5 Tips for Avoiding the AI Apocalypse: A Caribbean Community Guide

A community group uses AI to draft a notice. A small business asks it to explain a supplier contract. A professional association tries it for member enquiries. Each is an opportunity to save time, and each needs someone to ask what could go wrong before the answer reaches another person.

In this guide, “AI apocalypse” is a metaphor for workplace disruption and avoidable failures: private information escaping, confident mistakes spreading, or an automated action going further than anyone intended. These five habits give Caribbean professionals and community groups a practical starting point.

The NIST AI Risk Management Framework organises risk work around Govern, Map, Measure and Manage. It treats that work as continuous. The tips below are a small learning exercise inspired by that approach; completing them does not establish compliance or make a system risk-free.

1. Check the vendor, the model and the data

Before adopting a tool, identify its provider, the model or service behind it, and the information it will receive. Ask about retention, access, updates and available evidence of performance. If the supplier cannot explain a limitation, record that uncertainty. Supply-chain checks and documentation are part of the NCSC’s secure AI development guidance.

For a Caribbean learning group, try a fictional enquiry that includes local place names, currencies and everyday language. Ask a colleague familiar with the context to review it. A polished demonstration using a different country’s assumptions tells you little about your own task.

2. Keep private information out of practice prompts

Use fictional names and records when learning together. Avoid uploading member lists, customer documents or confidential correspondence just to try a feature. For real work, confirm that the tool and data use are approved, share only what is needed, and check the provider’s current settings and terms.

The NIST Generative AI Profile identifies privacy risks including unauthorised disclosure and sensitive inferences. Removing a name alone may leave enough detail to identify someone. A useful group exercise is to rewrite an enquiry as a fully fictional scenario before anyone enters it into a tool.

3. Verify important answers at the original source

NIST’s Generative AI Profile also describes confidently incorrect outputs, often called hallucinations, and fabricated citations. Open the source, check the date and confirm that it supports the claim. Recalculate important figures. Another chatbot repeating the answer is not independent verification.

Before sharing an AI-written notice, ask someone to check its venue, contact details and deadlines against the organiser’s original information. If a claim affects health, legal rights or money, get the appropriate qualified reviewer involved. Leave unresolved statements out of the final message.

4. Give AI limited access and people real authority

A tool that drafts a message does not need permission to send it. A tool that summarises records does not need permission to delete them. OWASP’s guidance on excessive agency recommends limiting functionality and permissions, enforcing authorisation in connected systems, and requiring human approval for high-impact actions.

Name the person who reviews a consequential action and give them the information and time to reject it. For a community association, that might mean a committee member checks the audience and final text before a notice is distributed. Start with a draft-only workflow while people learn its weaknesses.

5. Practise stopping, reporting and recovering

Keep a simple record of errors and near misses. Decide who can pause the workflow, who investigates and how people obtain help. After a model, data source or process changes, repeat the checks that matter. Logging, monitoring and incident management feature in the NCSC’s lifecycle guidance.

Try a fictional rehearsal: an AI-assisted notice has the wrong location. Who stops further messages? Who checks the correct information? Who approves a correction? Agree on a manual fallback before the tool becomes essential. Keep incident records access-controlled and avoid copying private information into a shared learning chat.

Turn the tips into a shared learning session

Bring one fictional task to your next professional or community meeting. Let one person operate the tool, another check the answer and a third review what information and permissions it needed. End with one decision: continue with safeguards, change the workflow, or stop. Share the lesson without exposing anyone’s data.

Explore Survival Guide to the AI Apocalypse: A Practical Guide to AI Risk Management and enquire with CAIA about course arrangements. For reading alongside that conversation, see my book, Survival Guide to the AI Apocalypse, on my public books page.

Find more of my work at Adrian Dunkley’s website, explore organisational AI work through StarApple AI, and connect with peers through CAIA membership. If you want to start studying today, CAIA’s separate free Introduction to AI Risk Management is available online.