Mauritius is turning campus AI rules into an operating manual

THE SIGNAL IN ONE SENTENCE

Most campus AI policy begins with a panicked email. Do not cheat. Cite the bot. Protect private data. Ask your lecturer if you are confused. Then the university discovers that everyone is confused, including the lecturer. Mauritius is trying a more deliberate route. It has national regulations for AI across higher education, detailed guidelines for institutions and now a consultation intended to turn those duties into a National AI Governance Framework and operational instructions. The Higher Education Commission convened the national workshop in Ebène on October 1. The Government Information Service reported the event the following day and said recommendations from participating stakeholders will directly inform the framework and its accompanying guidelines. The plain signal is that Mauritius has reached the difficult part of AI governance. The country has already written down what institutions owe students, staff and the public. Now it has to explain who does what on Monday morning when a student uses a model in an assignment, a lecturer redesigns an exam, a researcher uploads sensitive material or an administrative system makes a recommendation about a person. That distinction matters because the legal text is more substantial than a voluntary principles list. The Higher Education (Use of Artificial Intelligence) Regulations 2026 apply across the Mauritian higher education sector. They authorize the Higher Education Commission to issue standards and directives, prohibit uses that can cause bias or breach academic integrity, require periodic policy review and demand corrective measures within a prescribed time when an institution falls out of compliance. The rules also give every higher education institution a long homework assignment. Each institution must develop its own AI policy covering ethical use, academic integrity and governance in teaching, learning, assessment, grading and research. The policy must be updated as technology and law change. Institutions must promote equitable access to AI literacy, tools, services, research and faculty development regardless of socioeconomic status, location or background. They must teach both technical and ethical dimensions. They must require disclosure of AI use in coursework, assignments and research, including explicit acknowledgement of generated content and data. They must establish intellectual-property guidance, monitor AI use across academic and administrative work, protect privacy and security, mitigate harm and respect human rights and dignity. If misuse or noncompliance appears, corrective action is not optional. That is a serious skeleton. It still needs joints. Take disclosure. A requirement to acknowledge AI sounds simple until a student asks what counts. Spell-checking? Translation? Brainstorming? A model that summarizes the student's own notes? A research assistant that finds papers? A chatbot that drafts half the analysis? A coding model that writes a function the student then tests and explains? An operational guide has to draw useful boundaries without pretending every course needs the same rule. The sensible unit is the learning objective. If the point of an assignment is to practice drafting an argument, outsourcing the first draft may erase the work being assessed. If the point is to critique weak reasoning, a generated draft may be legitimate material. If a programming course is assessing system design, code completion may be allowed while unexamined generated logic is not. If a language course is measuring unaided writing, translation assistance changes the task. One national disclosure form cannot resolve every case. A common vocabulary can. Mauritius could require each assessment to declare an AI condition before students begin: prohibited, limited, permitted with disclosure or required. The course should name allowed functions, evidence the student must retain and the human work being graded. That would be more useful than making students guess which invisible line a lecturer has in mind. Monitoring needs the same care. The regulations tell institutions to establish mechanisms for monitoring AI use. That could mean sensible record keeping and review. It could also become a market for unreliable detectors, invasive surveillance and accusations based on a probability score. The current HEC guidelines themselves warn that AI can generate convincing errors and disrupt traditional assessment. An operational framework should not answer one unreliable system by purchasing another. No student should face discipline because a classifier says prose looks synthetic. Evidence should come from the learning process: drafts, citations, oral explanation, lab notes, version history, reproducible calculations and the student's ability to defend decisions. Automated indicators may trigger a conversation. They should not become the verdict. The appeals process belongs in the framework, not in a future footnote. A student needs to know who reviews an allegation, what evidence is available, how the student can respond, how disability accommodations and language differences are considered, and where a decision can be challenged. Staff need the same clarity when an approved tool changes its terms or a research workflow crosses a policy boundary. The regulation's promise of equitable access is equally practical. If one course quietly assumes every student can pay for a premium model, the policy has already failed. Institutions need a supported baseline tool or a non-AI path that does not punish the student. Campus connectivity, device access, disability support and local language performance all affect whether the same rule is actually the same opportunity. Mauritius is a multilingual island country with public and private institutions serving students from different backgrounds. A model that performs well in polished English but poorly in French, Mauritian Creole or domain-specific language can widen gaps while advertising personalization. Minister of Tertiary Education, Science and Research Kaviraj Sharma Sukon told the consultation that AI could identify individual strengths and weaknesses and deliver targeted learning materials. That is a plausible application. It is not yet an established outcome. Before an institution uses a model to guide a student's path, it should publish what data enters the system, what recommendation comes out, how accuracy is tested, whether groups experience different errors and who can override the result. Personalized learning becomes personalized surveillance when the institution collects everything and explains nothing. The operational guide should separate low-risk help from consequential decisions. A study assistant suggesting a practice question is different from a system steering a student away from a course, flagging academic risk, assigning a grade or influencing admission, scholarship or discipline. The second category needs stronger validation, human approval, an explanation and an appeal. Research brings another layer. The regulations require AI policies for research, disclosure of generated content and data, compliance with authorship and citation norms, and safeguards for privacy and intellectual property. A practical research procedure should tell a scholar which data may enter a public model, how confidential material is separated, how prompts and outputs are retained, when generated code must be reviewed, how model and version are recorded, and how an AI-assisted result can be reproduced. It should also protect the difference between assistance and authorship. A model cannot take responsibility for a paper, answer a research-integrity inquiry or consent to the use of a participant's data. The human authors and institution remain accountable even when the tool produced the sentence that caused the problem. The Commission's role now becomes important. The law allows it to issue common standards, convene working groups, conduct reviews and order correction. Minister Sukon proposed a dedicated committee that would meet every three to six months to keep institutional guidance current. That cadence is promising if it produces visible maintenance. The committee should publish its membership, meeting dates, conflicts of interest, revision log, incident themes and reasons for major changes. Students, faculty, researchers, disability advocates, privacy specialists and institutional technologists all need a seat somewhere in that loop. Vendors should provide evidence, not write the rules around their own products. The national framework can also prevent every institution from reinventing the same forms and controls. Mauritius could maintain shared templates for assessment declarations, research risk review, model procurement, incident reporting, student appeal, accessibility testing and vendor exit plans. The common layer should set a floor. Institutions can add stricter controls for medicine, law, engineering, education or sensitive research. Procurement deserves its own checklist. Before a campus tool is approved, the institution should know where data is processed, whether prompts train the vendor's model, how long records remain, which subcontractors receive them, what security evidence exists, how accessibility was tested, whether outputs can be audited and how the university retrieves its data when the contract ends. A cheap pilot can become expensive dependence very quickly. Environmental sustainability is already named in the regulations as a principle of responsible use. The operating manual can make that concrete by asking institutions to match the tool to the job, measure high-volume workloads, avoid wasteful default generation and include energy and infrastructure cost in procurement rather than treating compute as weather. None of this requires banning AI from campus. The rules explicitly call for AI literacy, curricular integration, research and faculty development. Mauritius is not trying to freeze higher education before the chatbot arrived. It is trying to give institutions permission and boundaries at the same time. That balance is the interesting bit. Training educators matters because the best policy is useless if lecturers cannot translate it into an assignment. Institutions should give staff examples, course-specific design help, protected time and a place to report failures without being punished for trying an approved tool. Students need the policy in plain language before the assessment, not buried in a governance portal after an allegation. Administrators need logs and escalation routes. Researchers need secure alternatives. Leaders need a dashboard that measures access, incidents, appeals and outcomes instead of counting how many tools were purchased. The consultation announcement does not publish the participant recommendations, final framework, implementation deadlines or compliance tests. It does not say how students will participate in ongoing review. Those are open questions, not small omissions. But Mauritius has already done something many education systems avoid. It put enforceable duties around disclosure, equity, monitoring, human dignity and correction before declaring the technology inevitable. Now the country has to make those duties usable. The real operating manual will not be judged by how often it says responsible AI. It will be judged when a lecturer can set a clear rule, a student can challenge a bad accusation, a researcher can protect sensitive data and an institution can stop a harmful system before the harm becomes a case study.