AI Benefits Do Not Decide Who Must Bear the Risk

Every consequential technology produces a mixture of benefit, uncertainty and harm. The existence of risk does not prove that a system should be rejected. The existence of benefit does not authorise a developer, government, school, employer or church to decide by itself that other people should carry the cost.

The neglected question in AI governance is therefore distributional: who receives the benefit, who bears the residual risk, who participates in the decision and who can obtain explanation, correction, remedy or withdrawal?

A responsible decision needs more than a favourable aggregate forecast. It requires a complete chain: defined problem → plausible benefit → affected people → alternatives → distribution of benefit and risk → legitimate authority → safeguards → independent evidence → consent or public justification → remedy → review → suspension or exit.

This chain matters wherever AI enters a classroom, workplace, public service or ministry. A system can create genuine value overall while transferring privacy loss, cognitive dependence, unpaid verification, exclusion or relational harm to people with the least power to refuse it.

The newest industry statement makes the trade-off explicit

Reuters reported on 4 October 2026 that OpenAI chief executive Sam Altman said AI's benefits justify accepting some harmful outcomes and argued that broad public access and individual agency are preferable to concentrating powerful systems in a single laboratory. He contrasted that position with more restrictive approaches to AI risk and supported lighter-touch regulation. Reuters

This is a reported interview and statement of institutional philosophy. It is not an independent estimate of AI's net benefit, a specification of acceptable harm or evidence that present safeguards distribute risks fairly. The article does not establish that any particular deployment is safe or that affected communities have agreed to its risks.

The statement nevertheless clarifies a real policy choice. Zero risk is not achievable, and excessive concentration of technological power creates dangers of its own. Broad access can support learning, creativity and economic participation. But the case for access does not settle who may expose children, workers, citizens or congregants to residual risk, nor which harms are impermissible even when benefits are large.

Evidence: a leading developer has publicly defended accepting some harms in exchange for broader benefit and agency. Interpretation: institutions now need an explicit account of risk-bearing authority. The company that develops a system, the organisation that deploys it and the person affected by it do not occupy the same position and should not possess the same decision rights.

Benefit is not a permission slip

A projected benefit becomes ethically meaningful only when its object, recipient, baseline and evidence are stated. Faster drafting, lower administrative cost, greater access, improved feedback and wider translation are different benefits. Each may accrue to a provider, institution, professional or learner while imposing new work or exposure elsewhere.

An aggregate claim can conceal this distribution. A university may save marking time while students lose clear authorship rules. A public body may process cases faster while citizens struggle to understand or appeal an automated recommendation. A church may publish more content while pastoral verification and confidentiality weaken. A provider may gain data and market share while the institution absorbs correction, training and exit costs.

Leaders should therefore ask four separate questions: Does the claimed benefit exist? Is AI necessary to obtain it? Who receives it? What cost, uncertainty or lost capability is transferred to others? A positive answer to the first question cannot answer the remaining three.

The discipline is not anti-innovation. It improves innovation by directing effort towards a real problem, comparing non-AI alternatives and preventing institutions from calling a transferred burden a shared benefit.

Risk tolerance belongs to accountable governance

The NIST AI Risk Management Framework defines residual risk as the risk remaining after treatment and notes that it directly affects end users and affected individuals or communities. It calls for organisations to define risk tolerance, document risk decisions and consider impacts on people, groups, society and the environment. NIST

NIST is a voluntary framework, not a law and not a universal moral theory. Its operational contribution is important: risk tolerance is an organisational decision that must be governed and documented. It cannot be inferred from a product's availability or from a developer's general optimism.

Legitimate authority varies by context. An adult may choose a low-stakes personal tool after meaningful notice. A school acts under duties to learners and families. A public institution needs lawful mandate, proportionality and procedural fairness. A church exercises pastoral and spiritual responsibility that cannot be transferred to a vendor. A board may approve an institutional risk, but it cannot waive another person's legal rights or treat vulnerable people as experimental material.

Some boundaries should be categorical: no automated prophecy or spiritual diagnosis, no unreviewable decision that determines a person's rights, no undisclosed use of confidential counselling data, and no child-facing system without safeguarding and a competent human route. Benefits may influence design within those boundaries; they do not erase them.

Education should begin with the problem, not the product

On 2 October 2026, UNESCO described a Digital Learning Week exercise in which policymakers and educators first defined educational problems and affected stakeholders, then assessed whether AI or a non-technological intervention was appropriate. UNESCO presented ethical assessment as a design activity before deployment rather than a compliance exercise after adoption. UNESCO

This is an account of a training programme, not comparative evidence that participants later made better procurement or learning decisions. Its sequence is nevertheless sound: beginning with a product encourages institutions to invent a use and discover harms after learners are exposed; beginning with the educational problem keeps pedagogy, alternatives and affected people visible.

A school considering an AI tutor should specify the actual problem. Is feedback delayed? Are class sizes too large? Are learners struggling in a particular concept or language? Is teacher preparation insufficient? Each diagnosis leads to different interventions. Additional educators, better materials, peer learning, formative assessment or connectivity support may outperform an AI purchase in a given context.

The UNESCO IITE 2026 global report similarly describes smart education as the integration of pedagogy, governance and technology around human capability, not technology adoption alone. Its seven regional perspectives show that no single pathway fits every education system. UNESCO IITE

The learner must not become the residual-risk account

Educational AI can provide explanations, examples, language support and rapid feedback. It can also shift verification work to teachers and learners, weaken independent practice, expose personal information, reproduce inaccessible assumptions or make participation depend on devices and connectivity that families must fund.

A preregistered randomised experiment involving 405 students aged fourteen to fifteen compared note-taking, chatbot use and a combination of both. Note-taking alone and the combined approach produced stronger comprehension and three-day retention than chatbot use alone, although learners tended to prefer the chatbot. Kreijkes and colleagues

One experiment in English secondary schools cannot establish the effect of every model, subject, age group or teaching design. It does show why preference, speed and immediate task completion are insufficient evidence of learning. A system may feel helpful while an educational cost appears only when assistance is withdrawn.

Institutions should measure delayed independent performance, source verification, transfer to an unfamiliar task, correction after feedback, accessibility, teacher workload and unequal effects. If the benefit is faster completion but the residual risk is weaker durable learning, the learner is carrying the institution's experiment.

Curriculum should teach who is allowed to accept risk

AI literacy needs to move beyond prompt technique. Learners should be able to identify the decision-maker, the people affected, the claimed benefit, the material harms, the evidence, the applicable rights, the safeguards and the route to challenge a decision.

A useful assessment would present a proposal for an AI-supported learning, employment or public-service workflow. Learners would construct a risk-benefit distribution map, compare an AI and non-AI alternative, identify a harm that cannot be traded away, propose safeguards, state what remains uncertain and defend who has legitimate authority to proceed.

Assessment should include an initial unaided analysis, bounded AI-assisted challenge, source verification, revision with reasons and an oral defence using an unfamiliar case. The purpose is not to reward a predetermined pro- or anti-AI position. It is to demonstrate that the learner can examine power, evidence and consequences rather than repeat the language of the provider.

The graduate capability is practical public reasoning: the ability to say not only that a benefit exceeds a risk, but whose benefit, whose risk, under whose authority, with what remedy and for how long.

E-learning procurement needs a distributional case

A conventional business case often compares price, functionality and anticipated efficiency. An educational AI decision needs a wider record. It should include the baseline learning problem, affected groups, alternatives, full cost, data flows, teacher capability, independent learning evidence, human support, appeal and exit.

Procurement teams should separate institutional convenience from learner benefit. Automated feedback may reduce response time while increasing false guidance. Translation may expand access while introducing theological, technical or cultural error in under-tested languages. Analytics may identify learners needing support while creating intrusive profiles or unjustified inferences.

UNESCO's current global consultation on AI and education places human agency, equity, inclusion, safety, ethics, procurement, total cost, explanation and education as a common good within one governance agenda. The consultation remains open through 15 October 2026; its background papers are policy inputs, not binding standards or completed outcome evaluations. UNESCO consultation

Contracts should support incident notice, audit access, data export and deletion, correction, suspension and exit. Institutions should preserve accessible non-AI routes where an automated pathway would otherwise make education conditional on accepting unnecessary data or technological risk.

Global South agency includes the right to reject transferred risk

Global AI benefit claims often invoke productivity, access and development. These possibilities matter in settings where educators, health professionals, public officers and pastors work under severe resource constraints. Scarcity, however, can weaken bargaining power and make institutions accept systems that would face stronger scrutiny elsewhere.

The African Union Continental AI Strategy links innovation with African agency, skills, infrastructure, inclusion, culture and responsible governance. African Union Agency is not merely access to tools. It includes power to define the problem, shape evaluation, protect local knowledge, test African languages, negotiate contracts, retain institutional capability and refuse an unsuitable system.

A provider should not count an African institution's access as the benefit while treating local-language errors, connectivity costs, data extraction, professional deskilling and vendor dependence as externalities. Those conditions belong inside the adoption decision.

Regional cooperation can make this feasible. Universities, churches, professional bodies, regulators and civil society can share evaluation protocols, translated test sets, procurement clauses, incident knowledge and independent expertise. Collective capacity can reduce cost without surrendering local decision power.

Privacy harm cannot be balanced away by convenience

A person who needs education, employment, pastoral care or a public service may have little practical freedom to reject a system. Consent in such settings cannot carry the whole governance burden. Institutions need a lawful purpose, data minimisation, privacy by design, processor oversight, security, retention limits and meaningful routes for access, correction and complaint.

Kenya's education-sector guidance requires lawful and transparent processing, child-specific protections, processor due diligence, security and impact assessment for high-risk digital and remote-learning systems. Office of the Data Protection Commissioner

These duties mean that an institution cannot justify unnecessary collection merely by pointing to improved convenience. It should know which learner or employee data enter the system, where they travel, whether they are used for model improvement, who can access them, how long they remain and what happens after withdrawal or breach.

Prayer requests, counselling records, safeguarding disclosures, health information and children's data require especially strict boundaries. A ministry or school should not ask vulnerable people to trade confidentiality for faster service when a safer human or low-data route is available.

CBTE can test stewardship under uncertainty

Competency-based theological education evaluates demonstrated competence in a community of learning rather than equating completion with seat time. It integrates knowledge, character and contextual practice. CBTE framework

That model is well suited to AI governance. A ministry learner can be asked to evaluate an AI-assisted teaching, communication or administrative proposal, identify the people who bear risk, protect confidential information, test theological accuracy, establish a human referral route and defend a decision to adopt, restrict or refuse the use.

Mentors should observe the learner's response when evidence complicates the promised benefit. Does the learner disclose uncertainty, listen to affected people, revise the plan, preserve pastoral presence and accept a slower process when human care requires it? These are performances of stewardship, not merely statements about ethics.

Portfolio evidence should be redacted and proportionate. No student needs to expose identifiable counselling, safeguarding, prayer, health or children's information in order to demonstrate responsible ministry competence.

Digital ministry must not transfer spiritual risk to the vulnerable

AI may support routine administration, public information, preliminary translation, study questions and critique of a draft. The apparent benefit can be real: scarce ministerial time may be released for teaching, visitation, mentoring and care.

The transfer is legitimate only if that human return actually occurs. If saved time produces more synthetic content, more automated engagement and less accountable presence, the ministry has counted output as benefit while congregants carry relational and theological risk.

AI should never be presented as a prophet, pastor, confidential counsellor or spiritual companion. It should not diagnose spiritual conditions, make safeguarding decisions, encourage exclusive dependence or publish testimonies, sermons, fundraising appeals and pastoral messages without named human approval. Materially synthetic media should be disclosed.

Pentecostal discernment offers a necessary corrective. Power and fluency are not self-authenticating. Claims are tested through Scripture, community, character and fruit. A model has no calling, anointing, covenantal responsibility or capacity for repentance. Technical assistance must remain under human pastoral and ecclesial authority.

Boards need a Risk-Benefit Distribution Record

Every consequential AI use should have a living record containing:

1. The problem, baseline and measurable benefit sought.

2. The AI and non-AI alternatives considered.

3. The provider, model, version, data, integrations and permissions.

4. The people who receive each benefit and the evidence supporting it.

5. The people exposed to each material risk, including residual risk after controls.

6. Rights or boundaries that may not be traded away.

7. The lawful and institutional authority for the decision.

8. Participation, notice and consent appropriate to the context.

9. Safeguards, responsible owners, independent tests and disclosed limitations.

10. Explanation, correction, appeal, incident response and remedy.

11. Measures of retained human and institutional capability.

12. Review, suspension, portability and exit conditions.

The record should be revised when the model, feature, data flow, affected group or context changes. A decision about one version and purpose does not authorise every future use.

Research should measure who gains and who pays

A useful research pathway is: problem definition → participation → benefit claim → risk distribution → safeguard quality → frontline practice → educational, ministry or public outcome → remedy → trust and retained agency.

Potential mediators include educator capability, assessment design, human support, procurement power, data governance, local-language performance, audit access and leadership responsiveness. Potential moderators include age, disability, language, rurality, connectivity, shared-device use, institutional size, professional scarcity, market concentration and vendor dependence.

An African comparative study could examine four conditions: conventional human practice; unstructured general-purpose AI use; bounded AI assistance; and a governed model using a risk-benefit distribution record with independent evaluation and exit rights.

Outcomes should include delayed independent learning, contextual transfer, source accuracy, correction time, professional workload, privacy incidents, unequal effects, human help-seeking, full cost, switching capability and trust. Results should be disaggregated so an average improvement does not conceal harm concentrated among a smaller or less powerful group.

Accept only the risks you are authorised to govern

Innovation always involves uncertainty. Responsible leadership does not promise that nothing can go wrong. It refuses to use uncertainty as permission to make other people bear unnamed costs without voice, evidence or remedy.

Select one current AI-enabled activity. State the problem before the product. Identify every group that gains and every group that carries risk. Compare a non-AI alternative. Name the responsible authority. Protect boundaries that cannot be traded away. Publish material limitations. Create a route for explanation, correction and appeal. Decide in advance who can stop the use.

The decisive question is not whether AI can create more good than harm in the abstract. It is whether this institution, for this purpose, has evidence and legitimate authority to ask these people to carry this residual risk—and whether it will remain accountable when the forecast is wrong.