Use the Caribbean AI Risk Index to ask better risk questions

Research status: Exploratory use only. Independent reconstruction and predictive validation remain unresolved.

Hero: AI-generated conceptual image prepared by AI Boss Limited. The compass and five geometric forms illustrate questions and categories, not measured results.

The Caribbean AI Risk Index (CARI) is a published exploratory comparison of AI exposure and institutional preparedness. Source inspection identified presentation defects, now corrected on the production site, and unresolved provenance gaps. Risk teams can use its categories to organise questions while keeping those limits visible. A country score should not decide whether an organisation approves a system, selects a supplier or accepts a risk. Those decisions still need evidence about the particular service and the people it affects.

What the work brings together

The published CARI dataset, labelled version 2.1 and dated 6 October 2026, contains profiles for 16 sovereign states and estimates for 15 territories. The territories sit outside the sovereign ranking. The project credits the Caribbean AI Association and Maestro AI Labs, with Adrian Dunkley as its lead.

Its structure separates exposure from institutional fragility. Exposure covers labour-market tasks, cyber and AI-agent threats, and information integrity. Fragility covers gaps in governance and economic resilience. Psychosocial material is contextual rather than a sixth scored pillar. This organisation gives a team a way to examine several channels instead of treating AI risk as a single technology question.

View the AI Risk Index. Treat its comparisons as exploratory and read the limitations before quoting a result; full calculation reconstruction remains to be verified.

A displayed score needs an evidence trail

The public display files store precomputed outputs. A static inspection of the repository files for this article confirmed the dataset counts and the distinction between the five scored pillars and contextual material. This check did not reconstruct the scores or test whether they predict future losses.

The initial review found reversed information-integrity and governance headers, an extra psychosocial header and provisional-status wording that did not distinguish missing essential inputs from low overall coverage. The production page and display code now contain the corrections. Repairing those presentation defects does not reconstruct the underlying calculations.

The inspected files do not establish the complete chain from original inputs to published scores. That requires the source register, permitted raw indicators, transformations and generating code. A public dashboard alone does not establish independent validation.

Before quoting a country comparison, check the edition, the relevant pillar, its source period and any coverage or provisional flag. If the original input and transformation cannot be traced, record that limitation rather than presenting the score as settled evidence.

From context to a useful check

1

Regional question

Choose the relevant pillar and edition.

2

Local evidence

Name the service, affected people and permissions.

3

Decision owner

Review tests and unresolved uncertainty.

4

Decision

Proceed, restrict, defer, decline or seek specialist review.

Turn a regional signal into a local check

Consider a hypothetical customer-service team introducing an AI assistant. Regional cyber exposure can prompt a useful question: what can this assistant access or change? The answer must come from the organisation's permissions, tests and supplier evidence, not its country's index position.

A practical sequence is:

1. Select the relevant index pillar and record the edition and limitations.
2. Identify the affected service, people and decision owner.
3. Gather service-level evidence, such as access permissions, test results and incident records.
4. Decide whether to proceed, restrict the proposed use or seek specialist review. Record the remaining uncertainty and a review trigger.

This is a proposed use of the index as a starting point for investigation. It is not evidence that the sequence has reduced losses.

Questions readers should ask

Does a score express the probability of an incident? No. The published scores are relative comparisons within the regional sample. They are not calibrated probabilities of harm.

Does a lower score establish that a business is safe? No. A national composite does not assess an individual organisation or system.

Can the public dashboard reproduce the calculation? The inspected display files store outputs. Full reconstruction requires the raw permitted inputs, source register, transformations and score-generation code. This source inspection did not establish that complete chain.

What should a risk team do next? Choose one relevant service and write down the evidence needed for its approval. Use the index to broaden the questions, then make the decision from evidence about that service.

Source note

Source: Maestro AI Labs × Caribbean AI Association, CARI release 2.1 dated 6 October 2026. The correction commit was checked against the production page and display code on 30 October 2026 at 03:08 UTC. Both matched the corrected files; the dataset and quarterly history hashes were unchanged. This verifies deployed file contents, not a rendered browser test. It does not reconstruct the scores or verify each national input.

Correction status

Corrections to the table mapping, provisional-status explanation and wording about the method were verified on the production site on 30 October. The numerical dataset and quarterly records are unchanged. These repairs do not establish independent reconstruction or predictive validity. The original input matrix and calculation code still need to be recovered and checked.