Founder
September 14, 2026
30 min read
Before an educational institution declares that it uses artificial intelligence, it must answer a more demanding question: Why should this system be used, whom will it affect and how, and under what conditions should it not be used at all? This article takes the governance space opened by YAZEK one step further by proposing “Artificial Intelligence Impact Assessment in Education” — briefly, EYZED. EYZED is a decision procedure that addresses pedagogical necessity, children’s rights, personal data, measurement validity, equality, human oversight and vendor risk within a single lifecycle, and that institutions can adapt to their own policy sets.
Conceptual status note: EYZED is a model proposal developed in this article; it is not an official system published by the Ministry of National Education (MEB) or an independent obligation defined in legislation. It does not replace YAZEK, KVKK processes or other statutory and administrative duties. It is designed as an institutional assessment layer that works before them and in connection with them.
Consider a school that wishes to use an artificial intelligence-based system to support students at risk of absenteeism or academic failure at an earlier stage. The system analyses attendance records, examination results, online platform activity and past support information to produce a risk level for each student. According to the vendor’s presentation, the model’s accuracy rate is 91 per cent. The aim is unquestionably legitimate: to reach the student before they are lost.
The model sends a notification to the counselling service for a student marked as “high risk”. There is as yet no grade, disciplinary sanction or decision to remove the student from the register. Even so, the label changes teacher expectations; the student is directed into a lower-achievement group, placed in a remedial programme instead of certain activities, and repeatedly encounters the assumption of being an “at-risk student” in meetings. Even if the system has made an error, that assumption can, after a time, produce its own outcome.
In this scenario, 91 per cent is only the beginning of the questions that must be answered. For which outcome and on which student group was accuracy measured? Is the educational cost of false positives the same as that of false negatives? Is it necessary to use counselling notes for this purpose? Can the student know with which data they were classified? When does the label lose its validity? Is the teacher genuinely free to reject the system’s recommendation? Will the warning be used for support, or will it gradually become an indicator that narrows the student’s opportunities?
Each of these questions belongs to a different discipline; yet the decision is single: Should the system be used in this school, for this student group, for this purpose and with these safeguards? The missing link in artificial intelligence governance in education is precisely this integrated decision.
MEB’s 2025–2029 Policy Document has placed high-impact areas such as learning analytics, development tracking, early warning, pattern recognition, decision support and automated examination assessment on the education system’s agenda.[1] YAZEK, opened for use on 2 February 2026, has tied applications involving student interaction, processing of student data, assessment and evaluation, and decision support to a declaration and monitoring regime.[2]
YAZEK’s function is important: it makes scattered uses visible, records responsibility and creates an institutional channel for ethical violation processes. Yet YAZEK is not a prior-authorisation system. According to current explanations, after making their declaration a teacher may begin implementation without awaiting separate board approval; the school artificial intelligence ethics team is not an approving authority but a guidance and monitoring unit. Where it sees a clear risk, it may warn that the application should not be started or should be stopped.[2]
Here we are speaking not of a deficiency but of the different functions of two instruments. A declaration is the visible record of planned use; impact assessment is the reasoned file of how the decision about that use was made. While YAZEK institutionalises the question “what do you plan to do?”, EYZED aims to answer “why, with what evidence, within which limits and with which criteria for withdrawal?”
The MEB Ethics Guide also provides a strong foundation in this direction. The Guide requires that artificial intelligence use be justified by evidence-based educational benefit rather than by temporary trend or technological popularity; that accuracy and reliability tests be carried out before implementation; that critical decisions remain under human oversight; and that appeal mechanisms be established for students and parents against unfair outcomes.[3] What EYZED does is not to repeat these principles one by one, but to convert them into a decision process that produces outcomes before implementation begins.
This proposal is also aligned with the practical direction MEB has previously set. YEĞİTEK’s Recommendations on Artificial Intelligence Ethics advises that, before a new artificial intelligence system is used, its academic, psychological and social effects on students be assessed, and that outcomes for motivation, self-confidence, social relationships and psychological condition be regularly observed.[13] What is missing is not the principle but the method that will bring these fine-grained reviews produced by different expertises together in a single file and bind them to a decision.
EYZED is an institutional procedure by which an educational institution examines an artificial intelligence application it plans to develop, purchase or use in its concrete context; identifies expected benefits and possible harms; tests evidence; records safeguards, responsible persons and monitoring conditions; and in the end produces a reasoned decision on use.
Every EYZED file must answer three fundamental questions:
Should we use it? Is the defined educational need real, and is artificial intelligence necessary to meet it?
If we use it, how should we use it? Within which limits should purpose, data, user group, decision effect and duration be kept?
When should we stop? Which error, harm, change or lack of evidence requires re-examination of the system or withdrawal from use?
In this respect EYZED is more than a risk inventory. It does not ask only “what could go wrong?”; it also investigates whether the expected educational benefit has materialised, whether the same outcome can be reached by a less intrusive method, and whether the resources allocated to technology entail giving up a more effective educational measure. An ineffective system does not consume only the procurement budget; it also consumes teacher time, student attention and the institution’s capacity for improvement.
EYZED is also not a new name for a data protection impact assessment. The Personal Data Protection Authority’s guide of November 2025 clearly states that Law No. 6698 does not impose a general data protection impact assessment as a mandatory obligation. At the same time, for generative artificial intelligence processing that may produce significant consequences for individuals, it characterises such an assessment as useful and the identification and management of risks throughout the lifecycle as good practice.[4] Data protection assessment is one of EYZED’s mandatory connections, but EYZED goes beyond it to encompass the quality of learning, the child’s development, equality of opportunity, decision procedure and continuity of education.
Similarly, a cybersecurity test shows the system’s resilience to attacks, vendor review the provider’s adequacy, and a pedagogical pilot the learning effect. None of these alone carries the whole decision. EYZED’s real value is to unite fragmented assessments around a common use scenario and a common decision record.
The readiness gap concerning educational institutions’ capacity to validate tools, and the human-centred, age-appropriate ethical validation and pedagogical design approach recommended in UNESCO’s guidance on generative artificial intelligence, also support this need for integration.[8]
The same model may be low-impact when a teacher uses it to prepare a worksheet draft without student data, and high-impact when it scores students’ texts or produces inferences about their psychological state. The question “which artificial intelligence tool are we using?” is therefore insufficient for impact assessment.
EYZED’s unit of review is this combination:
Educational need + purpose + affected group + input data + model output + effect on the decision + conditions of use + duration
Assessment should therefore begin with a brief use scenario passport. The passport should include at minimum the problem to be solved, the owner of the use, the targeted learning or management outcome, those directly and indirectly affected, data to be entered and produced, how the system output will affect which decision, whether use is mandatory, provider and model/version information, pilot duration and the alternative without artificial intelligence.
This record does not merely provide bureaucratic order. It converts vague promises into testable claims. When “we will provide personalised learning” is replaced by “we will recommend teacher-approved additional exercises based on error patterns in specific mathematics learning outcomes for eighth-grade students”, data need, success criterion, cost of error and human oversight become discussable.
The first gate looks not at technology but at need. The institution must set out the pedagogical or administrative problem it is trying to solve, why the existing method has proved inadequate and what concrete contribution artificial intelligence is expected to provide. Then the counterfactual question is asked: Can the same outcome be obtained to a reasonable degree by a simpler rule, instructional design, staff support or a tool that does not process personal data?
We may call this approach the principle of pedagogical necessity. The principle does not turn artificial intelligence into a last resort; it makes the educational justification of the technology choice demonstrable. Especially where student data and high-impact decisions are concerned, there must be a proportionate relationship between the convenience offered by the vendor and the intervention to which the student will be exposed.
The OECD Digital Education Outlook 2026 concretises the importance of this distinction: general-purpose generative artificial intelligence may improve a student’s task output; yet where pedagogical purpose and guidance are absent, that performance gain may not translate into genuine learning gain. Delegating the cognitive task to the tool can also weaken long-term skill acquisition.[7] EYZED’s success question should therefore be not “was better text produced?” but “can the student think, explain or apply better without the tool?”
If pedagogical necessity cannot be shown, assessment may end here. That a system can be lawfully established and technically operated does not mean it should be used in an educational institution; that is not opposition to innovation but discipline of purpose.
The Convention on the Rights of the Child makes the child’s best interests one of the primary considerations in matters affecting children; the child’s right freely to express their views and to have those views given due weight in accordance with age and maturity is a separate right.[6] This framework cannot be exhausted by saying “the parent consented”. Parental information or consent where legally required may be important; but it does not replace a concrete assessment of the child’s best interests or the child’s participation.
UNICEF’s 2025 Guidance on Children and Artificial Intelligence structures a child-centred approach on three axes: protection from harm, benefit to the child and inclusion of the child in the process. The guidance also states that childhood cannot be treated as a single user profile; that rules must be adapted to developmental stages; and that child rights impact assessment and effectiveness review are recommended for large-scale, insufficiently tested educational technologies.[5]
EYZED therefore examines not only physical safety but cognitive development, autonomy, emotional wellbeing, human relationships, space for play and rest, privacy and future opportunities together. For example, a learning assistant that continuously offers suggestions may blunt the student’s capacity to build their own strategy even if it presents correct content. That a chatbot is polite and fluent does not eliminate the risk that the child perceives it as human or develops emotional dependence.
Participation is not showing children a ready-made system and asking “did you like it?”. For high-impact uses, views should be sought in the design phase from students of different ages and with different disability, language and socioeconomic conditions, by safe and age-appropriate methods. The institution should record which view changed the design in what way. The outcome of participation is not the acceptance of every request but the real weight of children’s lived experience in the decision file.
An artificial intelligence system does not process only the data entered into it; it produces new inferences about the student’s probability of success, level of attention, areas of interest, behaviour patterns or risk status. Whether accurate or not, when these inferences are linked to the student they can affect their educational life. EYZED therefore includes in the data map not only inputs but also labels, scores, summaries and predictions produced by the model.
For each data element, purpose, legal basis, necessity, source, access authority, retention period, transfer chain, accuracy and deletion method must be determined. “So the system works better” is not a general purpose. A counselling interview note may technically contribute to predicting a student’s absenteeism risk; but that contribution does not of itself make processing that data for this purpose necessary and lawful. Especially for special categories of data, free text and third-party information, data category and purpose of use must be tested separately.
Here EYZED’s second policy proposal is the expiry date for educational inferences. A model output that a student is “low motivation”, “high risk” or “suited to a particular field” should not become an indefinite institutional fact. The validity period of the inference, conditions for update, who may see it and from which systems the previous label will be deleted must be determined in advance. UNICEF also emphasises that data collected in childhood should be kept for the shortest possible time and should not track the child into adulthood.[5]
This “expiry date” is not an independent right recognised by that name in current law. It is a proposed governance rule to avoid freezing the child in a past pattern, to operate data accuracy and storage limitation and to leave room for educational development.
Technical accuracy and pedagogical validity are not the same thing. An automated scoring system may reach the same result as human assessors to a high degree; yet it may systematically undervalue the responses of creative students, those with dialect differences or those with special needs. An early warning model may successfully predict absenteeism; yet it may be unable to explain why support is needed. A chatbot may give the correct answer; yet it may unnecessarily substitute for the student’s thinking process.
At this gate the institution first defines the success criterion. The criterion should be local indicators aligned with the educational purpose, not the vendor’s general accuracy rate. For learning, transfer without the tool, retention, conceptual understanding and higher-order thinking; for decision support, false positive and false negative rates, timely intervention and decision quality; for content generation, factual accuracy, age appropriateness and curriculum alignment may be assessed together.
Testing should be carried out on a local and adequate sample resembling the target group and should include a meaningful comparison with human implementation or the existing method. The NIST Artificial Intelligence Risk Management Framework also states that artificial intelligence risk varies by context of use, that laboratory measurements may diverge from real-world outcomes, and that systems supporting or substituting human activity need a comparison baseline.[9]
The cost of error must also be examined separately. A wrong suggestion on a worksheet cannot be managed with the same tolerance as a wrong classification affecting a student’s programme access. As impact intensifies, depth of testing, human review and standard of evidence should rise. Where evidence is insufficient, the solution is not to record uncertainty as low risk but to narrow scope, conduct a controlled pilot or defer use.
A model’s overall accuracy may be high while it consistently makes more errors for certain student groups. Differences in language, disability, socioeconomic status, geography and digital access can affect the outcome at every layer of the system, from data representation to interface design. Moreover, discrimination does not arise only when a protected characteristic is explicitly entered into the model. Postcode, device type, time online or absenteeism patterns can under some conditions become indirect indicators.
EYZED should make error rates across groups, levels of access to benefit and distribution of burden visible. Accessibility is not a technical feature added afterwards but part of the use decision. Whether a voice interface is an alternative for a hearing-impaired student, the cognitive accessibility of complex text, requirement of paid devices and availability of offline options should be assessed at the outset.
This testing should not create a new legal problem: collecting sensitive data for equality analysis also requires separate legal basis, necessity, security and access limits. The solution is not to ignore protected characteristics but to design measurement together with data protection. Where a difference is identified, narrowing the purpose of use or not using the system at all in certain decision areas are legitimate outcomes, not only adjusting the model.
The MEB Ethics Guide requires that in critical decisions affecting a student’s academic future or educational rights artificial intelligence should not alone be determinative, that humans should be able to correct the output and that appeal mechanisms should exist for students and parents.[3] Article 11 of the KVKK also grants the right to object to a result to the person’s detriment arising solely from analysis of personal data by automated systems. The Authority’s guide does not treat this right as limited to opposing the outcome but interprets it as a safeguard allowing request for re-evaluation of the grounds of the decision.[4]
Building on these sources, EYZED proposes a human second-look safeguard. This expression is not a claim to a new right regulated by that name in current legislation. For institutional policy it denotes this standard: the student or parent should be able to request that an important outcome affected by artificial intelligence be re-examined by another authorised person, without automatic adherence to the first output and taking relevant additional information and explanations into account.
Initial human approval and second look are not the same thing. In the first stage the authorised person must understand the system’s recommendation, compare it with independent sources, and have the authority and time to reject or change it. At the appeal stage, where possible a different evaluator should reconsider both the outcome and the currency of the data used and the model’s role. Rather than burdening the student with technical model detail, which data mattered, how far the system influenced the outcome, on what grounds the human adopted the decision and how to apply should be explained understandably.
The human oversight required by the EU Artificial Intelligence Act for high-risk systems also expects the overseer to understand the system’s limits, to notice tendency to over-rely on automation, to have the possibility to ignore or reverse the output and to stop the system where necessary.[10] This regulation cannot be treated as directly applicable to every school use in Turkey; but it is a strong comparative benchmark showing why “human in the loop” does not mean only a signature or approval button.
The developer of a purchased system may be another company; but the educational institution’s responsibility in its own context of use does not disappear. If the vendor says “data are secure”, “the model is unbiased” or “optimised for education”, EYZED asks with which document, test and contractual clause those statements can be verified.
Review should cover model and version information, intended and prohibited uses, local performance evidence, data retention and model training options, sub-processors, security incident management, age and content safety controls, logs, change notifications and exit from the service. A contract that does not allow the institution to conduct its own test, obtain necessary logs or learn of a significant change in advance is not only a commercial but a governance deficiency.
What the vendor cannot explain is for the institution not “zero risk” but unproven risk. Especially in areas such as student profiling, assessment and evaluation, biometric characteristics or psychological inference, uncertainty does not reduce the need for high assurance. If the provider does not supply necessary evidence, the institution should shut down the function, narrow data and impact scope, turn to an alternative vendor or not use the system.
An artificial intelligence system may keep the same name while the underlying model, data policy, security settings or output behaviour change. The student group, lesson purpose or weight of the system output in the decision may also expand over time. The first assessment therefore cannot be an indefinite certificate of suitability.
The NIST framework spreads risk management across the lifecycle with governance, mapping context, measuring and managing functions.[9] EU regulation treats risk management for high-risk systems as a continuous and iterative process, while the Council of Europe Framework Convention also envisages iterative assessment of actual and potential human rights impacts and establishment of necessary mitigation measures.[10][11] Canada’s Algorithmic Impact Assessment for public institutions bases assessment on doing it at the start of design and before production, and on updating when function or scope of use changes.[12] These cannot be presented as if they were Turkish education legislation; but the common direction is clear: a changing system is a system to be re-assessed.
EYZED files should therefore specify in advance indicators to be monitored, frequency of review, responsible person and stopping thresholds. Model/version change, new data category, new target group, expansion of purpose of use, unexpected group difference, increase in appeals, serious wrong output, security incident or significant change in the provider’s conditions is a trigger for re-assessment.
Here the third original principle is reversibility. The institution should design not only the ability to shut down the system but, to the extent possible, the ability to undo its effects: Can a wrong label be corrected and access to it limited? Can guidance based on that label be re-examined? Can data and backups be deleted? Can the student return to an alternative learning path? When the provider is changed, can records be transferred? The harder a decision is to reverse, the stronger pre-use evidence and human safeguards must be.
It is tempting to reduce impact assessment to a single arithmetic score; but that can be misleading. Pedagogical benefit does not compensate for lack of legal basis. High technical accuracy does not eliminate the absence of an effective appeal path. Numerous low-risk indicators cannot “average” a red line on the child’s safety into acceptability.
EYZED should therefore use decision gates before scores. Impact can be classified together with the weight and duration of the outcome, vulnerability of those affected, scale of use, nature of data, autonomy of the system, reversibility of error and the institution’s oversight capacity.
Light record is sufficient when there is no student interaction or student data and the output does not determine an educational decision; the minimum process is a use scenario passport with rules on confidential information and output accuracy, and the likely outcome is use with fixed scope. Standard assessment applies when there is student interaction or limited data, use is under teacher oversight, impact is low and easily reversible; proportional review of the eight gates, pre-class testing, information and monitoring are expected, and the outcome may be suitable or conditional use. Enhanced assessment is required when profiling, special category data, assessment, guidance, large scale, mandatory use or output affecting the student’s future is involved; multidisciplinary board, child/parent participation, local validation, discrimination testing, detailed data and vendor review, limited pilot and senior management decision are sought, and the outcome may be conditional pilot, redesign or non-use. Red line opens when no legal ground can be established, high-impact outcome is left solely to the system, purpose cannot be explained, serious harm cannot be managed or the provider does not supply critical data or performance information; where deficiency is remediable, return to design; where not, assessment stops and the outcome is non-use or termination of existing use.
These paths are not YAZEK’s official risk classes; they are an in-house policy proposal. Whether a YAZEK declaration is required is determined separately according to current MEB criteria. For example, student interaction, data processing or decision support in an application on the standard or enhanced assessment path will as a rule also bring a YAZEK declaration into play. By contrast, some preparatory uses that do not require declaration may still be subject to light record because of the institution’s confidential information or copyrighted content.
The gates below can be used not instead of the long assessment file but as a policy annex showing where which evidence should be sought.
1. Pedagogical necessity asks for which educational need and why artificial intelligence is used; expected evidence is problem definition, alternative analysis and measurable benefit claim; warning sign is that the solution was chosen before the problem. 2. Children’s rights and participation asks how the child’s best interests, development and views are safeguarded; evidence is age/development analysis, wellbeing assessment and participation record; warning sign is assuming a single type of “student” and reducing participation to introduction or satisfaction survey only. 3. Data and inferences asks whether input and derived data are lawful, necessary and time-limited; evidence is data flow map, legal basis, retention-deletion and transfer plan; warning sign is vague purpose, indefinite profile and unknown model training or sub-processor chain. 4. Pedagogical validity asks whether the system assesses the right thing, in the right group, with the right measure; evidence is local test, comparison with existing method, error and effectiveness analysis; warning sign is reliance only on the vendor’s general accuracy rate.
5. Equality and accessibility asks how benefit and error are distributed across groups; evidence is subgroup performance, accessibility and alternative access test; warning sign is average outcome concealing group harm and no path for students without digital access. 6. Human decision and appeal asks whether humans can understand, reject and re-examine the output; evidence is authority flow, decision record, explanation and second-look procedure; warning sign is formal human approval and repetition of the same automated outcome on appeal. 7. Vendor and technical evidence asks whether the institution can verify promises and manage change; evidence is test report, version record, audit and change clauses and security/exit plan; warning sign is critical information closed entirely as “trade secret” and updates not notified. 8. Lifecycle and exit asks how success will be monitored, when review or stop will occur; evidence is pilot plan, indicators, thresholds, re-assessment and deletion/return record; warning sign is indefinite suitability and no possibility to close and correct wrong outcomes.
When the opening early warning system is taken through EYZED, the purpose is first narrowed: the system will be used not to “identify the student who will fail” but to flag support needs that may be taken to additional human assessment. The language of use matters too; “at-risk student” is a permanent identity, “support signal to be reviewed” is a temporary procedural suggestion.
In pedagogical necessity review the institution compares a simple rule based only on an absenteeism threshold with the model’s contribution. If the model does not provide a meaningful early-detection advantage, complex profiling is unnecessary. If it does, which data yielded that contribution is examined. If free text from counselling interviews provides a small performance gain but seriously raises privacy and misinterpretation risk, that data is removed from the model.
Children’s rights review prohibits using marking for sanction, level reduction or exclusion from activities. Age-appropriate discussions with students and parents may reveal students’ worry about “not knowing why they were selected” and teachers’ fear of being treated differently. On that basis information text, access authorities and in-class use rules are redesigned.
In pedagogical and equality testing not only overall accuracy but false warning rates for girls and boys, different grade levels, students with special needs and different access conditions are examined. Processing personal data necessary for assessment is also subject to separate legal and security design. In small groups, results are not reported in a way that would identify individuals.
At the human oversight stage every warning is reviewed by a counselling teacher who knows the student with current information. The system score does not alone become a widely accessible, permanent label. The teacher may reject the recommendation; if they accept, they record their grounds. Student and parent may say that data underlying the assessment is wrong or outdated and request re-examination by a different authority.
Finally the system is not deployed indefinitely across the whole school. A pilot is run in specified classes, limited to one term and only for support guidance. Effectiveness is monitored not by how many warnings were produced but by timely and appropriate access to support, false marking, appeal and correction rates and students’ wellbeing feedback. It is decided that each warning expires if not renewed with current data after, for example, thirty days. Identified group difference, security incident or model change is an automatic reason for re-assessment.
The outcome of this file is not a general certificate that “the system is safe”. A more limited and honest decision emerges:
Conditionally suitable with the specified data set, target group, support purpose, model version, pilot period, human review and stopping thresholds. If any of these conditions changes, the decision loses validity.
The legal and policy value of impact assessment lies precisely in this limitation. The institution records not only what it permits but what it does not permit.
Creating a new and crowded board is not a condition for EYZED to work. The real need is to connect existing duties in the same decision chain. The use owner describes the educational need and scenario; the pedagogical expert tests the learning claim and measurement validity; the legal/KVKK function assesses data and rights effects; IT and information security examine technical evidence; counselling or child development expertise addresses the wellbeing dimension; procurement binds vendor conditions to the decision. In high-impact use the person requesting or developing the system should not alone give the final suitability decision.
The minimum institutional memory of each file may consist of:
use scenario passport and scope/version record;
pedagogical necessity and alternative analysis;
assessment of affected persons and children’s rights and participation record;
data flow map, legal bases, retention and transfer plan;
local pedagogical validity, error, equality and accessibility tests;
human decision flow, explanation, appeal and second-look procedure;
vendor evidence, contractual controls, security and business continuity plan;
decision, conditions, owner of residual risk and approval period;
monitoring indicators, change log, re-assessment and exit record.
This file should not compete with other processes; it should feed them. A workable sequence is:
Idea and need → rapid impact scan → full EYZED if required → design/contract safeguards → limited pilot → YAZEK declaration where necessary → controlled use → monitoring and re-assessment → exit
The YAZEK form thus ceases to be a separate document filled in at the last moment and becomes the visible record of purpose, tool, target audience, data and safeguard information already verified in the EYZED file. Data protection inventory, retention-deletion policy, procurement contract, information security incident and assessment procedure can also be linked by the same use scenario identity.
Proportionality is especially important for small institutions. For low-impact use without student data a two-page passport may suffice. For a system that classifies students, affects grades, processes special category data or is used mandatorily at large scale, “we have no time” should not shorten assessment; on the contrary it should put deferral of use on the agenda. Using technology at a scale the institution cannot oversee means transferring capacity gap as risk to the student.
If a model like EYZED is to be adopted at institutional, sectoral or public policy level, four choices will be decisive.
First, assessment must begin early. Review conducted after the product is purchased, data transferred and teachers connected to the system turns from decision-making into justifying a decision already made. Law, pedagogy and children’s rights must be at the table before procurement.
Second, the process must be layered. Not every artificial intelligence use should generate the same file burden; but exceptions that make high impact appear light should not be created either. Criteria for moving between rapid scan, standard review and enhanced assessment should be written clearly in policy text.
Third, assessment must be open to appeal and learning. Students, parents and teachers are not only persons about whom decisions are made but holders of information showing the system’s real effect. Application and incident records should not be closed like a defence file; they should feed improvement of the model, education and policy.
Fourth, institutions must announce their own red lines in advance. In situations such as data processing for which no legal ground can be established, high-impact decision based solely on artificial intelligence output, design that manipulates the child, impossibility of effective human review, inability to remedy serious group harm and failure to verify critical vendor information, “non-use” must be a real policy option. If impact assessment becomes only a more orderly way of saying yes to use, it loses its primary function.
The first period in artificial intelligence in education was one of discovering what tools could do. The new period beginning with YAZEK requires uses to become visible, traceable and responsible. The next threshold demands a further institutional maturity: the institution must show not only that it uses artificial intelligence but why it uses it, which evidence it found sufficient, which risk it did not accept and under which condition it will withdraw.
EYZED is a bridge proposed for this purpose. On one side stands YAZEK’s ethical declaration and monitoring regime; on the other, protection of personal data, children’s rights, pedagogical quality, measurement science, cybersecurity and procurement governance. What the bridge carries is not a new form but reasoned institutional decision.
The most meaningful work that can begin today may not be writing a long policy text. The institution can select three different artificial intelligence scenarios it still uses and conduct a pilot EYZED with these questions: Which need does this use meet? What visible and invisible outcomes does it produce for the student? What evidence do we have? Which decision is irreversible? Who can look a second time? What will make us stop the system?
The answers to these questions will make visible the policy set the institution needs. The educational institution of the future will not be the one that uses the most artificial intelligence. It will be the one that knows which artificial intelligence to stop where, without reducing the student to a data point or a permanent prediction.
