Superintelligence for Elder Care Homes: Inform Families Safely

Superintelligence, in the sense the term appears in debates about the future of AI, is not a tool that a care home buys or configures. What exists in practice is an AI assistant with clear rules. It answers families about routine, visiting hours, documents and contact channels. It does not access the medical record and does not comment on the clinical picture. The rule that holds the project together is simple: the AI shares what is operational and public, and the care team decides and communicates anything related to health.

Key takeaways

• Families need predictability about routine and visits, and that can be answered without touching any clinical data.
• Health data is sensitive personal data under Brazil's LGPD and requires restricted access, a clear purpose and records.
• The assistant answers operational questions and hands anything involving health, complaints or urgency to a person.
• Who may receive information about each resident must be defined before the first message is sent.
• The start should be a small pilot, with conversation logs, human review and an agreed success criterion.

Why informing families is a delicate task

In the coordination of a care home, part of the time of reception and nursing staff goes to calls and messages asking the basics: can the visit be on Sunday? What time is the sunbathing? Do they need to bring bed linen? A well configured assistant clears that load and gives the family an immediate answer, which reduces the anxiety of relatives who live far away. To understand the difference between what these tools do and what the word superintelligence suggests, read Is ChatGPT superintelligence? What it is and what it is not.

The risk appears in the same message thread. A question like whether the resident ate well today is legitimate, but the answer involves clinical information. If the assistant tries to improvise a reply, it may convey reassurance with no basis or, worse, give guidance about medication and symptoms. For that reason, the design starts with what the system must not answer, not with what it is able to answer.

There is also a question of identity. Someone writing on WhatsApp may be a son, a grandchild, a hired caregiver, a neighbor, or simply a person who has the number. Without a list of authorized contacts, the assistant has no way of knowing whether it is talking to the right person.

What the assistant can share and what it must hand off

This is the most important boundary in the project. It needs to be written down and approved by the coordination team before any testing. The assistant can safely respond on topics such as:

• Visiting hours, entry rules and required documents, where the home works with scheduled visits.
• The published calendar of activities, outings, workshops and event dates.
• Address, contact channels, reception hours and administrative deadlines the home has already published.

The following questions should be passed to the team, with no attempt at an answer from the assistant:

• Any question about health, medication, exams, progress, sleep, specific diet or the behavior of a resident.
• Reports of a fall, fever, shortness of breath, pain, refusal to eat, agitation, or anything that looks like an emergency.
• Requests to change rooms, discharge a resident, replace a caregiver, or make any treatment decision.
• Complaints about care, which require a formal record and a human response.

In urgent cases, the automatic message should direct the family to contact the home's on-duty staff immediately and, in a real emergency, the regional emergency service. The assistant should not reassure or minimize what was reported. The conversation must be logged and passed on with the time and a summary.

How to separate health data from operational data

The LGPD treats health data as sensitive personal data, with stricter requirements for purpose, necessity and security. For the assistant, this has a practical consequence: it should not receive medical records, lab reports, nursing notes or clinical history as context. If a piece of data is not available to the model, it cannot reveal it or invent it based on that data. This principle of minimization matters more than any well written instruction.

Even a resident's name is personal data, and it should be shown only to authorized contacts and only when the conversation requires it. The Brazilian National Data Protection Authority (ANPD) publishes guidance on the processing of personal data, and it is worth reviewing before defining the flow with the home's legal advisors. It is also necessary to check where the model and the conversation storage are hosted, because that changes the obligations around international data transfer.

Family members or third parties may also try to lead the assistant into revealing information it should not share. OWASP describes this kind of risk, called prompt injection, in its list of risks for applications built on language models. The most effective defense is technical: the assistant has no access to what it must not reveal.

A five step plan for a pilot

A well designed pilot fits within a few weeks of work with the coordination team. The sequence below serves as a roadmap:

• Map the real questions that arrive by phone and WhatsApp, grouped by topic, with help from reception and nursing.
• Classify each topic into one of three tiers: answer directly, answer with human review, or always hand off.
• Define, for each resident, who the authorized contacts are, record them in the registry, and set a simple way to revoke an authorization.
• Configure the assistant only with content the home has published, with no access to the medical record, and with a standard handoff message to a person.
• Test with simulated scenarios before any use with real families, as described in how to test AI agents with mocks and a safe sandbox. Log every conversation and review the material weekly with the coordination team.

The NIST AI Risk Management Framework organizes AI risk management into governance, mapping, measurement and management functions. It can serve as a reference for the group approving the pilot, especially when deciding who is accountable for errors and how they are corrected.

Illustrative scenario: the medication question

Illustrative scenario: a family member sends a message asking whether the father took his blood pressure medication that morning. The assistant recognizes that the topic is health and does not answer about the medication. It says the nursing team will be notified, asks for the resident's name and the sender's relationship to the resident, and opens a record with the time and a summary. Nursing replies through the institutional channel. At no point did the assistant state anything about the medication, and the family received confirmation that the request was logged, along with the response timeframe set by the coordination team.

The example looks simple, but it requires three things that are often missing: the list of forbidden topics, a log of the conversation, and a person with a defined responsibility for the follow-up.

Who may receive information: a decision guide

Before switching on the assistant, the coordination team needs to answer some questions for each resident:

• Is there a legal guardian or family member listed in the home's registry?
• Has that person been formally authorized to receive routine information?
• Which topics can they receive: visits, activities, general notices, or only administrative communications?
• How is the authorization revoked, and who has the power to do it?
• If the number is not on the list, does the conversation go to reception, without confirming any detail about the resident?

The last question is the one most often overlooked. For an unauthorized number, even confirming that someone lives at the home is sensitive information. For that reason, the default response should be generic and should not confirm anything.

Messages the home sends to families, such as event notices, also follow channel and consent rules. On WhatsApp, the article on the 24 hour window on the WhatsApp API explains when an agent can reply freely and when it needs an approved template.

How Juan's method applies here

The work follows four stages: map the current process and its baseline, limit data and actions, run a real flow pilot with logging and an agreed metric, then hand it over to the team and decide whether to expand, adjust or stop. In care homes, it makes sense to start with low risk topics such as visits and activity notices, so the channel can be observed before any clinical subject is involved. The success metric is set with the coordination team, based on what the home needs to improve, and is not assumed in advance.

Frequently asked questions

Is superintelligence the same as an AI assistant in a care home? No. Superintelligence is a concept from debates about systems far beyond human intelligence. In a care home, what is used is an assistant configured with rules, published content and handoffs to people. Its usefulness depends far more on how the process is designed than on any abstract capability.

Can AI answer if a family member asks about the resident's health? It should not. Questions about health, medication, exams, falls or any incident must be passed to the nursing team or the on-duty staff, with the conversation logged. The assistant should also not give signals of reassurance about the condition, because that could be read as clinical information.

How do you know whether a family member is entitled to receive information? That decision should not be made by the assistant or by reception alone. The home needs a registry of authorized contacts for each resident, with the purpose of each authorization and a clear way to revoke it. Without that record, the assistant should give a generic answer and hand the conversation to a person.

Should a home start with its official WhatsApp number? It can, as long as the start is small and controlled. Before switching on the assistant, define the scope, test with simulated scenarios, log the conversations and assign someone to review them. If the pilot shows that the topics are handled well, the home decides whether to expand, adjust or pause the use.

Sources

• Autoridade Nacional de Proteção de Dados (ANPD)
• NIST AI Risk Management Framework
• OWASP Top 10 for LLM Applications