The usual story about the Caribbean and artificial intelligence is that AI is the future, the region is behind, and it has to catch up. In that story the Caribbean is a customer. It waits for tools built elsewhere and hopes they fit.
My claim is a technical one. AI systems built without Caribbean speech, text and data perform worse on Caribbean people, and the gaps they carry make them less reliable for everyone else too. The region has something the industry needs, and it should be at the table when these systems are built.
Models trained on narrow data fail on the people left out
A model learns the patterns in its training data. When it meets inputs that look different, which happens the day it is deployed, its accuracy drops in ways that are hard to predict. Machine learning research has shown for years that more varied training data tends to produce models that generalise better to new cases.
The documented failures follow the same line. Facial analysis systems tested by Joy Buolamwini and Timnit Gebru in the 2018 Gender Shades study had far higher error rates on darker-skinned women than on lighter-skinned men. Dermatology models trained mostly on lighter skin have performed worse on darker skin. Speech recognition systems studied at Stanford in 2020 made roughly twice as many errors for Black American speakers as for white speakers.
None of these is a fairness issue sitting on top of a working product. The product does not work for part of its user base, and that is an engineering defect.
Caribbean language varieties are thin in language model training data
Jamaican Patois has more than three million speakers at home and large communities in the United Kingdom, Canada and the United States. It has its own grammar, sound system and vocabulary. Linguists treat it as a creole language, not as broken English.
Patois text does turn up on the open web, in song lyrics, social media and forums, so large language models have seen some of it. What they have not seen is curated, transcribed, well-labelled Patois at any scale. The result is a model that can imitate a few phrases and then misreads a customer complaint written the way people actually write.
Trinidadian English Creole, Bajan and the Creoles of the smaller islands are in the same position. Haitian Creole is better served, with more than ten million speakers and support in major translation tools for years, and even there quality lags behind the large European languages.
The cost shows up in voice systems first. A Jamaican caller who is misheard three times by an automated phone line hangs up. A visually impaired Trinidadian using dictation software gets garbled text, which means an accessibility tool that does not deliver access. As banks, utilities and government services move to automated channels, these failures decide who gets served.
How to test a model on Caribbean speech before you buy it
Any organisation choosing a voice or chat vendor can run its own check in about a week. Vendor accuracy figures are usually measured on American or British speech and tell you little about your callers.
- Record 50 short clips (15 to 30 seconds) of real customer-style requests from staff or volunteers across the islands and age groups you serve, with written consent.
- Have a person transcribe each clip exactly as spoken.
- Run the clips through each vendor's system and compare its transcript with yours.
- Record 20 clips of the same requests in standard American or British English as a baseline.
- If the error rate on the Caribbean clips is more than double the baseline, do not put that system on a customer line without a fast route to a human agent.
The same approach works for chat: collect 50 real messages written the way your customers write them and check whether the model identified the request correctly.
Who builds Caribbean language data, and what it takes
The gap exists because of choices about which data to collect and which communities to pay for annotation. Those choices are made mostly by companies in North America, Europe and China, whose teams do not speak Caribbean varieties and whose investors see a small market. The return on building a Patois dataset looks marginal from San Francisco.
There are two ways the gap closes. Caribbean institutions can build the data themselves: collect speech and text with consent, pay trained annotators, document it, and publish it under a clear licence where AI developers will find it. Or Caribbean markets can become valuable enough that vendors pay for it. The first route is within our control, and the region has a starting point in the University of the West Indies, whose Jamaican Language Unit has worked on Patois orthography and language rights for decades.
Artful Intelligence: why cultural practitioners belong on AI teams
Artful Intelligence is the name I give to a simple observation. The "A" in STEAM, the arts, is the part most AI teams leave out, and it costs them. AI is an engineering discipline, but its outputs are representations of people, and people are not only numbers.
Writers, musicians, linguists and cultural workers notice when a representation does not match the reality. A poet reviewing a model's Patois output spots a wrong register in seconds. A cultural practitioner writing test prompts catches failures an engineer does not look for, because the engineer is measuring scale and speed.
The Caribbean has spent four centuries combining African, European, Asian and Indigenous influences into new forms: reggae, calypso, soca, dancehall, a literature with two Nobel laureates in Derek Walcott and V. S. Naipaul. That habit of synthesis is a skill an AI team can use when it has to model human variety, and it is the practical case for putting Caribbean artists and linguists on data and evaluation work.
Caribbean seats in AI governance are few
The rules for AI are being written now: what systems are acceptable, how training data may be gathered, what transparency companies owe, and what protection vulnerable groups get. They are being written in standards committees, UN processes and industry forums with few Caribbean participants.
Standards reflect the interests of the people who draft them. The region has practitioners who can contribute, and the barrier is access. International bodies have not built routes in for small states, and Caribbean governments have not yet funded people to sit in those rooms for the years it takes to shape a standard.
Building the evidence in public
The argument that AI needs the Caribbean only persuades if Caribbean practitioners show working systems. I have built AI in Jamaica for fifteen years, and a system that solves a real problem earns more credibility than any conference speech.
That means publishing results, including failures. It means releasing Caribbean datasets where licensing and consent allow, and presenting evaluation numbers at international venues so the teams building foundation models have evidence they cannot wave away. I do not yet have a strong answer to who pays for that work across a region of small economies. Grants cover pilots, and nobody has found a model that keeps a language dataset maintained for a decade.
What to do next
- Heads of customer service at Caribbean banks, utilities and telecoms: run the 50-clip test above on your current voice or chat system this quarter, and publish the error rate internally before renewing the contract.
- Procurement officers buying AI tools: add a clause requiring the vendor to report accuracy on a Caribbean test set you supply, with a remedy if it falls below an agreed threshold.
- University linguistics and computing departments: pick one variety, record and transcribe 10 hours of consented speech, and release it with a data statement describing speakers, dialect region and licence.
- Ministries responsible for ICT: nominate at least one Caribbean expert to an international AI standards process through your national standards bureau, and fund two years of travel and time.
- AI teams in the region: add a linguist or writer from the community you serve to your evaluation step, paid, for every product that handles Caribbean text or speech.
Frequently asked questions
Is there a standard way to write Jamaican Patois for a dataset?
Yes. The Cassidy writing system, developed by the linguist Frederic Cassidy and promoted by the Jamaican Language Unit at the University of the West Indies, Mona, spells words consistently by sound. Most Patois on social media uses informal English-based spelling instead, so a good dataset keeps the original text and adds a normalised version. Record which convention you used in the dataset's documentation.
How is speech recognition accuracy measured?
The standard measure is word error rate: the number of substituted, deleted and inserted words divided by the number of words in a human transcript. A rate of 10% means roughly one word in ten is wrong. Always compare rates on the same kind of clip, because short commands and long conversations produce different rates.
Do I need consent to record people for an AI training dataset in Jamaica?
Yes, in practice. A voice recording of an identifiable person is personal data under Jamaica's Data Protection Act 2020, which came into force in December 2023. Get written consent that names the purpose, including sharing with third parties or publishing, and give people a way to withdraw. The Office of the Information Commissioner publishes guidance on the Act.
How can a Caribbean institution share a language dataset without losing control of it?
Choose a licence before release. Creative Commons licences are the usual choice: CC BY allows any use with credit, while CC BY-NC blocks commercial use without a separate agreement. Some community groups use a request-access model, where developers apply and agree to terms before download. Whatever you choose, publish a data statement so users know who is represented and who is not.
Where are international AI standards written, and how can a Caribbean expert take part?
The main international committee is ISO/IEC JTC 1/SC 42, which writes AI standards such as ISO/IEC 42001 on AI management systems. Individuals join through their national standards body, in Jamaica the Bureau of Standards Jamaica, which can register as a participating or observing member. UNESCO's Recommendation on the Ethics of AI, adopted in 2021, also runs readiness assessments that governments can request.