Lately I keep seeing the same advertisement: three senior women sitting around a phone, asking ChatGPT where to go to eat. No typing. No menus. No UI interactions shown.
I started this research because a version of that ad was happening in my own life. Older people I know were starting to pick up AI chatbots and make them a daily habit. For once, the marketing matched reality, which made me curious about the part it leaves out.
What I started off studying
For my first attempt at understanding how older adults use AI, I conducted a literature review. A common structure throughout the scholarly articles was AI technologies in elderly care broken down by function. AI appeared in older adults' lives in diverse roles: mobility assistant, emotional supporter, social facilitator, supervisor, and cognitive promoter. Multiple sources noted that by 2030, one in six people worldwide will be 60 or older, that care systems continue to be strained, and that AI integration in elderly care looks like hope for relief.
Initially, what interested me was whether older adults would readily accept these AI technologies.
That was the wrong question.
The people I was studying adopted just fine. What I instead kept noticing was how each paper ended with a section on safeguards including concerns on consent, data security, bias, dignity, and human connection.
So the question changed. If AI was going to become this embedded in older adults' lives, how should those safeguards actually be designed?
The paradox I found
For my second assignment, I moved from breadth to depth and focused on the cost-risk balance of conversational AI assistants in particular. I described a relationship between an older adult and an AI assistant over time, built from the literature and from watching people I know.
How the relationship changes over time
Older adults and conversational AI assistants: a three-phase trust paradox
Phase 1 · Trust
Introduction
A relative or caregiver sets things up, demonstrates a few prompts, turns on voice. Trust is scaffolded by a human being.
Voice lowers the entry barrier, and removes the visible seams.
Phase 2 · Reliance
Independent use
The AI assistant becomes an always-on substitute for search and, increasingly, for conversation: logistics, supplement questions, company.
A social surrogate that never reciprocates.
Phase 3 · Exposure
Concerns
The trust built in phases 1–2 becomes the liability. Users can't easily tell when it is wrong, and isolated conversations risk displacing human interactions.
Confidence doesn't track the ability to catch errors.
The same design choices produce both ends of the paradox.
Constant availability, warmth, and agreeableness ease loneliness and build trust, and enable overtrust, dependence, and exploitation. Benefit and harm aren't separate; they're two sides of the same features.
It came out in three phases.
A relative makes the introduction. The AI assistant becomes an accessible source for information and, increasingly, for conversation. A companion that is never impatient and always available. Then the trust built earlier becomes the liability: confidence in the tool doesn't track the ability to catch its errors, and the availability that makes companionship valuable is the same thing that threatens to displace meaningful human interactions.
This forced a conclusion I hadn't expected: benefits and risks aren't separate; they're two sides of the same design choices, seen at different points in the journey.
Constant availability, warmth, agreeableness, the features that solve problems like loneliness in the beginning, can start to produce dependence and overtrust. You can't just tack safety onto the product, it has to be designed.
Ethics stopped being my closing section and became the center of the problem.
The concept I built, and why I threw it away
My initial interaction concept was a "Senior Mode" for AI assistants, modeled on ChatGPT's mode for Teens.
Where ChatGPT for Teens has content filtering, Senior Mode would have scam and misinformation detection. Where ChatGPT for Teens has parental controls, Senior Mode would have caregiver controls.
The features have the right intention. Safeguarding seniors is at the top of mind.
But the architecture underneath is borrowed from a product built for people with limited legal autonomy. Parental controls exist because a minor cannot fully consent; mapping parent to caregiver carries that assumption across onto adults who never lost the capacity to decide.
It gets worse against my own evidence.
Caregiver controls require a caregiver, so the feature doing the most protective work does nothing for the older adults who don't have one accessible. And a control panel handed to a relative doesn't correct the power dynamic I documented. It formalizes it.
The question I couldn't answer about my own design was simple:
Who does this serve? The older adult, or the family's anxiety about them?
Nothing about me without me
The phrase comes from disability rights and patient-centered care. Oversight belongs to the person being overseen.
But what I ended up with isn't oversight at all.
It's sharing.
An older adult grants standing permissions by topic: his physician may receive conversations about symptoms; his daughter may receive conversations about money. Nothing moves on its own. When he wants his doctor to know what he's been noticing, he says so out loud: "Share this with Dr. Patel."
The system then assembles the parts of the conversation that fall inside that permission.
Three keys have to turn before anything leaves.
- The standing permission. Granted once and revocable. It sets a ceiling, not an action.
- The explicit request. Made in the moment by the user. Permission alone never sends anything.
- The preview. The system shows or reads aloud exactly what is about to be sent: "Here's what I'll send Dr. Patel. Anything you want out?"
This step is crucial because it creates a failsafe for situations where the AI incorrectly judges what counts as "medical". For example, a symptom conversation that also contains "I've been drinking more since my wife died" is clinically relevant, but possibly not what he meant to hand over.
Nobody can audit a classification policy, but anyone can look at four paragraphs and say, take that out.
You may wonder, should the system suggest moments of sharing? The system may offer, "Do you want your doctor to see this?" But rarely. Never twice after a no. Never urgently.
A suggestion is still the system deciding, and these users already trust it more than they should.
Two rules keep user agency at the forefront.
Sharing is one-directional. Recipients can never request it, and are never told that something shareable exists. A feature you can use is a feature you can be pressured into using.
The compartments are real. The doctor never sees the money conversations. The daughter never sees the symptoms.
What changes here is what the AI assistant is. In my trust paradox, it quietly displaces human contact. That was the harm. In this design, it routes into human contact instead.
Success is not how well the AI answers or how well the user can discern misinformation, but becomes how many well-informed humans are in the loop.
Why it shouldn't be a mode
Here is where I ended up furthest from where I started.
Everything above is good design for anyone.
Scoped sharing. A preview before you send. Voice-first access. None of it is medicine reserved for the old. Anyone would want to hand a doctor an accurate account of three weeks they can't quite reconstruct.
Putting these features behind an age restricted label makes protection contingent on identifying yourself as someone who needs protecting.
Nobody wakes up and enables Senior Mode. The needs may arrive at any moment, which is what universal design was invented for.
The uncomfortable collaboration lesson
There was one major insight I couldn't ignore.
For my second assignment, I wrote the narrative myself, used Claude to build the visual synthesis, then cut the claims that couldn't carry a citation. I kept the judgment and delegated the assembly.
It was the AI collaboration model I'd want for everyone, but it worked because I had the literature in my head and a reason to be skeptical.
The elderly user described in my paradox has neither.
Designing for verification, not just helpfulness, was essential to the solution.
I began this quarter asking how to get older adults to accept AI.
Acceptance was never the hard part. They accept it readily, sometimes more readily than they should.
The hard part is building systems that keep handing authority back, day after day.
We are all future elderly users. I'm designing now for the person I will one day be.