Badejo: Government AI Policy would put security, human oversight and risk controls first

GREAT BAY--St. Maarten’s proposed Government artificial intelligence policy would place security, human accountability and strict risk controls at the center of how AI is introduced across the public service, with high-risk uses requiring formal review and sensitive Government information barred from unauthorized public AI systems.

Head of ICT Femi Badejo told Parliament that AI is already being used within Government for productivity purposes and is expected to play a larger role in areas such as drafting, summarizing information, locating policies, routing applications and helping residents find information. He said the challenge is no longer simply whether Government should use AI, but how it can do so safely, responsibly and with clear rules that protect the public.

Badejo said the draft policy has been designed around international practices already being applied elsewhere, including principles reflected in European approaches to artificial intelligence. Rather than creating an entirely new model, Government is attempting to adapt proven safeguards to the St. Maarten public-service environment.

Under the proposal, Government employees would only be permitted to use AI tools approved by the ICT Department for official work. ICT would maintain a register identifying which tools are approved, what they may be used for and what information can or cannot be entered into them.

Badejo identified security as one of the most important parts of the framework. Personal information, Cabinet documents, official advice, confidential information and security-related material would not be allowed to be placed into public AI systems without the necessary authorization and safeguards. The intention is to prevent sensitive Government and personal data from being exposed through inappropriate AI use.

Government is currently considering Microsoft Copilot Chat for approved official use because it operates within Government’s existing Microsoft enterprise environment. Badejo stressed, however, that approval of a particular AI tool would not mean employees could use it without limits. The type of information being processed and the purpose for which AI is being used would still determine whether additional approval is required.

The draft policy also makes clear that AI cannot replace the responsibility of public officials. Employees would be required to review AI-generated material, verify its accuracy and reject recommendations where necessary. Badejo said an authorized official must remain accountable for decisions that affect residents, particularly where rights, services, permits, benefits, education or enforcement are involved.

If an AI system produces harmful answers, exposes information or behaves in an unexpected way, staff would be required to immediately report the incident to ICT. The department would then be responsible for investigating the matter and could restrict or completely disable access to the AI system if the risk could not be controlled.

Government is proposing a risk-based system for determining how AI may be used. Low-risk use could include drafting general text, generating ideas or summarizing non-sensitive material. These activities would still require staff to check the results but would not normally require the same level of approval as more consequential applications.

A second category would cover AI use under supervision, including information assistance, routing applications to the correct Government department, certain translations and similar functions where the risk can be managed through staff oversight.

Higher-risk applications would receive significantly greater scrutiny. Badejo identified uses involving eligibility for Government benefits, permits, education and enforcement decisions as examples where AI could have direct consequences for an individual. These applications would require review by the Enterprise Architecture Review Board before implementation.

Where a proposed AI use presents risks or exceptions beyond the authority of the review board, the matter could be referred to the Council of Ministers for a decision.

Some uses would be prohibited outright. Badejo said social scoring would not be permitted, nor would Government allow AI to make decisions affecting a person’s rights or access to services without meaningful human review.

The policy would also establish clear expectations for residents interacting with Government services supported by artificial intelligence. People should be informed when AI is being used, understand who remains responsible for the service and have access to a Government employee if they require assistance or wish to question an outcome.

Residents would also be able to request human review. Badejo said this is particularly important because AI can provide incorrect information in a convincing manner, meaning Government cannot simply accept an AI-generated answer as accurate without verification.

The proposed framework would also require Government to improve how it classifies information. Data would need to be categorized according to whether it is public, private, confidential or highly sensitive so the appropriate technical controls can determine what information an AI system is permitted to access.

Badejo said Government intends to introduce AI gradually rather than attempting a large-scale rollout at once. The approach would be to begin with supervised pilot projects, test whether they work, measure results and only then consider wider use.

One proposed pilot could involve an AI information assistant that answers common questions using only approved public Government information. Residents could potentially ask about requirements for a particular service, how long a process may take or which department handles a certain matter. More complicated questions would still be referred to Government staff.

Another possible pilot would help Government employees search for approved policies and information they are authorized to access, while safeguards would prevent access to confidential or restricted material.

AI could also be tested for application routing. Where a resident submits a request, the system could help identify which team should receive it, reducing the amount of time spent manually redirecting applications.

Badejo also pointed to Government’s planned grievance and feedback mechanism under the Digital Government Transformation Project. AI could potentially review a complaint or piece of feedback and suggest which Government department should handle it, particularly in cases where a resident may not know where the matter belongs. Staff would still review and confirm the AI recommendation.

Every AI pilot would require a responsible ministry owner, ICT leadership, a timetable for implementation and clear performance indicators. Badejo said Government must know what a pilot is expected to achieve and be able to determine whether those expectations were met.

If a pilot proves unsafe or unreliable, Government would have the ability to stop it immediately. Approval to conduct a pilot would also not automatically authorize permanent or wider use. A separate approval would be required before expanding the technology.

Government intends to use existing structures to oversee the policy rather than create unnecessary additional bureaucracy. The Enterprise Architecture Review Board would be given responsibility for reviewing and approving higher-risk AI applications, while ICT would oversee the technology, approved-tools register, security controls and implementation standards.

Legal and privacy specialists would also be involved where necessary to assess whether proposed applications comply with legal requirements, privacy protections, fairness standards and residents’ rights. Individual ministries would remain responsible for the services in which AI is used and for ensuring that meaningful human review remains available.

Training will be another major requirement. Government employees would need guidance on which AI tools are approved, what information may be entered, how AI-generated results should be verified and how incidents should be reported. Badejo said proper training is necessary because having an approved tool does not by itself guarantee responsible use.

The education sector is also expected to develop more specific rules around artificial intelligence. While AI can assist teachers with lesson preparation, support students with different learning needs and reduce administrative work, Badejo said educators must remain responsible for consequential decisions.

The education framework would also need to address student privacy, disclosure of AI use in schoolwork, safety and responsible use. Students, he said, should learn how to recognize AI errors and should not simply accept generated answers without questioning them.

Government is proposing that the policy take effect within approximately 60 days after formal approval by the Council of Ministers. Badejo stressed that this does not mean every proposed AI application would immediately be introduced. Implementation would be phased, with trials, risk assessments, approvals and staff training taking place progressively.

The financial impact will also have to be considered before AI is expanded throughout Government. Badejo told Parliament that artificial intelligence can be expensive and that Government must consider licensing, implementation, training, technical support and total long-term ownership costs before deciding whether an AI project should proceed.

The Council of Ministers would ultimately be asked to approve the policy, give the existing review board authority over AI approvals, authorize the maintenance of an official register of approved tools and uses, require formal approval for higher-risk applications and allow supervised trials to begin.

The overall objective, Badejo said, is to take advantage of AI’s potential to reduce paperwork, improve access to information, provide faster responses and help civil servants work more efficiently, while ensuring that security, privacy and accountability are not sacrificed in the process.

The proposed system would keep people at the center of Government decision-making. AI may assist with information, recommendations and routine work, but officials would remain responsible for checking its output and for every consequential decision affecting the public.

‍