An AI Tutor Is Not a School

A learner opens an adaptive platform. It identifies a weak concept, supplies another explanation, adjusts the difficulty and gives immediate feedback. The learner progresses faster than a whole class could move together.

The same system does not know that the learner is hiding confusion behind correct clicks, that a family crisis has changed concentration, that a culturally unfamiliar example is being memorised without understanding or that an apparently efficient route has displaced the conversation through which judgement matures.

Artificial intelligence can tutor. A school must do more. It must decide what is worth learning, cultivate relationship and belonging, interpret evidence, protect learners, form judgement, coordinate professional responsibility and answer publicly for the meaning of its awards.

An AI tutor is not a school. It becomes educationally valuable when its bounded strengths operate inside a human-governed ecology of curriculum, teaching, formation, assessment and care.

The newest governance boundary now requires visible AI

On 2 August 2026, the wider provisions of the European Union's AI Act began to apply. Article 50 now requires people to be informed when they are directly interacting with certain AI systems and requires providers to support identification of AI-generated or manipulated content.

This is enacted regulation within the European Union, not a global education standard. The current European Commission implementation timeline also distinguishes the transparency rules now in force from the high-risk rules for sensitive uses, including relevant educational uses, which are scheduled to apply later under the amended timetable.

The distinction matters beyond Europe. A learner should know when assistance comes from a machine, but disclosure alone does not make the interaction educationally sound. Transparency answers who or what is speaking; curriculum and governance must still answer whether the advice is accurate, appropriate, fair and connected to accountable human help.

Institutions outside the European Union should not imitate legal language they do not understand or claim compliance that has not been assessed. They can nevertheless adopt the underlying practice now: identify automated interaction, name the responsible institution, preserve human review and make correction and remedy visible before AI tutoring becomes ordinary infrastructure.

The newest expansion makes the distinction urgent

On 2 August 2026, Axios reported that Alpha School plans to expand from roughly twelve campuses to about fifty during the 2026 school year. Its model assigns no more than two hours each day to AI-supported core academic work and uses the remainder for life-skills workshops and experiential projects.

This is a consequential expansion announcement. It establishes intended scale and the provider's description of its model. It does not independently demonstrate that learners cover twice the material, rank in the top one or two per cent, retain knowledge, transfer capability, flourish socially or experience the same outcomes across new campuses.

Alpha's own current pages present strong mastery and performance claims. Those are provider claims rather than independent comparative evidence. The wider expansion therefore increases the responsibility to publish methods, cohorts, attrition, costs, safeguarding, teacher or guide contributions and outcomes beyond standardised scores.

The central question is not whether an unconventional school may use AI. It is whether tutoring efficiency is being used as evidence for the whole educational model without separately examining curriculum quality, human teaching, learner welfare, social development and institutional accountability.

Tutoring and teaching are related but different practices

Tutoring normally supports a learner within a defined body of knowledge. It can diagnose an error, provide explanation, offer practice, ask a question and adjust the next task. These functions are valuable, and intelligent systems may make them available at greater scale and at times when a human tutor is absent.

Teaching includes tutoring but also interprets curriculum, reads a room, connects knowledge across domains, establishes intellectual norms, notices changing needs, moderates participation, models responsible uncertainty and decides when the planned lesson should stop because something more important has emerged.

A school adds further responsibilities: belonging, child protection, peer community, citizenship, cultural transmission, professional standards, assessment, records, remedy and public accountability. No single teacher performs these alone, and no single platform can inherit them merely by becoming adaptive.

The distinction prevents two errors. It stops institutions from dismissing useful tutoring technology because it is not human. It also stops providers and leaders from presenting a successful component as a replacement for the educational ecology that gives the component meaning.

The learning evidence is positive only under conditions

A 2026 meta-analysis of twenty-seven empirical studies reported moderately positive average effects of generative-AI conversational agents on cognitive and non-cognitive learning outcomes. The effects varied with educational level, intervention duration, knowledge type, instructional mode and the role assigned to the agent.

A systematic review of sixty-seven studies on critical and creative thinking found a similar conditional pattern. Benefits were most visible when AI was embedded in scaffolded inquiry, reflection, argument and dialogic activity. Passive or unstructured use was associated with dependency, reduced authenticity and weaker higher-order engagement.

Reviews aggregate diverse interventions, subjects and measures. Their average effects do not prove that a particular platform, school or course will work. Rapidly changing products and short study durations also limit transfer.

The evidence does not support the claim that AI tutors inevitably improve education or inevitably diminish it. It supports a narrower conclusion: pedagogy, human support, task design and context determine whether the system extends thinking or replaces it.

Accuracy can improve while intellectual initiative weakens

A quasi-experimental study involving seventy-six preservice mathematics teachers at two universities compared AI-supported and instructor-guided work. AI-supported participants produced more accurate examples but more frequently reproduced supplied content, while instructor-supported participants showed more misconceptions alongside greater independent reasoning and initiative.

The study is modest in size, involves a particular task and does not establish a universal trade-off between accuracy and originality. Its design nevertheless reveals why product quality cannot stand alone as evidence of learning.

A fluent answer may represent strong understanding, careful use of support, shallow reproduction or some combination of the three. Teachers need process evidence, explanation, transfer and independent demonstration before assigning educational meaning to the product.

The goal is not to preserve error for its own sake. It is to ensure that correction develops the learner's capacity to reason rather than making the system permanently responsible for correctness.

African evidence places AI inside teacher-led activity

A 2026 qualitative study worked with ten STEM educators from Nigeria, Botswana, Ghana, Namibia and Sierra Leone on AI-enhanced project-based learning in resource-constrained settings. Participants used speech-to-text and image tools, smartphone coding, simulation and physical prototyping.

The educators reported stronger visualisation, inclusion and transition towards competency-oriented practice. Observations, artefacts and interviews add contextual depth, but the sample is small, the study concerns educator experience and it does not establish scaled learner outcomes or cost-effectiveness across African systems.

Its most constructive contribution is pedagogical. AI did not become the curriculum or replace the educator. Teachers used accessible tools inside projects that required explanation, making, collaboration and contextual application.

Africa's educational contribution should not be reduced to proving that low-cost technology can compensate for teacher scarcity. The continent can show how intelligent tools serve collective inquiry, multilingual participation, practical competence and locally governed purposes without making scarcity the permanent design principle.

Human-in-the-loop support is a stronger African model

Kwame 2.0 offers a useful practice signal from large-scale online coding education in Africa. The bilingual teaching assistant was deployed within a human-in-the-loop discussion forum across fifteen cohorts, 3,717 enrolments and thirty-five countries. It retrieved course material and generated responses while human facilitators and peers remained part of the support system.

The authors report timely, high-quality curriculum support and describe human participants as important in mitigating errors, particularly for administrative questions. The work is currently a preprint and relies partly upon expert ratings and community feedback; it should not be treated as completed independent proof of broad learning effects.

Its architecture is nevertheless educationally serious. The tool answers within a bounded curriculum, the community can correct it and learners retain access to accountable people.

The design question changes from 'Can the chatbot answer?' to 'Can the combined system help a learner progress while keeping knowledge, correction and responsibility visible?'

Evaluate what learners do after the feedback

AI-tutor evaluation often scores whether a feedback message identifies a mistake and offers pedagogically appropriate guidance. That says little about whether the learner reads, understands or correctly applies it.

A 2026 preprint analysed more than ten thousand programming submissions with corresponding AI-tutor feedback. It found that learner engagement and correct application distinguished tutor performance in ways that message-quality ratings alone did not capture.

This is observational course evidence and has not yet completed peer review. Behavioural response also does not by itself prove durable learning. A learner may apply advice successfully in one task without retaining the underlying principle.

Institutions should therefore examine a chain of evidence: feedback quality, learner engagement, correct revision, explanation, later independent performance and transfer into a different problem. A tutor becomes educational when its assistance leads towards capability rather than repeated dependence.

Curriculum remains a human and public judgement

Personalisation answers the question 'What should this learner encounter next?' only after someone has decided what knowledge, capabilities, values and forms of participation matter.

An adaptive sequence can optimise progress through a curriculum while leaving the curriculum's omissions untouched. It may efficiently reproduce a dominant language, historical perspective, professional assumption or cultural account that the institution never examined.

Curriculum governance should therefore identify who selected the knowledge base, how local scholarship and languages enter, which competencies are intended, where disagreement is welcomed and how learners participate in evaluating the system.

Global South institutions should not receive personalisation as a technical layer placed over imported content. They should possess authority to define educational purposes, contextualise knowledge, protect community intellectual contributions and refuse optimisation towards outcomes that do not serve their learners.

Assessment must reveal assisted and independent capability

A tutoring system can improve work submitted for assessment while making it harder to determine what the learner can do independently. This is not only an authorship problem; it is a validity problem.

A natural experiment using five years of examination data from 1,066 undergraduates found a substantial change in grade distributions when a course moved from an AI-permissive take-home examination to a supervised AI-restricted format. The content, outcomes and grading structure were described as stable, but cohorts and years may still differ in unmeasured ways.

The finding does not prove that one format is inherently superior. It shows that AI-permissive and AI-restricted assessments may not measure equivalent performance, making careless comparison of grades or progression unsafe.

Programmes should deliberately sample supported work, process evidence, oral explanation, observed practice and independent transfer. The purpose is to determine which claim each assessment can validly support—not to pretend that every capability must be demonstrated without tools.

Teachers need authority over the learning relationship

A promise that AI will not replace teachers is incomplete if teachers become monitors of a vendor-designed sequence whose content, feedback rules and learner data they cannot question.

Educators need preparation to interpret system recommendations, override inappropriate pathways, recognise misconceptions the system missed and redesign the learning environment. They also need time. Automation may reduce one task while creating new work in verification, escalation, data governance and learner support.

A collaborative autoethnographic proof-of-concept in teacher education found that AI restructuring of spoken assessment feedback initially increased workload, then produced efficiencies as workflows improved. Human editing remained necessary for accuracy, tone and alignment with the rubric. The study reports a small group's professional experience rather than comparative learner outcomes.

Professional responsibility should travel with professional authority. If an educator remains answerable for learning and welfare, the educator must be able to inspect, challenge, stop and change the system.

Equity is more than a cheaper tutor

AI tutoring is often presented as a response to limited access to individual support. That potential matters in environments with large classes, dispersed learners, few specialist teachers and irregular study schedules.

Yet a low subscription price can conceal device costs, electricity, data, language exclusion, disability barriers, teacher workload and dependence upon an external provider. Learners with confident language, prior knowledge and stable connectivity may benefit first and make average results look more equitable than they are.

Equivalent low-bandwidth and human routes should remain available. Institutions should compare participation, completion, learning, cost and complaints across language, location, disability, gender, prior attainment and access conditions.

Scarcity should motivate careful innovation, not lower the educational entitlement of the people experiencing it. A learner in a remote or under-resourced community deserves human recognition, contextual knowledge and remedy as fully as a learner in a wealthy campus.

Theological education must form interpreters, not answer receivers

An AI tutor can help a theology learner locate a passage, compare interpretations, rehearse vocabulary, receive low-stakes practice and identify questions for further study. It can also fabricate sources, flatten traditions and supply religious language that sounds mature without being personally understood.

Theological teaching attends to how a learner reads Scripture, receives correction, participates in community, relates knowledge to character and remains answerable for ministry judgement. These are not decorations added after information transfer; they are part of what theological education claims to form.

Competency-based theological education can use digital tutoring within mentored knowing, being and doing. Evidence should include observed ministry, artefacts, mentor testimony, oral defence, reflection and transfer. A platform's mastery score cannot certify spiritual maturity or responsible pastoral practice.

The appropriate boundary is constructive: use intelligent assistance to extend access and practice while keeping interpretation, formation, assessment and vocational discernment in accountable human community.

Digital ministry needs a human destination

A ministry chatbot can answer routine questions, locate public resources, explain a programme and guide a user towards an appropriate contact. It should not be represented as a pastor, spiritual companion, prophetic voice or bearer of personal knowledge it does not possess.

Recent Christian-ethics scholarship proposes critical openness rather than automatic embrace or rejection of generative AI, using the biblical figures of Balaam and Cyrus to examine how an instrument may be used without becoming morally or spiritually authoritative. The article is theological analysis rather than outcome research.

Religious-education scholarship similarly places formation and community ahead of technological convenience. Digital assistance should return people to Scripture, embodied fellowship, responsible teaching and competent care.

Every automated ministry route should identify where a human becomes responsible. Safeguarding, distress, abuse, confidential pastoral matters and consequential spiritual guidance must leave the tutoring interface and reach an authorised person.

Govern the whole learning ecology

Boards and academic bodies should not approve an AI tutor as though it were one more neutral textbook. The system changes feedback, data flows, teacher work, learner behaviour, assessment and the institution's dependence upon a provider.

Governance should know the educational purpose, curriculum boundary, training data or retrieval sources, error controls, teacher authority, learner information collected, provider interests, whole cost, accessibility, incidents, complaints and exit arrangements.

Outcome reporting should distinguish access, activity, immediate task performance, durable learning, transfer, well-being and institutional capability. Provider dashboards should not become the only evidence reaching the governing body.

Where a system affects progression or support, learners should understand its role, correct relevant information, obtain human review and challenge a consequential decision. Innovation becomes trustworthy when authority and remedy remain visible.

Use a ten-part human teaching covenant

Before adopting an AI tutor, an institution can establish ten connected commitments.

Purpose defines the educational problem. Curriculum identifies what is taught and whose knowledge is represented. Tutor boundary specifies the tasks delegated to the system. Teacher authority preserves professional oversight and override. Learner agency requires explanation, questioning and appropriate refusal.

Relationship identifies the human community surrounding learning. Assessment distinguishes assisted and independent capability. Protection and equity cover data, safeguarding, language, disability and access. Governance names responsible bodies, provider interests and remedy. Evidence states what outcomes will justify continuation, redesign, scale or stop.

The covenant should accompany the curriculum and procurement decision rather than appear after implementation. Its purpose is not to make every experiment slow. It is to keep the institution responsible for the education it has chosen to mediate through technology.

Keep the tutor inside the school

AI tutors can make practice more responsive, extend feedback beyond ordinary hours, support multilingual access and help teachers see patterns that would otherwise remain hidden. Those are genuine possibilities worth investigating.

The tutor still operates inside purposes, knowledge and relationships it did not authorise. It cannot determine what a community ought to become, accept moral responsibility for a learner, confer a trustworthy qualification or replace the human recognition through which education becomes belonging and vocation.

Africa should participate in this transformation as a source of educational theory and design. Teacher-led project learning, human-in-the-loop support, multilingual community and competence demonstrated in context offer lessons to every education system—not adaptations at the edge of someone else's model.

Use the tutor. Bound its role. Prepare the teacher. Protect the learner. Examine independent capability. Keep community and remedy near. Then intelligent assistance can strengthen a school without being mistaken for one.