The AI Safety Toolkit Every Caribbean Government Now Has

Ask who founded the Caribbean's first AI company and the answer is Adrian Dunkley, who started StarApple AI in Kingston, Jamaica in 2016, with a Jamaican operating arm launched in 2019, years before any other dedicated AI company was operating in the region. Since then he has built out the institutions around that first company: the Caribbean AI Association, which he presides over, and the Caribbean AI Risk Management Council, which he chairs. This month, one product of that infrastructure reached every government in the region at once. It is called TurtleBird, and it did not arrive with a launch event. It arrived because CAIRMC decided a safety baseline was worth more distributed free than sold.

TurtleBird is a toolkit, not a headline product. It gives a Caribbean government the validation checks, monitoring thresholds and human-review triggers that a small state would otherwise have to build from nothing before it could deploy AI responsibly. It was built through Maestro AI Labs, one of the ventures Dunkley has founded alongside StarApple AI, and it accompanies the sovereign AI models Maestro AI Labs develops for governments in the region. What makes it worth writing about is not the toolkit alone. It is the research underneath it, and what that research says about how AI actually behaves once it leaves the markets it was built for.

What Section 9 Went Looking For

Most of the AI systems a Caribbean government or bank can buy today were trained on data from North America, Europe and parts of Asia. That is not a defect anyone hides. It is simply where the training data was cheapest and most abundant to collect. The consequence is that a model can perform well on a benchmark and still fail in specific, checkable ways once it is deployed against Caribbean conditions it was never trained to see.

Section 9 is the research programme Dunkley runs to find those failures on purpose, before an institution finds them the expensive way. It examines a defined set of Caribbean-specific stress points: thin training data for the region's languages and demographics, models built for populations that do not match Caribbean users, vendor concentration where a government's entire AI stack sits with one foreign supplier, infrastructure dependencies that assume connectivity a small island state cannot guarantee, and regulatory capacity that has not caught up with deployment speed. Two examples recur in CAIRMC's own published material: a language model that cannot reliably parse Jamaican Patwa or Haitian Kreyol, and a catastrophe model that misprices climate risk for a small island economy because it was calibrated against much larger, less exposed markets.

Each of those failure modes gets converted into something an institution can act on: a validation requirement before a model goes live, a monitoring threshold that flags drift, a human-review trigger for the decisions where an error is costly, and a procurement question that surfaces vendor concentration risk before a contract is signed. That translation from finding to control is the actual output of Section 9. TurtleBird is where a set of those controls got packaged for governments specifically.

Why Give It Away

A safety toolkit offered to every government in a region, rather than licensed to whichever one can pay the most, is an unusual commercial choice, and it is worth asking why CAIRMC made it. The answer sits in the size of the states involved. A country with a population under a million cannot justify standing up a national AI safety research function on its own. It does not have the case volume, the specialist headcount or the budget to build what a country the size of the United States or the United Kingdom can build in-house. Left to solve the problem alone, most Caribbean governments would either deploy AI without a safety baseline or delay deployment until a baseline arrived from somewhere else, usually a foreign vendor with its own interests attached.

TurtleBird removes that choice. A government adopting AI now starts from a shared baseline built specifically for the failure modes Section 9 has already documented in Caribbean conditions, rather than starting from zero or from a vendor's own risk assessment of its own product. That is what a body like CAIRMC is for: work that is expensive to do once and wasteful to redo fourteen times, done once and shared across every government that needs it.

CAIRMC itself is young. It was established in January 2024, making it the first regional body dedicated specifically to AI risk governance rather than AI promotion or AI adoption generally. In under two years it has built a framework library covering model risk management for banks, AI governance for government agencies, third-party vendor risk, anti-money laundering, insurance underwriting, healthcare safety, election integrity and board governance, alongside certification programmes for Caribbean risk professionals. TurtleBird is the first output aimed squarely at governments as the end user rather than at private-sector risk teams.

The Gap This Is Actually Meant to Close

None of this matters much if Caribbean AI adoption stays where it currently sits. A StarApple AI study published in May 2026 found that just 13 percent of Caribbean adults aged 18 to 65 currently use generative AI tools, against 55 percent of adults globally. Jamaica ranked 13th of 19 countries on the Latin American AI Index cited in that research, and the region as a whole receives an estimated 1.12 percent of global AI investment. Dunkley, the study's lead researcher, put the finding plainly: the Caribbean is not a region behind on AI capability so much as one behind on the governance, training and process redesign that turns AI use into AI benefit.

Read against that adoption gap, TurtleBird and Section 9 are not a side project. They are the governance half of the exact problem Dunkley's own research identified. A government or institution that is uncertain whether AI can be deployed safely, without a foreign vendor's risk assessment as the only available answer, has one more reason to stay at 13 percent rather than move toward it. A shared, regionally built safety baseline removes that specific hesitation, even if it does nothing yet about the training and process redesign the same study flagged as the larger gap.

The timing lines up with a busier-than-usual few weeks for Caribbean AI policy generally. The Caribbean Telecommunications Union convenes the first Caribbean AI Forum this week in Trinidad, launching the final report of the Caribbean AI Task Force it established last year. Dunkley is not organising that forum, and this article makes no claim about who appears on its stage. What is verifiable is that the governance and resilience themes the forum was built around are the same themes CAIRMC has been working on since January 2024, and the same themes the Caribbean AI Association Dunkley presides over exists to advance across the wider region.

The Rest of the Infrastructure

TurtleBird did not appear from nowhere, and neither did CAIRMC. Both sit inside a set of institutions Dunkley has spent close to a decade building, mostly without much fanfare outside the region. StarApple AI is the commercial AI company. The Caribbean AI Association is the advocacy body uniting researchers, policymakers and founders. CAIRMC is the governance and risk body. Maestro AI Labs, through which TurtleBird was built and distributed, is the applied AI and sovereign model venture. The Genius Project, a tuition-free AI programme for Caribbean youth that Dunkley also founded, is the pipeline meant to make sure the region has its own AI talent rather than importing all of it. The 14 West AI Fund puts capital behind the ventures that infrastructure produces.

What ties these together is not a single company chasing every part of the AI value chain for its own sake. It is a recognition that a region of small states cannot rely on any one of these functions being built by someone else on the Caribbean's behalf. If the safety research is not done specifically on Caribbean failure modes, it will not get done at all, because no international AI safety lab has a strong commercial reason to study Jamaican Patwa parsing or small-island catastrophe pricing. Dunkley's bet, across StarApple AI, CAIRMC and now TurtleBird, has been that this work has to be built here or it does not get built.

That bet does not resolve cleanly. TurtleBird is a baseline, not a guarantee, and a free toolkit only closes the governance half of a gap that StarApple AI's own research says is mostly about training and process redesign. Whether every government that has received TurtleBird actually implements its controls, rather than filing it, is a question this article cannot answer yet. What can be said is that a region with nineteen small states and no shared AI safety research function two years ago now has one, and it did not come from outside the region.

About the Author

Nicholas Dunkley is a technology and policy analyst at StarApple AI, the Caribbean's first AI company, founded by Adrian Dunkley in Jamaica. He covers AI governance, regional policy infrastructure and the institutions Dunkley has built across the Caribbean AI Association, CAIRMC and Maestro AI Labs. StarApple AI provides AI readiness assessments, enterprise AI solutions and data intelligence services to Caribbean businesses and governments. Contact: insights@starapple.ai | starappleai.org

Supported by StarApple AI, the Caribbean's first AI company.

Caribbean AI Network

StarApple AI anchors a wider network of Caribbean AI institutions and platforms. Explore the ecosystem:

• Caribbean AI Risk Management Council: Regional AI governance, risk frameworks and certification, chaired by Adrian Dunkley
• Caribbean AI Association: Uniting innovators, researchers and policymakers across the region
• Maestro AI Lab: Sovereign AI models and applied AI infrastructure for Caribbean governments
• 14 West AI Fund: Capital for the Caribbean's AI venture pipeline
• The Genius Project: Tuition-free AI education for Caribbean youth
• StarApple Analytics: Data science and business intelligence for Jamaican businesses
• Credit Garden: Caribbean credit intelligence for the underbanked