AI and Research Integrity

This is a journal about artificial intelligence, which puts it in an awkward and interesting position: the thing it studies is also the thing most likely to corrupt its own record. Every scholarly publisher is now handling manuscripts written by language models and submitted as original research. A journal in this field cannot treat that as somebody else's problem.

This document covers three distinct things that are often muddled together: how authors may use AI tools in producing a manuscript, what a paper about an AI system must report, and what editors and reviewers may do with AI tools in editorial work. The rules differ sharply across the three.

1. AI Cannot Be an Author

Authorship carries accountability, and accountability requires someone who can be asked a question and give an answer. A language model cannot approve a manuscript, cannot declare a competing interest, cannot respond to a post-publication challenge, and cannot be held responsible for a fabricated citation.

• No AI system, tool or model is listed as an author or co-author. Submissions doing so are returned without review.
• Every named author is a person who meets the authorship criteria in the editorial policies and can defend the work.
• AI tools are credited in the methods and the AI use statement, the way a statistical package or an instrument is credited. Naming the tool is required; naming it as an author is not permitted.

2. What Authors May and May Not Do

We are not hostile to these tools. Refusing to use them in AI research would be a strange position for this journal to hold, and a futile one. The line is not whether you used them but whether a human verified what came out and whether you said so.

Synthetic data is a separate matter and is permitted where it is the object of study or a documented part of the method. It must be labelled as synthetic at every point, its generation process must be described, and it must never be presented as observation.

The AI use statement

Every submission carries one, published with the article. It names the tool, the version, and the task. "ChatGPT was used" is not a statement; it is an evasion. This is:

Write "no generative AI tools were used" where that is true. Disclosure has never cost anyone a publication at this journal. Concealment ends one, and the detection section explains why concealment is a worse bet than it looks.

3. Reporting Requirements for AI Research

A paper about a model is not reproducible because you named the architecture. These requirements apply to every submission that trains, fine-tunes, evaluates or deploys a machine learning system, and the handling editor checks them before review.

Data

• Provenance: where every dataset came from, who collected it, when, under what licence, and with what consent. A dataset with no stated provenance is a finding you cannot check.
• Composition: size, the split between training, validation and test, class balance, and the demographic or geographic composition where the work concerns people or places.
• Preprocessing: cleaning, exclusion criteria, imputation and augmentation, stated in enough detail to be repeated.
• Known limitations: coverage gaps, sampling bias, label noise, and what the dataset cannot support a claim about. Caribbean datasets are frequently small, unevenly collected and interrupted by a hurricane season; say so rather than hoping nobody notices.
• Availability: a repository, a DOI and a licence, or a stated reason the data cannot be released, with what can be released instead.

Models

• Architecture and size, including the base model and version for anything fine-tuned
• Training procedure: objective, optimiser, hyperparameters, schedule, stopping criterion, and how hyperparameters were selected
• Compute: hardware, wall-clock training time, and an estimate of energy or carbon where the scale makes it material. In a region where compute is expensive and electricity more so, what a result cost to produce is part of whether it is useful.
• Random seeds and variance: results from a single run are reported as such. Where multiple runs are feasible, report the mean and the spread. A single-seed result claiming a point-two improvement is not a result.
• A model card for any system proposed for real use, covering intended use, out-of-scope use, performance across subgroups, and known failure modes.

Evaluation

• Baselines that are real. A comparison against an undertuned baseline is not a comparison. Say how the baselines were tuned.
• Metrics chosen before the results were seen, with the choice justified. Accuracy on an imbalanced Caribbean health dataset is close to meaningless and we will say so.
• Uncertainty: confidence intervals or standard errors on every headline number.
• Subgroup performance wherever sample size permits, with explicit acknowledgement where it does not.
• Test set hygiene: state that the test set was held out, and address contamination directly for any work using a pretrained model on public benchmarks.
• Negative and null results are welcome here. A well-designed study showing that a method which works elsewhere fails on Caribbean data is exactly the kind of paper this journal was founded to publish, and it will not be rejected for the direction of its finding.

Transfer and external validity

A model trained in one context and applied in another must report what changed and what broke. Where a paper claims a system transfers to Caribbean conditions, it must show evidence from Caribbean data rather than an argument that it ought to. This is the single most common reason a methodologically sound manuscript is returned for revision at this journal.

4. Dual Use and Foreseeable Harm

Some research enables harm as readily as benefit. Submissions with a plausible harmful application must include a short impact statement covering the foreseeable misuse, who would be affected, and what the authors did to reduce the risk.

Areas that routinely trigger this at a Caribbean journal: surveillance and biometric identification; predictive policing and border systems; credit and insurance scoring where exclusion has a documented history; systems that infer immigration status, sexual orientation or political affiliation; and models trained on data from populations small enough that the model itself leaks its training data.

The editorial board may decline work whose foreseeable harm outweighs its contribution, or may require that a component be withheld. That is a rare decision, it is taken by the board rather than a single editor, and the reasons are given to the authors in writing.

5. AI in Editorial Work

What editors and reviewers may do:

• Not permitted: entering any part of an unpublished manuscript, a review report, or an author's personal data into a system that retains or trains on inputs.
• Not permitted: generating a review report, or any part of one, with an AI tool. A review is the reviewer's own judgement, and a generated one is worse than no review at all because it looks like one.
• Not permitted: any automated editorial decision. Desk rejections, review outcomes, moderation and appointments are made by a named human being who is accountable for them. This is also a data protection commitment.
• Permitted: language editing of the reviewer's own already-written text, using a tool with no retention.
• Permitted: similarity screening, reference checking and image forensics through services under a written processor agreement. These flag passages for a human to read; they reject nothing.
• Required: a signed confidentiality undertaking covering all of the above, before an editor or reviewer handles their first manuscript.

Where any tool has been used in handling a manuscript, it is disclosed to the authors on request. There is nothing here we would be uncomfortable telling an author about.

6. Detection, and Why We Do Not Rely On It

AI text detectors do not work reliably. They produce false positives at rates that make them unusable as evidence, and they are documented to be systematically worse on text written by authors whose first language is not English. At a Caribbean journal, using one as evidence would mean accusing exactly the authors this journal exists to serve.

We therefore do not use AI detection scores as evidence of misconduct, and we will not act on one. What we do instead:

• Verify citations. Fabricated references are the most reliable signal in practice, they are trivial to check, and a reviewer finds them in minutes.
• Ask the authors. An author who did the work can explain a choice made in their methods. One who did not, cannot, and the conversation resolves quickly.
• Check the data and code. Results that cannot be reproduced from the materials provided are the concern, whatever produced the prose around them.
• Rely on the declaration. Every submitting author declares that a named human stands behind the analysis and results. A false declaration is misconduct in itself, independent of how the text was produced.

If an editor raises a concern, the author is told what the concern is and given a real opportunity to answer it before any decision is taken. Nobody here is refused on a detector's say-so.

7. The Preprint Server

The moderation criteria on the preprint server include a disclosure check, and a manuscript whose argument, analysis or results were produced by a model that no author has verified is refused. Moderators are looking for whether a human stands behind the claims, not for a detector score. The registration checks exist for the same reason: an unauthenticated server drowns in machine-generated submissions, and a contributor account is the cheapest defence that does not also exclude legitimate researchers.

8. Breaches of This Policy

9. Review of This Policy

This policy is reviewed every six months rather than annually, because the technology and the practice around it change faster than an annual cycle can track. Next scheduled review June 2027, with an interim review in December 2026. Changes are published here and recorded in the document register.