AI Must Not Manufacture Competence

An institution can receive a polished assignment, a convincing lesson plan, an eloquent sermon outline or a technically complete portfolio and still possess weak evidence of competence. Generative artificial intelligence has widened the distance between the quality of a product and the capability of the person whose name appears on it.

The central assessment question is therefore not whether artificial intelligence touched the work. It is whether the evidence justifies the claim the institution intends to make. If a school records that a learner can reason mathematically, a college certifies that a minister can interpret Scripture, or a professional body concludes that a candidate can exercise safe judgement, the performance used for that decision must remain attributable, explainable and sufficiently independent.

AI may help a learner practise, receive feedback, translate instructions, test an argument or organise evidence. It becomes educationally corrupting when it produces the very reasoning or performance that the assessment exists to observe and the institution then reports the machine-assisted result as the learner's competence.

Kenya's current assessment moment makes authorship consequential

Kenyan education reporting published on 25 September states that the Kenya National Examinations Council warned teachers and other people not to use artificial intelligence to complete or assist with school-based assessment tasks intended to demonstrate learners' own competencies. It reports that the September assessment tools for several grades become accessible through the Competency-Based Assessment Portal on 28 September, with score uploads due by 23 October. Education News

Evidence status: the time-sensitive warning is reported journalism summarising KNEC guidance. Schools should use the official KNEC circulars and guidelines page and the official Competency-Based Assessment Portal for the operative documents, tools and instructions. The report does not establish how frequently adults or AI have produced assessed work, whether malpractice has increased, or which safeguards will improve reliability.

The underlying principle is nevertheless sound and urgent. In competency-based education, the score should represent what the learner demonstrated. A teacher may explain a task, create suitable conditions, observe performance and apply a rubric. The teacher must not replace the learner and then certify the replacement as evidence. AI does not change that principle; it makes the substitution easier to conceal and therefore increases the need for sound assessment design.

The product and the evidence are not the same thing

A submission is an artefact. Assessment evidence is the defensible connection between that artefact, the process that produced it and the capability the institution claims the learner possesses. Before generative AI, extensive help from tutors, relatives, peers or commercial services could already weaken this connection. AI increases the scale, speed and apparent coherence of outsourced performance.

A 2026 systematic conceptual review argues that AI-rich assessment should move from authentic products towards authenticated processes. Its central concern is the inference from observable work to the learner's capability: a realistic-looking product can remain weak evidence when judgement, verification and responsibility are invisible. Authenticated processes review

This is a conceptual review rather than a trial proving that one design always produces better learning. Its contribution is analytical. Educators should ask what the task permits them to infer, what alternative explanation AI assistance introduces, and which additional evidence would make the judgement fairer. The response is not maximal surveillance. It is better alignment between the competence claimed and the evidence collected.

Learning performance can improve while learning itself weakens

A preregistered field experiment involving nearly one thousand secondary-school mathematics learners compared ordinary resources, a general-purpose GPT interface and a safeguarded GPT tutor. Access to AI improved practice performance substantially. When AI was removed for the examination, however, learners who had used the unrestricted interface performed worse than the control group; the safeguarded tutor largely mitigated the harm. Bastani and colleagues

This was one subject in one school context, using particular interfaces and safeguards. It does not prove that every use of generative AI reduces learning or that the same effect will appear in theology, language, counselling or professional formation. It does demonstrate why success while the tool is present cannot by itself establish learning.

Assessment systems should therefore distinguish assisted performance from retained capability. If AI is part of instruction, learners still need an opportunity to explain, transfer and perform with assistance reduced or absent. Otherwise an institution may reward a successful human-machine interaction while making an unsupported claim about the human alone.

Assessment begins by defining the claim

Every consequential task should begin with a clear statement of what successful evidence would justify. Is the learner being assessed on recall, interpretation, source judgement, communication, tool supervision, pastoral presence, collaborative problem-solving or independent performance? These are different claims and require different conditions.

If the outcome is the ability to use AI responsibly in professional practice, forbidding the tool may make the task less valid. If the outcome is the ability to formulate an initial interpretation, calculate without assistance or respond personally to a pastoral situation, allowing the system to generate the answer may destroy validity. The role of AI should follow the learning outcome, not the other way around.

A useful assessment record states: intended competence, permitted assistance, evidence produced directly by the learner, process evidence, verification method, independent transfer task, judgement criteria, responsible assessor and appeal route. This makes the educational logic visible before a disputed submission forces the institution to invent it retrospectively.

A task-level AI rule is more useful than a general slogan

Institution-wide statements such as use AI responsibly or do not use AI to cheat are too vague for learners and educators. The same tool may be prohibited during an initial attempt, permitted for formative feedback, required during a professional simulation and unavailable during later independent verification.

The University of Manchester's policy, effective from 1 September 2026, requires each summative assessment to be assigned one of four categories: AI prohibited, minimal, permitted or integrated. Academic staff make the disciplinary decision, and the category must be communicated in the assessment brief and learning platform. The university describes the framework as a workable starting point that will be reviewed after its first semester. University of Manchester

This is an institutional policy, not outcome evidence. Its transferable strength is that permission becomes task-specific and visible. Global South institutions should adapt that principle to their curricula, languages, connectivity and professional responsibilities rather than import the four labels mechanically. A useful rule explains not only what is allowed but why the boundary protects the intended learning.

Use an evidence sequence that makes learning visible

A strong sequence begins with direct encounter and an initial learner attempt. Bounded AI may then challenge, question, translate or provide feedback. The learner verifies claims and sources, revises with reasons, discloses material assistance and later applies the learning to a new situation. An oral, practical or delayed task checks whether the capability remains when prompts, examples and system support are reduced.

The sequence can be expressed simply: intended outcome; initial attempt; bounded assistance; verification; reasoned revision; disclosure; unfamiliar transfer; independent performance; human judgement; appeal. Not every low-stakes activity needs every stage. The greater the consequence of the score or credential, the stronger the evidence should become.

Digital platforms can support this without capturing every keystroke. They can stage access to materials, preserve selected drafts, collect a concise declaration, schedule an oral defence and record rubric-based observations. Data minimisation matters. Authenticating a process should not become a licence for continuous monitoring of learners' private devices, conversations or unrelated intellectual work.

Teachers must not become hidden co-authors of learner evidence

Teacher support is essential, particularly for younger learners, learners with disabilities and those encountering a task in an additional language. Good facilitation explains expectations, provides accessible resources, gives formative feedback and creates a fair opportunity to demonstrate competence. It does not complete the performance and then attribute it to the learner.

AI complicates this boundary because a teacher can rapidly generate model answers, project text, feedback or scoring explanations. The professional question is not whether the teacher used a tool. It is whether the teacher's action changed the authorship of the evidence or the fairness of the judgement. A generated rubric draft that the teacher validates is different from generated learner work submitted for a score.

School leaders should therefore prepare teachers before policing students. They need examples of permitted preparation, prohibited substitution, accommodations, documentation, moderation and response to suspected malpractice. A warning without workload support, reliable instructions and professional learning risks inconsistent enforcement across classrooms.

Fairness requires more than detecting prohibited use

AI detectors cannot prove authorship reliably enough to carry a consequential decision alone. Language, writing style, disability, prior educational opportunity and the use of legitimate editing tools can all shape the signals that detection systems interpret. A school should never convert a probability score into an automatic accusation, mark or sanction.

A fair inquiry begins with the stated task rule and the evidence needed to establish the learning outcome. It may examine selected drafts, source notes, version history, oral explanation, prior work, observed performance and the learner's response. The institution should state the concern, protect confidentiality, allow explanation, give a reasoned decision and provide an accessible route for review.

The purpose is not to catch as many users as possible. It is to protect trustworthy assessment while avoiding false accusation. Better task design and visible reasoning usually provide more educational value than expanding surveillance after submission.

E-learning should preserve low-bandwidth and accessible evidence routes

An AI-resilient assessment can still be inequitable if it assumes continuous connectivity, a personal laptop, paid subscriptions or fluent use of a dominant language. A learner on a shared phone may leave less version history than a learner using an integrated platform. An oral defence scheduled synchronously may exclude someone facing unreliable electricity or caring responsibilities.

Equivalent pathways should assess the same competence with comparable support and status. Downloadable instructions, asynchronous audio, supervised local performance, mentor observation and scheduled telephone or in-person verification can complement platform-based evidence. Accessibility should be designed into the task rather than treated as an exception after failure.

The OECD Digital Education Outlook 2026 distinguishes AI-assisted performance from learning and calls for uses grounded in learning science, teacher agency and rigorous evaluation. OECD For resource-constrained institutions, the smallest sufficient technology is often the most responsible starting point.

CBTE needs witnessed competence, not generated portfolios

Competency-based theological education integrates knowing, being and doing in ministry context. A learner's portfolio can organise evidence, but the portfolio is not the competence. AI may help the learner structure reflection, compare interpretations or prepare questions; it cannot demonstrate character, pastoral presence, teachability, safeguarding judgement or faithful response when a real situation changes.

Mentors and ministry communities should witness performance over time. Evidence can include observed teaching, supervised pastoral encounters, feedback from people served, correction after error, theological explanation and a new performance in a different context. Consequential judgements should be moderated so that one mentor's affection, authority or technological enthusiasm does not determine the whole result. CBTE framework

AI disclosure remains necessary where assistance is material, but disclosure alone does not authenticate competence. The learner should be able to defend the interpretation, explain rejected suggestions, identify confidential information that was withheld and act responsibly without the system. Formation is demonstrated through accountable practice, not the fluency of a generated reflection.

Digital ministry must not convert synthetic output into spiritual evidence

A ministry may use AI to organise public information, translate routine communication, test the clarity of a lesson or generate questions for study. It should not present generated prophecy, testimony, pastoral advice or theological claims as evidence of a person's calling, spiritual maturity or encounter with God.

The distinction matters when churches train leaders. A candidate may submit an eloquent sermon produced largely by a model and still be unable to interpret the text, understand the congregation, respond to correction or speak truthfully when the prepared plan fails. Assessment should protect direct engagement with Scripture, accountable use of scholarship, contextual discernment, embodied delivery and response to the community.

Sensitive prayer, counselling, health, safeguarding and children's information should remain outside unapproved systems. Disclosure cannot make unsafe processing acceptable. No assessment convenience justifies exposing a vulnerable person's story to establish somebody else's competence.

Global South institutions need authority over what counts as competence

Imported AI systems embody languages, datasets, professional assumptions and patterns of explanation formed elsewhere. If institutions let those systems define exemplary writing, credible sources or professional judgement, they may quietly transfer curriculum authority away from teachers, communities and local knowledge traditions.

African institutions should specify the evidence their own educational, professional and theological purposes require. This includes competence in African languages, contextual application, ethical relationship, community accountability and the ability to work under actual conditions of connectivity and resource constraint. A system's fluent response should remain a proposal for examination, not the hidden standard against which African learners are judged.

The African Union Continental Artificial Intelligence Strategy connects adoption with African agency, skills, culture, infrastructure, inclusion and responsible governance. African Union Assessment sovereignty is part of that agency: institutions must retain the authority and capability to determine what a person needs to know, become and do.

Boards should govern the claims their credentials make

A governing body does not need to approve every assignment, but it is responsible for the integrity of the institution's awards and public claims. It should know how programmes define competence, where AI may enter the evidence chain, which decisions require independent performance, how assessors are calibrated, how unequal effects are monitored and how disputed judgements are reviewed.

Useful assurance measures include the proportion of major tasks with explicit AI rules, moderation findings, delayed or unfamiliar transfer performance, source-verification errors, substantiated and overturned misconduct decisions, accessibility concerns, assessor workload, privacy incidents and the performance of graduates in supervised practice. Detector usage and declarations completed are control measures, not learning outcomes.

Boards should also ask whether revenue, enrolment pressure or completion targets reward weak evidence. An institution that knowingly awards competence it has not observed transfers risk to congregations, employers, the public and the learner. Quality assurance must follow the claim beyond the platform and into the contexts where people rely upon it.

Research should test whether assistance transfers into independent capability

A useful research pathway is: intended competence; assistance design; learner effort; process evidence; verification and revision; independent transfer; professional performance; trust. Potential mediators include prior knowledge, teacher and mentor capability, feedback quality, assessment clarity, disclosure safety, data governance and access to human support.

Language, disability, connectivity, shared-device use, learner age, discipline, institutional size, authority distance and task stakes may moderate the relationship. An African comparative study could examine conventional support, unrestricted general-purpose AI, bounded AI assistance and a complete evidence-centred model with delayed independent performance.

Outcomes should include source accuracy, reasoning quality, contextual transfer, unsupported performance, correction after feedback, learner confidence, false accusation, teacher workload, privacy burden, unequal effects and stakeholder trust. Immediate grades are insufficient. The research question is whether assistance becomes capability that remains accountable when the assistance changes or disappears.

Use a competence-evidence record

For each consequential task, record the competence being assessed, the context in which it matters, the learner's required contribution, permitted and prohibited assistance, protected information, process evidence, verification, transfer task, responsible assessor, moderation, feedback, appeal and retention period.

Apply the record proportionately. A low-stakes practice quiz may require only a clear instruction and feedback. A final ministry performance, professional placement or public credential may require observed practice, multiple sources of evidence, a new-context task and independent moderation. The design should become stronger as the consequence of error grows.

Review the record with learners and educators after use. Did it produce credible evidence? Did the AI boundary support learning? Were some learners disadvantaged? Did verification consume more time than expected? What should be revised, suspended or removed? Assessment integrity is a continuing institutional capability, not a one-time policy statement.

Certify the person, not the performance supplied to the person

Begin today with one important assessment. State the exact claim the institution will make if the learner passes. Identify which part of the evidence must come directly from the learner and which assistance can deepen learning without replacing that evidence. Add one unfamiliar or delayed performance that tests transfer.

Tell learners the rule before they begin. Give educators examples and workload support. Protect confidential data. Use selected process evidence rather than indiscriminate surveillance. When a concern arises, investigate fairly and allow human explanation and appeal. When the evidence cannot support the claim, revise the assessment rather than pretending that a polished product proves competence.

Artificial intelligence can support formation. It can widen access to feedback, help learners practise and expose them to alternative questions. It must not manufacture the very capability that a school, college, church or professional body certifies. A trustworthy institution awards what it has good reason to believe the person can actually do.