International Women's Day 2026: Why AI Built Without Women Works Worse for the Caribbean

The shortage of women in AI is an equity issue, and it is also an engineering defect. A system built and tested without the people it will serve makes more mistakes about them. For the Caribbean, where many AI tools arrive already trained on other people's data, that defect lands hardest on Caribbean women.

I lead with the engineering case because it is harder to wave away than a representation statistic. Discrimination in technology careers is real and documented; this piece is about what it does to the systems themselves.

How missing data and missing reviewers produce worse AI

A model learns patterns from its training data and applies them to new cases. If a group is thin in the data, the model is less accurate for that group, and the errors often go unnoticed unless someone on the team looks for them. Four well-documented cases show the pattern:

• Hiring. Amazon trained a CV-screening model on ten years of its own hiring, which had favoured men. Reuters reported in 2018 that the model penalised CVs containing the word "women's", as in "women's chess club captain", and downgraded graduates of two all-women's colleges. Amazon dropped the tool.
• Medicine. Clinical research has long under-enrolled women, and heart disease often presents differently in women. Diagnostic models trained on that research can carry the gap into every hospital that uses them.
• Faces. The 2018 Gender Shades study by Joy Buolamwini and Timnit Gebru found commercial gender classification systems misclassified darker-skinned women up to 34.7% of the time, against under 1% for lighter-skinned men.
• Language. Translating gender-neutral Turkish into English, Google Translate used to render a doctor as "he" and a nurse as "she". Google changed the product in 2018 to show both versions.

Each of these failures was preventable. Each one is more likely when nobody on the team is positioned to notice it.

Why Caribbean women are exposed twice

Caribbean users already get worse results from tools trained elsewhere. Speech and text tools mishandle Patois and other Caribbean creoles because there is little of them in training data. Face systems perform worse on darker skin. Credit models built on formal-sector borrowers in richer countries misread households that earn in the informal economy.

Caribbean women sit at the intersection. Women are under-represented in AI development worldwide, and Caribbean people are under-represented in global AI. A tool built without either group, then imported into Kingston or Georgetown, is least likely to have been tested on a Caribbean woman's face, voice, medical history or income pattern.

What StarApple AI has got wrong

Since I founded StarApple AI in 2016 I have run bootcamps, assessments and deployments across the region. Women come to the training, do the work well and ask the questions that catch failure modes. Yet the training and networking channels in Caribbean AI have served them poorly, and that includes ours.

Our programmes have not always recruited women actively, made them feel they belonged, or prioritised the applications most relevant to their working lives. We are changing that, starting with the first target in the list below.

Where local knowledge changes the model

My idea of Artful Intelligence is that useful AI combines technical skill with knowledge of the community it serves. That knowledge is concrete:

• A health tool designed with Guyanese women nurses and community health workers is more likely to reflect local chronic disease patterns and how women actually use clinics.
• A credit model reviewed by Jamaican women who run market stalls or partner schemes is more likely to count income a bank form ignores.
• A language tool tested by speakers across Trinidad and Tobago is more likely to handle Trinidadian English and creole.

My argument has a weak point, and it should be stated. Adding women to a team does not by itself fix a model; a team of any make-up can miss a bias if it never measures accuracy by group. Representation raises the odds that someone asks the question. Disaggregated testing is what answers it.

An AI system that has never been tested on Caribbean women will make mistakes about Caribbean women, and nobody will know until the mistakes pile up.

What to do next

1. AI training providers, including StarApple AI: set a target share of women in each cohort, publish it, and pair recruitment with scholarships and flexible formats such as evening and online sessions.
2. Any organisation buying an AI tool: before deployment, ask the vendor for accuracy figures broken down by gender and skin tone on data like your own users, and test it yourself on a small local sample if they cannot supply them.
3. Banks and credit unions using scoring models: compare approval and default rates for women and men every quarter, and investigate any gap the default data does not explain.
4. Governments drafting national AI strategies: include women practitioners and women's organisations on the drafting committee, and require bias testing by gender for AI used in public services.
5. Conference organisers and media in the region: make sure women doing AI work in the Caribbean appear as speakers and sources, so girls choosing subjects can see them.

Frequently Asked Questions