Superintelligence in Higher Education: Registrar, Retention, Ethics

Superintelligence in Higher Education: Registrar, Retention, Ethics

Superintelligence in higher education is not something a college operates today. What exists are increasingly capable AI systems, and the practical decision is to define in advance what they may do in the registrar's office, in student support and in academic integrity. An institution that sets limits, owners and records now is ready for whatever level of capability comes next.

Key takeaways

• Superintelligence is a long-term hypothesis. The institution runs narrow-scope AI agents today, and governance should be designed for them.
• In the registrar's office, AI works well to inform, guide and organize requests. Decisions that change a student's academic standing stay with people.
• On dropout, the value is in noticing signals early and triggering human outreach, not in automatically classifying students.
• Academic integrity depends on clear AI-use rules by course and by type of assessment, more than on detectors.
• The safe path is to map, limit, pilot one real workflow with records, and only then decide whether to expand.

What does superintelligence mean for a higher education institution?

In the technical debate, superintelligence describes a system that outperforms humans at practically every relevant cognitive task. That does not describe any tool a college can buy today. Chat assistants, support agents and faculty copilots are powerful but limited systems. They make mistakes, invent information and depend on the context they are given. To keep the concepts apart, read the difference between AGI, ASI and AI agents and the piece on myths that get in the way of business decisions.

Why talk about it at all? Because the direction of travel is toward more capable systems, and governance decisions at a university last a long time. An AI use policy, a data policy and a service design written today will be applied to different tools a few academic terms from now. What has to stay stable are the principles: who answers for a decision, which data can be used, what the student can contest and how the institution records what was done.

A good question for the board is not when superintelligence arrives. It is this: if the system we use became much more capable tomorrow, would the limits we set still protect the student and the institution? If the answer depends on the goodwill of the tool, the design is incomplete.

The registrar's office: what AI can do and what needs a human decision

The registrar handles a large volume of repetitive questions and rule-based processes: enrollment, re-enrollment, leave of absence, transfers, transfer credit, certificates, transcripts, graduation clearance. It is good ground for AI precisely because much of the work is applying known rules to a case with documents that can be checked. If you have looked at this in schools, compare it with the piece on AI for schools, admissions and parent communication, keeping in mind that higher education has more course rules, calendar constraints and financial ties.

A useful way to split the work is by type of action:

• Inform: explain deadlines, required documents, process steps and where the student can follow a request. This can be automated as long as the answer comes from current regulations and calendars, not from the model's memory.
• Collect: open the request, receive documents, check that it is complete and route it to the right queue. This can be automated with simple validations.
• Look up: tell the student the status of their own request or a pending item. This requires authentication and access limited to that student's data.
• Decide: approve a leave of absence, grant transfer credit, clear graduation, allow an exception. This stays with the responsible person, with AI at most preparing a summary for them.
• Communicate a sensitive decision: failure due to absences, dismissal, a hold that blocks enrollment. Prefer human contact or, at minimum, a reviewed message.

The central point is that AI should not be the source of the rule. The source is the regulation, the calendar and the academic system. The agent retrieves the current passage, answers based on it and, when it finds nothing, says it does not know and hands off. When a rule changes, you update the document, not the behavior of the model.

Dropout: using AI to look after students without labeling them

Dropout rarely has a single cause. It can involve financial difficulty, a clash with work schedules, performance in foundation courses, lack of belonging, health problems or plain disorganization with deadlines. AI helps more in noticing scattered signals and organizing outreach than in predicting who will leave.

Signals that are usually available in the institution's systems and can feed an alert:

• a drop in access to the virtual learning environment;
• prolonged absence from in-person or live online activities;
• document requests that tend to come before a transfer or cancellation;
• recurring payment issues combined with less contact;
• questions to the registrar about leave of absence or cancellation.

The risk is in the next step. A model that assigns a risk score to each student can become a label that, without anyone noticing, shapes the attention the student receives or how they are treated. To avoid that, the alert should point to the observable reason (for example, no access for several weeks) and not to an opaque probability. The next action should be human: a coordinator, tutor or support office makes contact, listens and offers real paths, such as financial guidance, flexibility already allowed by the rules or academic support.

A proportionality rule also helps: if the alert leads to a benefit, such as an offer of support, the cost of an error is lower. If it leads to a restriction or worse treatment, it needs review and a right to contest. The piece on automated decisions under LGPD and the right to review covers this point in detail.

Academic integrity and ethics: rules that survive more capable AI

The temptation is to treat integrity as a detection problem: buy a tool that says whether a text was written by AI. Detectors err in both directions, and an unfair accusation weighs heavily on a student. Also, the more capable the model, the harder it is to tell the text apart. A policy that depends on detection ages quickly.

The more durable path is to define, by type of activity, what is allowed, what is allowed with disclosure and what is prohibited. Some of the decisions the institution needs to make and publish:

• In which assessments AI use is free, in which it requires a statement of what was used and in which it is not allowed.
• How the student discloses use: a footnote, an appendix with the prompts used, or an oral conversation about the work.
• Which assessment formats weigh more: oral defense, in-class production, projects with versioned stages, cases using original data.
• How the process works when there is suspicion: who reviews, what evidence counts, the timeline and how the student responds.
• What faculty may do with AI, such as grading or giving feedback, and which student data may not be pasted into external tools.

The last point is often forgotten. Ethics applies to the institution too: using AI to pre-grade papers, summarize student reports or analyze performance has privacy and fairness implications. A simple rule helps: if a decision that affects a student's grade, enrollment or reputation went through AI, the student should be able to know and ask for human review. References such as the NIST AI Risk Management Framework and the OECD AI Principles offer a shared vocabulary for this governance, without replacing internal policy or Brazilian law.

Student data, LGPD and the service channel

The registrar and student support deal with personal data, sometimes sensitive: health information in special-regime requests, financial status, data on minors in outreach programs, identity documents. When AI enters, three questions need written answers before any pilot. (LGPD is Brazil's general data protection law.)

• What data does the agent see? Each workflow should access the minimum necessary. An agent for general information does not need to see individual records.
• Where does the data go? If the model runs on a service outside the country, international transfer questions arise. The piece on international data transfer under LGPD when using AI providers helps structure the analysis, and the ANPD is the official reference for guidance.
• Who confirms the person is the student? Before showing a grade, a hold or a request number, the agent should authenticate through a method the institution already uses, not through a name and birth date that are easy to find out.

On WhatsApp, a channel where many students already prefer to talk, the channel's own conduct rules also apply: outbound messages need permission and approved templates, and conversations about delicate subjects deserve a quick handoff to a person. Agents that read documents and third-party messages are also exposed to malicious instructions hidden in text, which OWASP describes in its list of risks for applications built on language models. In practice, this means limiting the agent's tools and never giving it the power to change academic records on its own.

How to start: map, limit, pilot and hand over

The method we apply to other sectors works here without major adaptation. Each step has a concrete result:

• Map: choose one workflow, such as a certificate request or re-enrollment questions, and describe how it works today. Who receives it, through which channel, under which rules, in how much time, with which frequent errors. This picture is the baseline for knowing later whether things improved.
• Limit: define which data the agent sees, which actions it can take, which topics it always hands to a person and what message it uses when it does not know. Record all of this in a short document approved by the registrar and by legal or the data protection officer.
• Pilot: run the workflow with a small group, recording each conversation, each handoff decision and each correction made by humans. Agree in advance on a success measure, for example requests completed correctly on the first attempt, without promising a number.
• Hand over: give the institution's team the records, the rules and the adjustment procedure, and decide together whether the workflow expands, is adjusted or stops.

When the institution has no in-house team to monitor it, managed operation can be provided by Be Human, with permissions, observability and human approval. The decision to expand remains the institution's.

Illustrative scenario: a leave of absence request on WhatsApp

This is an illustrative scenario, unrelated to any real institution. A student writes in the evening asking how to take a leave of absence. The agent identifies itself as a virtual assistant, confirms the student's identity with a code sent to the registered email and explains, based on the current regulation, the deadlines, the effects on scholarships and financing, and the documents required.

The student says she is facing financial difficulty. At this point the agent offers no exception and promises nothing. It records the request, attaches the conversation summary and routes it to the student support office, telling her that a person will contact her during office hours. The next day, the registrar reviews and decides, and the agent only communicates progress. If the student questions the outcome, the history shows what was said, based on which rule and who decided.

The design has three safeguards: the answer comes from an official document, the decision stays with a person and everything is recorded. They hold regardless of how capable the model behind them is.

Frequently asked questions

Is superintelligence already being used in colleges? No. What institutions use are narrow-scope AI models that answer questions, summarize documents and support tasks. The term superintelligence is useful for planning long-term governance, but it does not describe any tool you can buy today.

Can AI decide on a leave of absence, transfer credit or dismissal? It should not decide alone. These acts change the student's academic standing and need an identifiable owner, with the possibility of review. AI can organize the request, check documents and prepare a summary for the person who decides.

Do AI text detectors solve academic integrity? Not on their own. They can err and accuse an innocent student, and they tend to lose effectiveness as models improve. It is safer to combine clear use rules by activity, disclosure of use, assessment formats that require students to show their reasoning and a fair review process.

Where should a college start? With a single registrar workflow that has clear rules and low risk, such as guidance on documents and deadlines. Map how it works, set limits on data and actions, run a pilot with records and an agreed measure, and only then decide whether to expand, adjust or stop.

Sources

• NIST: AI Risk Management Framework
• ANPD: Autoridade Nacional de Proteção de Dados
• OECD AI Principles
• OWASP Top 10 for LLM Applications