Summary: Design With, Not For

The room has eight engineers, three product managers, and two clinical advisors. The average age is 34. The whiteboard shows a product roadmap with a launch date six months out. The product is a personal AI health companion for adults over 65. No one in the room is 65. No one in the room speaks African American Vernacular English. No one lives on $1,140 a month. No one conducted their last medical visit in a language other than English. No one has an ITIN instead of a Social Security number. No one uses a wheelchair. No one has hidden a relationship to survive in an institutional care setting. No one lives on a reservation.

The product will launch in six months. It will work well for people who look, speak, earn, and age like the people in the room. For the people this series has named, the gaps will arrive with the product. They will be visible six months after launch, when the adverse event reports begin arriving. They will be harder to fix then than they would have been to prevent now.

The synthesis opens by assembling what the four preceding articles established. Denise Watkins, flagged by a speech analysis system that heard her African American Vernacular English as cognitive anomaly, spent a weekend wondering whether she was losing her mind. She was not. The system could not hear her correctly because it was not trained on people who talk like her. Marvella Johnson, $1,140 a month, no smartphone, no broadband, a community health worker named Raymond who visits twice a month with a clipboard and a caseload of fifty-four people: the minimum viable AI ecosystem costs more than Marvella has after rent and food. Carmen Gutierrez, screened in English, flagged as borderline, diagnosed two years later in Spanish with MCI that was identifiable at year zero: the two-year delay narrowed her intervention window. Rosa Mendoza, thirty-one years in the United States, twenty-eight years of taxes on an ITIN, ineligible for the programs the benefits agent navigates; LGBTQ+ elders re-closeted in institutional settings where the AI’s relationship questions become risks; elders with disabilities excluded by interfaces designed for sighted, hearing, mobile bodies; Indigenous elders whose relationship to land, data, and care does not map onto the ecosystem’s assumptions.

These four people are not edge cases. Denise is part of the 13 percent of Americans over 65 who are Black. Marvella is part of the 23 percent who live on less than $1,500 a month. Carmen is part of the 22 percent who speak a language other than English at home. An elder with a disability is part of the 46 percent of Americans over 75 who have at least one. The edge cases, collectively, are the majority of the population the product claims to serve.

The “for” model fails for a structural reason. By the time the product is stable, the architecture has encoded the exclusions. The training data is fixed. The clinical validation has been conducted on the populations the system was designed for. Retrofitting inclusion into a product designed without it is exponentially harder than designing it in from the beginning. The speech analysis system trained on standard American English cannot be made dialect-aware by adding a patch. The training data must be rebuilt. The validation must be re-run. The cost of the retrofit grows with every month of deployment.

The “with” model is simple to state and difficult to execute: design with the excluded populations before the first line of code is written. Community advisory boards with genuine decision-making authority, not consultative roles — boards with the authority to say no change what gets built. Training data collected from the populations the system will serve, not only from the populations cheapest to recruit: a speech analysis system that will monitor Black older adults must be trained on Black speech. Testing protocols that include performance validation across the full demographic range before deployment, not after launch. Accessibility as a design requirement from the first prototype, not a compliance retrofit.

The history of accessible design supports the case. Closed captions were developed for deaf and hard-of-hearing viewers; they are now used by millions of hearing people in gyms, airports, and living rooms where someone else is sleeping. Curb cuts were designed for wheelchair users; they are now used by parents with strollers and travelers with luggage. Voice interfaces were developed for people with mobility impairments; voice interaction is now the dominant mode for smartphone users across all demographics. Features designed for the margins often become the standard for the center.

The institutional barriers are real. The market rewards building for the affluent first. Inclusive design takes longer and costs more in the pre-launch phase. Changing this requires institutional choices: funding agencies requiring inclusive design as a condition, not a value; the FDA making bias testing before deployment non-negotiable rather than aspirational; Medicare and Medicaid including demographic performance data in coverage determinations. These choices have not been made. Until they are, the market incentive is to build for the comfortable and call it universal.

The synthesis turns the equity test on the publication itself. BlueMirror.life is written in English. It assumes broadband access and device ownership. The populations named in this series, the populations most excluded by the AI ecosystem, are also the populations least likely to read the publication that documented their exclusion. The free-pathway standard established in 13.02 is the publication’s standard for itself: every recommendation includes the accessible alternative. The language limitation is named: BML currently publishes in English, Spanish-language publication is a goal not yet achieved, and the gap between the goal and the reality is the same kind of gap this series has named in the products it reviews.

Return to the room. Now change it. Add Denise’s community clinician who has been documenting adverse events because someone has to. Add Raymond, who knows what the technology needs to do because he does it by hand for fifty-four people. Add Jorge Gutierrez, who knew the cognitive screening was wrong before anyone with a degree did. Add a disability rights advocate. Add an Indigenous health worker from Montana who can explain why tribal data sovereignty is a principle to honor rather than an obstacle to route around.

The product roadmap changes. Some features move up. Some that were not on the roadmap appear. The launch date moves from six months to eight. The two additional months are not a delay. They are the cost of building a product that works for the people it claims to serve. The product that launches at eight months includes dialect-aware speech analysis, an offline-capable version, validated cognitive monitoring in Spanish, a privacy architecture that does not require Rosa to disclose her immigration status, and accessibility from the first screen. The product that launches at six months has a plan to add these things later, after the adverse event reports arrive. The later version costs more, works less well, and reaches the excluded populations after the harm has already been done.

The two months are the argument. Who is in the room when the first line of code is written determines what gets built.

Read the full article at BlueMirror.life.