Research Ethics and Human Participants

Machine learning research is rarely done on abstractions. It is done on hospital records, on credit files, on satellite imagery of people's homes, on speech recordings, on the movement of individuals through a small country where almost nobody is genuinely anonymous. This policy sets out what AIC requires of authors whose work touches people, and what we require when it touches communities rather than individuals.

It draws on the Declaration of Helsinki, the Belmont Report, the CIOMS International Ethical Guidelines and the CARE Principles for Indigenous Data Governance, and it adapts them to the specific problem of doing quantitative research in populations small enough that statistical disclosure is a live risk rather than a theoretical one.

1. When This Policy Applies

It applies to every submission on every route — peer-reviewed articles, preprints and student papers — where the research involves any of the following:

• Data collected directly from people: surveys, interviews, experiments, observation, user studies
• Secondary use of data about identifiable people, including administrative, clinical, financial or educational records
• Data that is individually identifiable, or that could be re-identified by combining it with something else
• Human tissue, genetic data, or biometric data of any kind
• Data about a community, a settlement or an indigenous population, whether or not individuals are identifiable
• Deployment of a system that makes or informs decisions about people
• Scraped or repurposed content created by identifiable individuals

It does not apply to work using only aggregate published statistics, physical or environmental measurements with no human subject, simulation, or purely theoretical and methodological research. If you are unsure which side of the line your work falls on, write to ethics@caribbeanaijournal.org before you submit. That question has never delayed anyone's submission, and getting it wrong has.

2. Ethical Approval

Research involving human participants requires approval from a properly constituted research ethics committee or institutional review board, obtained before the data was collected. The submission must state the approving body, the protocol number and the approval date, and we may ask to see the approval letter.

If you have no institutional ethics committee

Many of the researchers this journal exists for — independent researchers, practitioners in small agencies, secondary school students, staff at institutions that have never constituted a board — have no committee to go to. Refusing their work outright would exclude the people the journal was built for, so instead:

1. Say so plainly in the ethics statement, rather than leaving the field empty or implying an approval you did not obtain.
2. Set out what you did instead: how participants were recruited, what they were told, how consent was recorded, what data you held, how it was secured, and what you did to prevent identification.
3. The handling editor applies a proportionate review against the principles in section 3, drawing in a second editor where the risk is not obvious.
4. The published paper carries the same statement, so readers can judge for themselves. Transparency about the absence of a committee is acceptable; silence about it is not.

Absence of an ethics committee is never a reason to skip consent, to skip disclosure control, or to take a risk with participants that a committee would have stopped. It changes who does the review, not whether one happens.

3. The Principles We Apply

4. Informed Consent

Consent means the participant understood what they were agreeing to. A signature on a form nobody read is not consent, and neither is a checkbox in a terms of service document.

• Informed: purpose, what participation involves, what data is collected, who sees it, how long it is kept, and what happens to it afterwards.
• Freely given: no coercion, and particular care where the researcher holds power over the participant — a teacher over a student, a clinician over a patient, an employer over a worker.
• Specific: consent to one study is not consent to every future use of the data. Broad consent to unspecified future research must have been explicitly agreed and must be disclosed in the paper.
• Recorded: in writing, or by an auditable process where literacy or language makes writing inappropriate. Say which in the paper.
• Withdrawable: participants could stop and have their data removed, and the paper says what the deadline for that was.
• Comprehensible: in a language and register the participant actually uses, including Creole and other vernaculars. A consent form in academic English handed to someone who speaks Jamaican Patois is a compliance artefact, not consent.

Participants under 18

Consent comes from a parent or legal guardian, and assent from the young person themselves, in language they understand. Both are required and one does not substitute for the other. A young person who declines is not enrolled, whatever their guardian has signed. Where the authors are themselves under 18, the safeguarding policy applies in addition.

Waiver of consent

Consent may be waived only where a research ethics committee has approved the waiver, the research could not practicably be done otherwise, the risk to participants is minimal, and the waiver is disclosed in the paper. Secondary analysis of an existing de-identified dataset under a licence permitting research use does not require fresh consent, but the paper must name the dataset, its licence and its original consent terms.

5. Identification and Disclosure Control

This is the section most likely to result in a request for revision at this journal, and it is worth reading before you write the paper rather than after.

• No direct identifiers in the paper, the figures, the tables, the appendices or the supplementary files: names, addresses, national identification numbers, patient numbers, exact dates of birth, telephone numbers, email addresses, vehicle registrations, facial images.
• Quasi-identifiers are the real risk. The combination of parish, age, occupation and one clinical fact can identify a single person in a Caribbean population. Authors must assess re-identification risk against the actual population size, and say how they did it.
• Small cell suppression. Cells with counts below five are suppressed or aggregated, with a lower threshold used only where an ethics committee has approved it. This applies to figures as well as tables: a scatter plot with eleven points identifies eleven people.
• Geographic granularity is reported at the coarsest level the analysis supports. Parish rather than community, community rather than street, and never a coordinate pair for a household.
• Free text and qualitative quotation is redacted of identifying detail. In a small community a quotation can identify its speaker by phrasing alone, and the paper should say how that was handled.
• Photographs of people are published only with written consent from every identifiable person, and never where the participant is under 18.
• Data availability does not override this. Where the underlying data cannot be shared without disclosure risk, say so and deposit what can be shared: a synthetic dataset, aggregate tables, or the analysis code with a controlled-access route to the data.

6. Research on Deployed Systems

Papers reporting a system in live use carry obligations beyond those on a retrospective analysis, because real decisions were made about real people while the research was going on.

• State whether the affected people knew. Whether individuals subject to the system's decisions were informed, and what recourse they had.
• Report the failure modes, not just the headline metric. False positives and false negatives have names attached in a deployment, and a paper that reports only aggregate accuracy is incomplete.
• Report subgroup performance wherever sample size permits, and say so explicitly where it does not. A model that works on average and fails on one parish is a finding, not a footnote.
• Describe the human oversight actually in place, not the oversight the design assumed.
• Describe the harm that would follow a wrong output, and what was done to bound it.
• A/B tests and field experiments on live users require ethical approval and must say what the control condition cost the people in it.

7. Community and Indigenous Data

Research can harm a community without identifying a single individual in it. A published finding about a settlement, a fishing community, an indigenous population or a parish can affect insurance rates, investment, policing and reputation for people who never consented to being studied.

Where research concerns a identifiable community, authors must state how that community was engaged, who spoke for it, and what was agreed about how findings would be published. For research involving indigenous peoples — including Kalinago, Garifuna, Lokono, Maroon and Taíno communities — the CARE Principles apply: Collective benefit, Authority to control, Responsibility, and Ethics. Data about a people belongs in a governance relationship with that people, and a licence from a government agency is not a substitute for it.

8. Animal Research

Research involving animals requires approval from an institutional animal care and use committee or its national equivalent, and the paper must report compliance with the ARRIVE 2.0 guidelines and describe how the three Rs — replacement, reduction, refinement — were applied. This arises rarely at this journal, most often in fisheries, agriculture and wildlife monitoring work, and the requirement is the same as at any life sciences journal.

9. The Ethics Statement

Every submission touching people carries an ethics statement, published with the article. It is a short section, and it needs to cover eight things.

10. What Happens If a Submission Falls Short

Most shortfalls are omissions rather than misconduct, and are handled as such.

1. Missing statement or unclear detail: we ask you for it. This is the outcome in the large majority of cases and delays a submission by days rather than ending it.
2. Identification risk in the manuscript: revision required before review continues. We tell you which table, figure or quotation is the problem and what would fix it.
3. Consent absent where it was needed: the submission is declined. We say why, and we say whether a different analysis of the same data could be submitted.
4. Approval claimed but not held: this is misconduct, not an omission, and it is handled under the publication ethics policy, which includes contacting the institution named.
5. A concern raised after publication triggers the corrections, retraction and expression of concern process in the editorial policies. Ethics concerns about published work go to ethics@caribbeanaijournal.org and are investigated whoever raises them, including anonymously.

11. Review of This Policy

Reviewed annually by the editorial board. Changes are published here and recorded in the document register.