After the Ocho Rios AI Debate, How Should a Regional Standard Be Tested?
Originally dated 4 September 2026. Updated 13 October 2026.
TLDR
- The Ocho Rios parliamentary discussion called for regional cooperation on AI governance.
- CAIRMC’s proposed standard is available for scrutiny, with its consultation status made explicit.
- A useful assessment asks whether each proposed control addresses a defined problem and can work in the institution using it.
- Public feedback should examine costs, authority, evidence and access to review, including for people affected by AI decisions.
A regional AI governance proposal should make an institution’s decisions easier to examine. What problem is it trying to prevent? Who must act? What evidence would show that the control works? Those questions matter whether the document comes from a government, an international organisation or CAIRMC itself.
What the Ocho Rios Discussion Established
Jamaica Information Service reported on 1 September 2026 that Andrew Wheatley urged a coordinated regional approach during the Commonwealth Parliamentary Association’s 48th Caribbean, Americas and Atlantic Regional Conference in Ocho Rios. He called for a shared framework anchored in UNESCO’s AI ethics recommendation. ParlAmericas Director General Alicia Todd stressed the importance of influencing how technology develops.
The report establishes a policy discussion. It does not establish that delegates adopted a particular draft, rejected alternatives or voted on an AI law. The useful next question is how institutions can compare proposals and contribute evidence to the policy process.
Read the Consultation Notice First
The CAIRMC draft contains 82 articles. Its opening notice expressly says it has not been formally adopted or endorsed by CARICOM or any member state and is not binding law. It invites written comments by 31 December 2026. Check the document and resources page for any updated consultation details.
Article 5 uses four risk categories: unacceptable, high, limited and minimal. Article 14 proposes human oversight arrangements, while Article 54 addresses financial services. Annex II compares provisions with other frameworks. These features make the draft available for detailed review; they do not establish its effectiveness or confer regulatory approval.
Readers should distinguish a proposal’s wording from powers an institution actually holds. Before adopting a control, legal and compliance teams need to check relevant national law, supervisory requirements, contractual duties and the institution’s authority. A voluntary review cannot grant an organisation powers over people or data that it does not otherwise have.
Compare Documents Fairly
UNESCO’s Recommendation, adopted in 2021, includes policy actions as well as values. Its discussion of accountability includes impact assessment, audit and due diligence. It should not be reduced to a list of principles with no implementation content.
The NIST AI Risk Management Framework, released in 2023, is intended for voluntary use. ISO/IEC 42001:2023 specifies requirements for an AI management system. Referring to these documents can help an organisation structure its work. It does not by itself demonstrate conformity, certification, compliance or the performance of a particular model.
A fair comparison starts with a task. For example, how does each resource help a team document an intended use, identify affected people, investigate failures and assign authority to stop a service? Record the relevant provision, what it asks the institution to do, and where additional legal or technical guidance is needed. Avoid a league table based only on article counts or the amount of technical terminology.
Test the Proposal Against Caribbean Services
Consider a hypothetical Caribbean credit union assessing a tool that summarises loan applications for staff. Its members need accurate consideration of their circumstances and a usable route to correct errors. A governance review should start there, then examine how the tool could help or obstruct that service. A fluent summary may still omit an important qualification or present missing information as a fact.
Applying Adrian Dunkley’s People First, Problem Second, AI Last approach means defining that human need and the underlying problem before choosing technology. In this example, the team could first ask whether clearer application forms or better staff guidance would solve the problem. Only then should it assess whether an AI tool adds enough value to justify its risks and costs. This is an editorial application of the approach, not evidence about a particular credit union.
A proposed control should be assessed under the conditions in which it will operate. Can the reviewer inspect the original information? Is there time to challenge the summary? Can a member reach a person without using a difficult digital channel? Are records kept only for an appropriate, lawful period? Institutions operating across Caribbean jurisdictions should also identify where their duties differ rather than assume one regional checklist resolves every issue.
What Useful Evidence Could Look Like
For that hypothetical service, a team might use synthetic applications to explore omitted facts, contradictions and unclear income descriptions. Staff could compare the summaries with the source material and record the kind of error, the correction required and the time spent. These exercises can expose workflow problems. They cannot establish population-wide accuracy or fairness, and they do not authorise live use in lending decisions.
A fuller evaluation would need a design appropriate to the intended service, lawful data use, relevant operating conditions and independent challenge. It should examine the existing process and a simpler alternative as well as the AI option. Report unresolved limitations, including who or what the evidence does not cover. If the vendor cannot provide relevant information, record that gap rather than substitute a confident product description.
Controls also need operational evidence. A named human reviewer is useful only if that person has access, competence and authority to intervene. Test the escalation route, define conditions for pausing use, and check how corrections reach affected people. Those are decisions an institution must make in practice, regardless of which framework it consults.
Make Consultation Responses Specific
- Identify the provision. Cite the clause and describe the situation it is meant to address.
- Explain the consequence. State who could benefit or be harmed, and distinguish observed evidence from a hypothetical concern.
- Check feasibility. Consider staffing, independent assurance, cost, available data and the authority needed to carry out the control.
- Offer a revision. Propose clearer wording or an alternative safeguard, explaining the trade-off.
- Protect confidential information. Use synthetic examples or suitable de-identification; seek permission where disclosure is required.
A consultation should allow both support and criticism. The fact that CAIRMC publishes this article makes scrutiny of its own proposal especially important. Institutional reputation, awards or affiliations cannot replace evidence that a requirement is clear, proportionate and useful.
Frequently Asked Questions
Did the Ocho Rios discussion adopt a regional AI law?
The cited Jamaica Information Service report describes calls for regional cooperation and a shared approach. It does not report a legislative vote or adoption of CAIRMC’s document.
Is the CAIRMC document binding?
No. Its opening notice identifies it as a consultation draft without formal CARICOM or member-state adoption or endorsement. Read its proposals separately from applicable national law and supervisory requirements.
Does UNESCO provide only broad ethical values?
No. Its Recommendation also sets out policy actions, including impact assessment, oversight and due diligence. Different documents should be compared against the particular task rather than dismissed because their formats differ.
Does a framework mapping prove a system is reliable?
No. A mapping can help organise a review. It cannot establish legal compliance, certification, local accuracy or the effectiveness of a control without relevant evidence.
What makes a useful consultation response?
Identify a clause and an operational problem, explain who could be affected, provide appropriate evidence and suggest a workable revision. Use synthetic or properly de-identified examples instead of disclosing confidential information.
What should an institution do before testing an AI proposal?
Define the human need, consider alternatives, check applicable duties and decide what evidence would justify proceeding. Exploratory tests may reveal problems but do not replace validation appropriate to the intended use.
Publisher Disclosure
This article is published by the Caribbean AI Risk Management Council, which issued the draft discussed above. CAIRMC is chaired by Adrian Dunkley, founder of StarApple AI and president of the Caribbean AI Association. The review questions and hypothetical examples are editorial recommendations, not independent validation of CAIRMC’s proposals.