept open

A student invisible to the model must also be able to request support; false negatives must not remain outside institutional help. Similarly in high-impact guidance the student must be able to access an assessment path not tied solely to the algorithmic profile. This is not disabling artificial intelligence entirely; it is ensuring that institutional attention and opportunities do not have the algorithm as their only door.

A control map from prediction to intervention for institutions

An educational institution can examine its learning analytics use with the following six-stage map.

1. Purpose asks exactly which educational outcome is being predicted and why. Evidence to keep is clear purpose, target group and alternative method analysis; red line is an unmeasurable purpose such as “general student success”. 2. Prediction asks with what data, period, model and uncertainty the score is produced. Evidence is data source, version, local performance and error records; red line is a label presented as timeless, sourceless and certain fact. 3. Intervention asks which action each risk level triggers. Evidence is intervention matrix, responsible person and distinction between support and restriction; red line is the score alone closing opportunity.

4. Participation asks what the student knows, when they were heard and how they appealed. Evidence is age-appropriate explanation, views and second review record; red line is hidden profile or ineffective appeal. 5. Outcome asks whether the intervention genuinely helped and whether difference arose between groups. Evidence is intervention-based effect, wellbeing, error and opportunity distribution measures; red line is counting only model accuracy as success. 6. Reuse asks when the score will be deleted, updated or carried to another process. Evidence is duration, firewall, deletion and re-assessment record; red line is inference automatically carrying to a new term or decision.

This map does not replace a YAZEK declaration or personal data compliance documents. It makes visible the link they do not answer: What institutional behaviour did the prediction trigger and how did that behaviour change the student’s subsequent data?

The success indicator must be redefined

The success of early warning systems is not giving many alerts or finding at-risk students with high accuracy. Indeed if all high-risk students in an institution fail, that may at first glance look like extraordinary model accuracy; in reality it may indicate that the support system did not work or that the label narrowed opportunities.

More meaningful indicators are:

• proportion of students who accessed appropriate support after alert,

• distribution of students who benefited from support and types of intervention,

• students wrongly marked or needing support without being marked,

• educational options opened and closed because of the score,

• rates of appeal, correction and changed decision,

• whether labels were deleted within the foreseen period,

• stigmatisation, anxiety or loss of autonomy experienced by students in the process,

• whether difference in support and opportunity arose between different student groups.

These indicators do not make technical performance unimportant. They link technical measurement to the educational institution’s real aim: learning, development and fair opportunity.

Conclusion: The most dangerous prediction may be the one that looks correct

Learning analytics can strengthen the right to education by recognising the student’s need early. But the system’s legal and pedagogical value cannot be measured by how accurately it classifies the student. The real value emerges in what support this classification provided the student, which options it affected and how honestly the institution tracked the consequences of its own intervention.

Therefore the new compliance question for educational institutions must not end at “are we profiling?”. It must continue with: Whose attention does the profile direct where? Which student becomes visible, which invisible? Does the alert add support or reduce opportunity? When does the old prediction expire? Can the institution separate its own intervention from the model’s success?

The policy counterpart to these questions is not only an artificial intelligence inventory but a Student Profiling and Algorithmic Guidance Policy. Such a policy must bring the YAZEK process together with assessment and evaluation, counselling, personal data, equality, appeal and records management procedures in the same decision chain. Work can begin by mapping the prediction–intervention–outcome flow for three uses: early warning, course recommendation and level placement.

The educational institution of the future will not be distinguished only by how much artificial intelligence it uses. The real difference will be whether the institution can use algorithmic prediction without turning it into a judgment passed on the student. Because if the institution has intervened in shaping the future that materialises, that future cannot be counted as the algorithm’s neutral certificate of accuracy.