OpenAI built an advertising campaign for its voice product around older women. A woman knitting a sweater for her grandson, thinking out loud about needle size. A woman thirty minutes out from a lecture on the history of sound, asking the system to fact-check her dates. A woman haggling with a rare book dealer in Paris through live translation.

The representation is real, and people were right to praise it. Scroll into the comments and you find the other reading, which calls it predatory.

In this one I argue that both readings stall in the same place, because both are asking whether this is good for grandmother or bad for grandmother. The question underneath it is about authority, and the most charming moment in the entire campaign is the one where the elder corrects the machine and the correction sticks.

Then I take you to where these same systems are already touching the same demographic, and show you what happens to that authority there. And I tell you about a treaty being drafted in Geneva right now that almost nobody in the AI governance field is watching.

Full written version on theajayieffect.com.

As always, hit that like button, subscribe for more, and remember: you are capable of big things.

#AIGovernance #OlderAdults #AgingAndAI #AIEthics #TheAjayiEffect #HumanRights #MedicareAdvantage

Word count: 215.

LONG-FORM BLOG POST

The Grandmother in the Commercial and the Treaty Nobody Is Watching

By Dupé Ajayi | The Ajayi Effect

A woman is knitting a sweater for her grandson and she is worried about needle size.

She says it out loud, the way you would to somebody sitting across from you, and the voice on the other end tells her that going up a size is going to make the thing baggy, and she says back that baggy is popular now anyway. That is the opening of an advertising campaign OpenAI built for its voice product, a campaign constructed almost entirely around older women. In another sequence a woman is half an hour out from delivering a lecture on the history of sound and she asks the system to check her dates while she simultaneously asks whether the trains are delayed and whether it is going to rain that afternoon. The system reports back that one of her dates needs correcting, that Thomas Edison’s tinfoil phonograph belongs in 1877 rather than 1865. She laughs and tells it she knew that, she was testing it. Elsewhere a woman haggles in real time with a rare book dealer in Paris through live translation, holding her price because the book is marked up. The campaign, developed with a creative agency and covered across the trade press, was built to make the technology feel warm and to unsettle the assumption that older people arrive last to everything.

I want to credit what the campaign actually accomplishes before I complicate it, because drive-by criticism is not the practice.

The representation is real and it is not nothing. Older women appear in technology marketing overwhelmingly as the confused party, the one who needs the grandchild to come fix the router, the setup rather than the person delivering the line. These spots refuse that entirely. The women in them are quick and curious and dry, and they are unmistakably running the interaction. The woman who says she was testing the machine is the most self-possessed person in the campaign. People who praised it for that were reading it correctly.

Scroll beneath the video and you find the other reading, delivered with considerable heat. People calling the campaign predatory. People calling it bleak. People arguing that the family warmth is doing concealment work, that a company has walked a product engineered to sound like a friend directly toward people who may have fewer and fewer friends left, that systems built to agree with whoever is talking to them are a specific hazard for somebody already alone in a house. People were not being unreasonable. They were describing something they have watched happen.

So the same thirty seconds produces dignity to one viewer and predation to another, and I think both readings stall in the same place, which is that both of them are asking whether this is good for grandmother or bad for grandmother. That is a question about a product. The question underneath it, the one that determines what actually happens to millions of older people over the next decade, is a question about authority. In the world these systems are genuinely being deployed into, who holds the power to correct them, and what happens when the correction is refused?

The commercial accidentally filmed the answer, and then sold it as a feature.

Look again at the most charming beat in the entire campaign. The machine tells her the date is wrong. She holds her ground. She says she knew, she says she was testing it, and the moment lands because she is unmistakably the final authority in the exchange. She checks the system. She weighs what it says. She decides what to do with it.

That relationship is the product being advertised. And it is precisely the relationship that older people are denied nearly everywhere else these systems have already arrived.

What happens when the correction has nowhere to land.

There is federal litigation in the District of Minnesota involving UnitedHealth Group and a subsidiary over an algorithmic tool called nH Predict, used in connection with post-acute care for Medicare Advantage beneficiaries. Post-acute care is the skilled nursing and rehabilitation that older people receive after a fall, a fracture, a stroke, the weeks that determine whether somebody walks again or does not. The tool projected how many days of recovery a patient would require by drawing on historical data from patients it treated as comparable, rather than on the clinical assessment of the physician standing in the room.

The complaint, filed on behalf of the families of two deceased beneficiaries, alleges two figures that ought to be read together and almost never are.

It alleges that when a denial was actually appealed, it was reversed roughly nine times out of ten. And it alleges that only about two tenths of one percent of people ever appealed.

The company disputes the characterization. It has stated that the tool does not make coverage determinations, that it operates as a guide to inform providers, families, and caregivers about the assistance a patient may need in the facility and after returning home, and that coverage decisions rest on Centers for Medicare and Medicaid Services criteria and the terms of the member’s plan. The company has said the suit lacks merit and that it will defend itself. In March of 2026 a federal magistrate judge granted broad discovery into the insurer’s algorithmic claims processes, so considerably more of this record will become public than is public now. Separately, the United States Senate Permanent Subcommittee on Investigations has examined how major insurers deployed automated technology in Medicare Advantage prior authorization for post-acute services.

I am recording the denials beside the allegations because a clean record is the discipline, and because the allegation is not actually the center of my argument.

The center is the distance between those two numbers. A determination that loses nearly every contest it enters, pointed at a population that nearly never enters one. That distance is not a technical defect awaiting a patch. It is a governance design, and it is what you produce when a system is correctable in principle and uncorrectable in practice, aimed at people who are recovering from a broken hip, who are frightened, who are managing an insurer on the telephone while a facility asks their daughter how the family intends to pay.

Set the two images beside each other and the shape of the thing is unmistakable.

In the marketing, the older woman is the authority and the machine defers to her. In the deployment, the machine renders a determination about her body and her recovery and her savings, and her authority to correct it survives only on paper, inside an appeals process that two tenths of one percent of people ever reach. Same demographic. Same family of technology. Opposite standing.

There is a treaty being written about this right now, and the AI field is not watching it.

In April of 2025 the Human Rights Council at the United Nations adopted by consensus a decision establishing an intergovernmental working group with the mandate to elaborate an international legally binding instrument on the human rights of older persons. His Excellency Ambassador Carlos Mario Foradori of Argentina was elected Chairperson-Rapporteur. The working group held its organizational session at the Palais des Nations in Geneva in February of 2026 and meets for two five-day sessions each year, moving through the framework, the architecture, and the guiding principles of the instrument before drafting specific articles.

This is the same class of instrument as the Convention on the Rights of the Child. It binds. It creates obligations for states rather than aspirations for them. It exists because protections for people in old age have been distributed thinly across other instruments and applied inconsistently, which is exactly the fragmentation the convention is meant to repair.

Human Rights Watch submitted a written statement to the organizational meeting arguing that older persons and their representative organizations should participate on a treaty drafting group, and that drafts, working documents, and submissions should be published well ahead of each session so civil society can prepare rather than react.

Read that again slowly. An international human rights organization found it necessary to formally request that the people the treaty is about be seated where the treaty is written. The necessity of that request is the entire diagnosis. It tells you what the default posture is, and the default posture is that older persons are the subject of the instrument rather than parties to it.

The Institute for Ethics in AI at the University of Oxford has convened directly on this intersection under the title Convention on the Rights of Older Persons in the Age of AI, organized around the proposition that artificial intelligence offers tools to support intercultural dialogue and to gather the voices of older persons worldwide to inform the drafting.

That instinct is sound and the people behind it are doing serious work. Using these tools to reach older people who cannot travel to Geneva, who do not speak the working languages of the United Nations, who are in rural districts and care facilities and in countries whose delegations will never raise them, is a legitimate and potentially significant application.

It is also, standing alone, participation infrastructure and nothing more. It gathers voice. Gathering voice is not conferring authority, and the entire history of the field I work in is the history of that substitution being mistaken for progress. Sherry Arnstein made this legible for urban planning in 1969 and we are still relearning it, one domain at a time, at the expense of whoever is currently being consulted.

The framework, applied.

I have been building an AI governance framework I call the Translation Layer™, which I define as the governance infrastructure that enables communities most affected by AI systems to exercise actual decision-making authority over the development, evaluation, deployment, and nurturing of those systems. It has three components, and I insist on all three because the first without the other two is precisely how sincerity becomes documentation of sincerity.

Participation infrastructure. Older persons and their representative organizations hold reserved standing wherever the decisions are made. That means seats on a treaty drafting group in Geneva, and it equally means seats in the design and evaluation of any algorithmic system determining eligibility, care duration, benefit continuation, housing, or legal capacity. The people whose recovery is being projected were not present when the projection rules were written, and there is no chair currently reserved for them.

Evidence redefinition. When an older person or their family says the determination was wrong, that testimony carries evidentiary weight against the model’s output, with a documented obligation to respond in writing. Not an anecdote. Not a support ticket. Not a data point subordinate to the aggregate. Evidence, weighted as evidence, on the record.

Enforcement authority. Somebody accountable to older people can pause a deployed system, compel an audit, require retraining, or decommission it. Absent this third component, the first two generate a well-attended consultation while the system continues operating exactly as designed, which is the arrangement we already have and the one that produced a ninety percent reversal rate met by a two tenths of one percent appeal rate.

A convention drafted with those three written into its articles would be among the most consequential AI governance instruments of this decade. Most of the AI governance field is not tracking it, because most of the AI governance field is tracking the labs, and the labs are not where this particular authority is going to be settled.

Where this leaves the sweater.

The commercial is not the problem and the commercial was never the point. The woman with the knitting needles is going to be fine, the sweater will come out slightly baggy, and baggy is popular now.

What the campaign did, without meaning to, was film the exact thing older people are being denied everywhere the technology has already landed. It put correction authority on screen, made it charming, and sold it as a feature, during the same period in which a binding international instrument that could make that authority a right is being drafted by a body that had to be formally asked to invite older persons into the room.

We are not short on understanding here. The remedy is not mysterious and it is not novel. It is participation carrying evidentiary weight and backed by enforcement, written into the instrument while the instrument is still open, which it is, which is the window, and windows on treaties close for decades at a time.