The AI That Assumes You Exist

Rosa Mendoza is 71 years old. She has lived in the United States for thirty-one years. She raised three children here. All three graduated from American high schools. Two graduated from college. She paid taxes using an Individual Taxpayer Identification Number for twenty-eight of those years. She cleaned houses, then office buildings, then worked in a restaurant kitchen until her knees gave out at 64. She has contributed to a country that does not, in the formal language of its benefit systems, know she exists.

The benefits navigation agent from Series 2, the AI that identifies every program a person qualifies for and helps them apply, gave Rosa an honest answer. Medicare requires forty quarters of covered employment with a Social Security number. Rosa has a taxpayer identification number, not a Social Security number. She is ineligible. Social Security retirement benefits require the same qualifying quarters on the same record she does not have. The Affordable Care Act marketplace is closed to undocumented immigrants regardless of how long they have lived, worked, and paid taxes in the country.

The agent is not broken. It is working correctly. The programs it navigates were designed for people with documentation Rosa does not have. The agent can tell her what she qualifies for: emergency Medicaid for acute care, community health center services on a sliding fee scale, and a small number of state-funded programs that do not require immigration status verification. That is the list. It is short. It does not include the programs that would most help a 71-year-old woman whose body has been worn down by five decades of physical work.

Rosa is not invisible to the healthcare system. She is visible as a patient who cannot pay. She is invisible as a citizen owed anything in return.

The Undocumented Elder

#

Roughly 600,000 undocumented immigrants in the United States are over the age of 60. The number is an estimate because the population is, by definition, difficult to count. They have lived here for an average of more than twenty years. Many have worked continuously, paid taxes, and raised families who are themselves citizens. They age with the same bodies, the same chronic conditions, the same cognitive vulnerabilities as everyone else. They age without the safety net.

The specific barriers are worth naming precisely. Medicare Part A requires forty quarters of covered employment credited to a Social Security record. An ITIN does not generate Social Security credits regardless of how much tax is paid on it. Medicare Part B and Part D are available only to individuals eligible for Part A or who have been lawful permanent residents for at least five years. Social Security retirement benefits follow the same eligibility structure. The Affordable Care Act marketplace explicitly excludes undocumented immigrants from purchasing coverage, even without subsidies.

What remains: Federally Qualified Health Centers provide primary care on a sliding fee scale regardless of immigration status. Emergency Medicaid covers acute care in an emergency. Some states, California and New York among them, have extended Medicaid-like coverage to undocumented immigrants in certain age groups. These programs are real and they matter. They do not provide the coordinated, continuous care that the AI ecosystem this publication has described is designed to support.

The benefits navigation agent from Series 2 works by mapping available programs to the person’s eligibility profile. For Rosa, the map is mostly empty. The agent’s honesty is the problem. There is nothing wrong with the software. There is something wrong with the landscape the software navigates.

The Privacy Architecture Problem

#

An AI benefits navigator that requires disclosure of immigration status to evaluate eligibility creates a record. For Rosa, that record carries risk. Any system that stores her immigration status and connects it to her identity creates a database entry that, depending on the political environment and the data-sharing practices of the platform, could become a liability rather than a resource.

Privacy-protective architectures that evaluate eligibility without creating a deportable record are technically achievable. The system can assess what a person qualifies for based on age, income, and state of residence without asking whether they have documentation. The system can present available programs without requiring the disclosure that makes Rosa vulnerable. Differential privacy techniques and secure computation methods exist that would allow Rosa to interact with a benefits navigator without the navigator creating a record of her undocumented status.

These architectures are not standard. Most benefits navigation systems, whether AI-powered or human-operated, begin with eligibility determination, and eligibility determination begins with documentation status. The design choice to put documentation first is a choice. A different design choice would put the person first and determine what can be offered without requiring the disclosure that creates risk.

The Formerly Incarcerated Elder

#

The prison population in the United States has aged dramatically. Roughly 200,000 state and federal prisoners are over 50, and the number of formerly incarcerated people over 65 is growing as the mass incarceration wave of the 1980s and 1990s produces an aging cohort of people with criminal records who are now navigating elder care systems.

Criminal records create benefit eligibility gaps that compound with aging. Public housing authorities in most states can exclude applicants with certain criminal histories. Some state benefit programs have eligibility restrictions tied to criminal records. The home modification support from Series 3, the programs that fund grab bars and ramp installations and bathroom modifications that keep an older person in their home, may be administered through public housing agencies that exclude people with the records this population carries.

The AI ecosystem assumes a clean background. The benefits agent navigates programs that were designed for people without criminal histories. The home modification agent identifies funding sources that may be closed to someone with a record. The cognitive monitoring system assumes the user can interact with healthcare institutions without the hesitation that comes from years of institutional distrust built in environments where institutions are not safe.

Designing for this population requires specific choices. Benefits navigation that accounts for record-based eligibility restrictions rather than presenting programs the person cannot access. Trust-building interfaces that acknowledge institutional distrust rather than assuming institutional comfort. Connection to reentry-specific services that the mainstream benefits landscape does not include.

The LGBTQ+ Elder

#

The LGBTQ+ elder population faces a specific version of the assumption problem. The personal AI that coordinates couples-based care, that manages shared medical decisions, that involves a partner in cognitive monitoring and advance care planning, assumes a relationship that can be safely disclosed in every care environment the system touches.

For LGBTQ+ elders in many institutional care settings, that assumption is wrong. Research documents that LGBTQ+ older adults in long-term care facilities report high rates of discrimination, harassment, and the pressure to conceal their identity and relationships. The phenomenon has a name: re-closeting. An older adult who lived openly for decades returns to concealment upon entering institutional care because the institution is not safe.

A personal AI that requires disclosure of relationship status or gender identity to access couples-based care coordination assumes a safe environment for that disclosure. In a nursing facility where a same-sex couple has learned to introduce each other as friends, the AI’s assumption becomes a risk. The system that asks about the user’s partner, that offers to coordinate with their spouse, that includes relationship status in care planning documents, is doing exactly what it was designed to do. It is doing it in an environment where the design assumption is wrong.

The design choice that addresses this is separation. The information the AI needs for care coordination can be collected and stored separately from the information that creates risk. Relationship-based care functions can be offered without requiring the disclosure to be recorded in systems that institutional staff can access. Privacy architecture that protects the user’s identity from the institution that is supposed to care for them is a design requirement, not a feature request.

The Elder with Disability

#

A person who is blind cannot use the visual scaffolding system from Series 5. The family photo browser that triggers memory, the life story timeline displayed on a screen, the visual medication organizer that shows which pills to take at which time, none of these work for someone who cannot see the screen. A person who is deaf cannot use the speech-based daily check-in from Series 1. The voice-activated home controls from Series 3 do not work for a person who cannot speak.

These are not edge cases. Roughly 46 percent of Americans over 75 have at least one disability. Vision impairment, hearing impairment, mobility limitation, and cognitive disability are prevalent in the population this ecosystem was designed to serve. An ecosystem that does not accommodate them is an ecosystem that excludes nearly half its target population.

Aging compounds existing disability in ways the AI ecosystem was not built to handle. A person who has used a wheelchair for twenty years and now develops cognitive impairment faces compounded accessibility challenges that the system treats as two separate problems, each addressed by its own silo of adaptive technology. The person experiencing them does not have two problems. They have one life with multiple barriers, and the AI that serves them needs to understand the intersection rather than treating each dimension separately.

Accessibility as a design standard rather than a retrofit is the requirement. The difference matters because retrofits are compromises. A screen reader added to a visual interface after the interface was designed produces a functional but degraded experience. An interface designed from the beginning to work for both sighted and blind users produces a different product entirely. The accessible-first product is usually better for everyone, including the sighted users, because the design constraints force clarity and simplicity that purely visual interfaces often lack.

The Indigenous Elder

#

The relationship to land, community, and care in Indigenous communities does not map onto the assumptions embedded in the home and geography series of this publication. The ambient home monitoring from Series 3 assumes a single-family dwelling with broadband access. Many Indigenous elders live in multi-generational households on reservations where broadband penetration is among the lowest in the country. The fall prediction system assumes indoor environments that match the training data. Housing conditions on some reservations do not match that data.

Indian Health Service provides healthcare to American Indian and Alaska Native people. IHS is chronically underfunded, with per-capita healthcare spending roughly one-third of the national average. The AI health monitoring that this publication has described as a near-term supplement to clinical care assumes a clinical infrastructure that IHS cannot consistently provide. Adding AI to a healthcare system that does not have enough physicians, enough facilities, or enough funding does not solve the underlying problem. It adds a layer of technology to a foundation that cannot support it.

Tribal sovereignty creates specific governance challenges for AI systems operating on tribal land. Tribal data sovereignty, the principle that data generated by tribal members on tribal land belongs to the tribe, not the AI company that collected it, is an emerging legal and ethical framework that most AI health companies have not engaged with. An AI system that collects health data from Indigenous elders on tribal land without tribal consent and stores it on corporate servers is not a neutral technology deployment. It is an extraction.

This publication does not claim expertise in Indigenous health governance. What it can name is the structural mismatch between the ecosystem it has described and the reality of Indigenous elder care, and the specific design choices, tribal data partnerships, offline-capable systems, housing-adaptive monitoring, that would begin to address the gap.

The Design Question

#

Each exclusion named in this piece is addressable with a specific design choice. Privacy architectures that do not require documentation disclosure. Benefits navigation that accounts for record-based restrictions. LGBTQ+-affirming interfaces that separate care information from identity risk. Accessibility built in from the first prototype. Tribal data partnerships that respect sovereignty.

None of these are impossible. All of them are uncommon. The question is not whether the technology can accommodate the people it currently excludes. The question is whether the people building the technology consider them people worth designing for.

Rosa Mendoza paid taxes for twenty-eight years to a country whose AI health systems cannot navigate her eligibility because she does not have the number the system requires. She is 71 years old. Her knees are worn. Her children worry. The agent that could help her most was designed for someone else. The design that would include her is a choice that has not been made.