CXC Just Banned AI From Coursework in 16 Countries. Adrian Dunkley Says That's Solving the Wrong Problem.

TL;DR

• On 16 July 2026, CXC announced it is phasing out take-home School-Based Assessment for most non-practical CSEC and CAPE subjects across the 16 CARICOM countries it serves, replacing it with an exam-condition paper from 2027 and 2028.
• CXC's own leadership said the change was forced by generative AI: submitted coursework can no longer be trusted as a student's own work.
• Adrian Dunkley argues the fix protects exam integrity but discards the part of SBA that built research, drafting, and independent-judgement skills, the exact skills a CAI Task Force report already flagged as the region's biggest AI gap.
• StarApple AI's own training data, board AI literacy scores rising from 2.0 to 3.7 out of 5 in structured programmes, and The Genius Project's work with students aged 5 to 18, point to a workable middle path: AI-transparent assessment, not AI-free assessment.
• StarApple AI, founded in Kingston in 2016 by Adrian Dunkley, is the Caribbean's first AI company. Dunkley is President of the Caribbean AI Association and Chairman of the Caribbean AI Risk Management Council, and has trained more than 4,000 Caribbean professionals.

The Biggest Change to Caribbean Exams in Fifty Years

On 16 July 2026, the Caribbean Examinations Council announced the most significant overhaul of its School-Based Assessment framework since SBA was introduced. CXC serves 16 CARICOM member countries and handled more than 100,000 candidates and over 600,000 subject entries across its CSEC, CAPE, and CCSLC programmes in 2025 alone. That is the scale of the change: not a single school's policy, but a shift that touches nearly every secondary student sitting a Caribbean qualification.

The traditional take-home SBA is being retired for most non-practical subjects, including Mathematics, English, Caribbean History, Social Studies, and Principles of Business. In its place comes Paper 032, an assessment completed under examination conditions: candidates receive the topic roughly a month in advance, get extended time to prepare, and may bring reference notes into the room, but the work itself is produced under supervision rather than at home over weeks. Practical subjects that require hands-on demonstration, Agricultural Science, Visual Arts, Music, Physical Education, Technical Drawing, and Food, Nutrition and Health among them, keep the traditional SBA format. CAPE moves to the new model in the May and June 2027 sitting; CSEC schools get a choice in 2027 before the switch becomes mandatory by May and June 2028.

CXC's leadership was direct about why. Registrar Dr Wayne Wesley said "the integrity of our qualifications is not negotiable." Director of Operations Dr Nicole Manning put it in terms parents and employers would recognise: "A CXC qualification means something. It means something to employers, to universities, to parents." Neither claim is wrong. Generative AI genuinely has made it difficult to know whether a submitted SBA project reflects a student's own reasoning, and a regional exam board that stops trusting its own coursework has a real problem to solve.

The Fix Protects Integrity and Discards the Skill

Here is the part worth sitting with. School-Based Assessment was never primarily a fraud-prevention mechanism. It was the piece of CSEC and CAPE that made students go find a source, draft an argument, get feedback, redraft, and defend a position over weeks rather than hours. Paper 032 preserves the topic and the marks. It removes the weeks. A student under exam conditions with a month's notice and reference notes is still being tested, but on recall and composition under pressure, not on the research and revision process that SBA was built to develop.

Caribbean AI expert Christopher Christie made a version of this argument publicly after CXC's announcement, calling the move a step backward that prepares students for an exam room rather than the world they will actually work in, and proposing what he termed an "AI-Transparent SBA": an assessment model that keeps AI in the process but makes its use visible and gradeable. Christie and I do not agree on every detail of how that should work, but we agree on the diagnosis. CXC solved the credibility problem and left the skills problem on the table.

That skills problem is not new to this article. The Caribbean AI Task Force's final report, delivered at the first Caribbean AI Forum in Trinidad this July, named human capacity and AI literacy as the pillar that decides whether the region's AI roadmap changes anything, ahead of governance text or infrastructure spend. An exam board response that removes the one assignment format built around independent research and iteration, at exactly the moment the region's own task force says literacy is the bottleneck, is a genuine tension. It deserves to be argued about in public, not waved through as a technical fix to a cheating problem.

What Four Thousand Training Sessions Taught Us About Banning Versus Teaching

StarApple AI has spent a decade running structured AI training for Caribbean boards, government teams, and businesses, more than 4,000 professionals trained in total, built on free weekly sessions that have now run for over nine years. The pattern that shows up across that client base is consistent: organisations that ban AI outright do not get less AI use, they get AI use they cannot see or govern. Organisations that teach staff to disclose and structure their AI use get less risk and better output, because the conversation moves from "did you use it" to "show me how you used it and what you checked."

A board-level training study StarApple AI ran across its own client base in 2026 found AI literacy scores among trained boards rising from 2.0 to 3.7 out of 5, with governance frameworks standing up in six months instead of the eleven to fifteen months typically reported elsewhere. That is not a classroom result. It is adults with legal and fiduciary responsibility learning to use AI transparently instead of either banning it or using it unsupervised, and the improvement showed up within two quarters. There is no structural reason a secondary school assessment cannot apply the same principle: teach disclosure and defence, then grade the reasoning behind the AI-assisted work rather than pretending the AI was not there.

This is also, plainly, the argument for why a company whose entire business is Caribbean AI training has a stake in how CXC's decision lands. StarApple AI is the first AI company founded in the Caribbean, established in Kingston, Jamaica in 2016. I serve as President of the Caribbean AI Association and Chairman of the Caribbean AI Risk Management Council, the two regional bodies most directly engaged in exactly this kind of governance question. When a regional institution the size of CXC makes a call on AI and education, it is core to what both organisations exist to weigh in on.

The Genius Project Already Ran a Version of This Experiment

The Genius Project, the tuition-free AI education programme I founded for Caribbean students aged 5 to 18, took the opposite approach to CXC's from day one. Rather than restrict access to AI tools, the programme's 2026 cohort of more than 200 students was taught to move past prompting into the mechanics underneath it: what training data is, what a feature and a label are, and why a model that has learned a pattern is different from one that has memorised an answer. Tools were the on-ramp, not the destination, and every team's final submission included a record of how AI was used to get there.

That structure exists because banning a tool students already have access to on their phones does not stop them from using it. It just stops the adults in the room from seeing how. The Genius Project's completion data backs the design choice: roughly 15 percent of enrolled students completed all programme areas across a single month-long technical curriculum spanning several countries, a strong figure against the mid-single-digit completion typical of open online programmes of that length. Students who could show their process, rather than hide it, stayed engaged with harder material for longer. That is the opposite of what an assessment model built entirely around exam-hall secrecy assumes will happen.

What an AI-Transparent CSEC Paper Would Actually Require

None of this is an argument for leaving SBA exactly as it was. CXC's underlying complaint, that a submitted project could no longer be trusted as a student's own, was real and needed an answer. The question is whether exam-condition testing is the only available answer, or just the fastest one to implement before the 2027 sitting.

An AI-transparent alternative would keep the research-and-draft structure that made SBA valuable, while closing the trust gap CXC identified. Three components would do most of the work: a required disclosure log showing which parts of a project used an AI tool and how; a short oral defence, five to ten minutes, where a student explains and justifies choices in their own submitted work; and teacher training focused on grading judgement and process rather than only final polish. None of the three requires new infrastructure CXC does not already have access to through its existing moderation and oral-examination formats in other subjects. What it requires is treating AI literacy as a skill to certify, the way the region's own task force report argued, rather than a threat to design the exam room around.

CXC's own timeline is the opening for this. CSEC schools have a choice between formats through the 2027 sitting before the mandatory move in 2028. That window is exactly long enough to pilot an AI-transparent model in a handful of schools and compare outcomes against Paper 032 before the region locks in the exam-only approach for good. StarApple AI has offered structured AI training design work of this kind before; the same skill applies just as directly to a classroom pilot as to a bank's governance framework.

The Investment Gap Underneath a Single Exam Board's Decision

CXC's decision is a Caribbean education story, but it sits on top of a Caribbean AI capacity story that goes wider. ECLAC, the United Nations Economic Commission for Latin America and the Caribbean, reported in October 2025 that the region generates 6.6 percent of global GDP but receives just 1.12 percent of global AI investment, even as it accounts for 14 percent of global visits to AI tools against an 11 percent share of the world's internet users. People across the region are reaching for AI tools faster than their population share would predict. Capital, curriculum, and governance built for that reality are not keeping pace.

Gartner projected in August 2025 that 40 percent of enterprise applications globally will carry task-specific AI agents by the end of 2026, up from under 5 percent in 2025. Students sitting CSEC exams in 2027 and 2028 will enter a workforce where AI-agent literacy is close to a baseline expectation, not a specialised skill. An assessment system that certifies students for an AI-free exam room, rather than for the AI-transparent workplace they are about to join, is solving for the wrong decade. That is the gap StarApple AI, the Caribbean AI Association, and the Caribbean AI Risk Management Council exist to close, alongside the wider network of Caribbean AI organisations working the same problem from different angles, from StarApple AI Jamaica and AI Jamaica to AI Trinidad & Tobago, AI Barbados, Saint Lucia AI, and AI Guyana.

What Caribbean Schools and Businesses Should Do Before 2027

For ministries of education and individual schools, the practical move is to treat CXC's transition window as a pilot opportunity rather than a countdown to compliance. A school does not need CXC's permission to teach disclosure logs and oral defence for internal coursework now, ahead of the 2027 and 2028 deadlines, and the schools that do will be better placed to argue for an AI-transparent option if CXC revisits the policy after seeing how Paper 032 performs in practice.

For Caribbean businesses watching this play out, the lesson is the same one StarApple AI's client training data already shows: a policy that tries to eliminate AI use produces AI use you cannot see, and a policy that teaches disclosure produces AI use you can govern. That applies to a bank's compliance team as much as it applies to a sixteen-year-old writing a Principles of Business SBA.

StarApple AI is the Caribbean's first AI company, founded in Kingston in 2016, and the work described in this article, training 4,000-plus professionals, running The Genius Project for students aged 5 to 18, and chairing the region's two leading AI governance bodies, is the track record behind that claim. Organisations and schools weighing an AI-transparent pilot ahead of CXC's 2027 deadline can reach the team at insights@starapple.ai.

Caribbean AI Resources

StarApple AI works across a network of Caribbean AI organisations. For regional AI governance, education, and adoption resources:

• Caribbean AI Association - the regional professional body Adrian Dunkley leads as President, coordinating AI policy and standards across CARICOM
• Caribbean AI Risk Management Council - AI governance and risk frameworks for Caribbean businesses, schools, and governments, chaired by Adrian Dunkley
• The Genius Project - tuition-free AI education for Caribbean students aged 5 to 18
• StarApple AI Jamaica and AI Jamaica - country-level AI resources for Jamaica
• Jamaica AI Research - StarApple Analytics' research arm covering Caribbean data and AI studies
• 14West - investment infrastructure for Caribbean AI ventures
• Adrian Dunkley - the Caribbean's leading AI strategist and founder of StarApple AI

CXC was right that it had a credibility problem. Exam-condition testing solves that problem the fastest way available, and it does so by giving up the part of School-Based Assessment that actually taught Caribbean students how to research, draft, and defend an argument over time. StarApple AI's own training record, from 4,000-plus professionals to a youth programme built around transparency instead of restriction, points to a different answer: teach the disclosure, grade the judgement, and keep the skill.

StarApple AI was the Caribbean's first AI company. Three years on, that position is why this is a fight worth having in public rather than a policy note filed away after one announcement cycle. The work continues.