Source: Evelyn, ElliQ member, daily interaction with an AI companion,
New York State programme. elliq.com
The companion question: Is AI a different thing entirely when you're 80?
A close friend of mine lost her best friend recently.
Her friend's mother has dementia. A significant stage of it. She wasn't told that her daughter had died. She wasn't invited to the funeral. When she asks about her daughter, the family now tells her she's at work.
When you work it through, this is not a cold decision. It's a careful one. A kind one. The family and carers know that telling her would cause acute grief and then, because of how the disease works, that grief would need to be caused again and again, every time she asked. The merciful fiction then holds.
We are already curating reality
There is a question that this story opens up, gently but directly. If we are already making decisions to protect someone's reality, if we are already curating what is true for someone whose grasp of truth is no longer intact, does the arrival of AI change anything? Is protecting someone's fragile inner world fundamentally different when it is a human making that call, versus a piece of technology doing it?
If you are a daughter, a son, a carer, sitting with a parent whose memory and identity are becoming harder to hold, what would you actually want AI to do?
What would you want it never to do?
A personal note on why this series exists
My mother was diagnosed with mixed dementia and Alzheimer's last June. In the weeks after that, I did what most people do. I spoke to others, and I researched.
I looked for anything that might help her stay in her home for longer, anything that might slow things down, or make daily life better for her, or make the relentless weight of it a little lighter for us around her.
What I found in technology was a confusing, blotchy landscape of products across categories; some looked helpful but were not specific for a user with Dementia, many had claims. Most were not available in the UK.
I had no framework to navigate. I had no idea that Companion AI, Assistive AI, and Robotic AI were entirely different things, built on different evidence, for different stages, with different risks.
Nobody had written the guide I needed.
That is what this series is.
The more I have researched this space, the more I have come to believe it is not just a practical story, it is a genuinely surprising one. What’s emerging quietly, in research labs, care homes and the private homes of families navigating dementia, cuts against almost every assumption that the wider culture currently holds about AI and human connection.
When I founded Age and AI, the premise felt obvious but slightly uncertain. Your age shapes how you experience technology. Younger people adopt faster, older people slower, and the middle generation navigates both. It was interesting, but was ‘age’ in itself a story of real depth?
It turns out it is.
Age is a more striking story than I expected.
The confounding bit
Age does not just change how we use AI.
It changes whether AI is helpful or harmful.
For younger users, relational AI risks distorting reality. For older adults with dementia, relational AI may be one of the few tools left that can stabilise it.
Evidence is beginning to show that AI can help older adults. Ageing may also be one of the first contexts where the technology's capacity for connection is genuinely more helpful than harmful. That is a significant claim.
This series is an attempt to test it honestly.
A note on what we’ll cover:
AI is arriving in the world of dementia care and older adult support in three distinct forms, and they are not the same thing. Over three editions, we will cover each one honestly. This edition is about Companion or Conversational AI: conversational tools designed to talk with your parent, stimulate their thinking, lift their mood, and report back to you. It is the most emotionally complex category, the most commercially active, and the one where the questions are hardest. It is also where the evidence is beginning to tentatively and carefully form. The other two categories are Assistive AI that supports independence and safety, and Physical Robot Companions, which will follow in editions 2 and 3. In edition 4, we will provide a full practical guide to what exists in the UK and the US right now.
The numbers, quickly
For context, AI companion apps are now a $580 million global industry.
Fifty-two million people use them. The concern about what Companion AI does to us, particularly to young people, is now loud enough that California has legislated, the FTC has opened investigations, and Common Sense Media has called for an outright ban for under-18s.
The fear is that this technology is more intimate than social media ever was. If, as adults, we are already struggling to control our addictive relationship with that, how will those less self-assured be able to manage a Companion AI? The type of AI that is designed to pull us back into conversation to further its relationship with us.
So why are researchers and clinicians beginning to argue that AI companionship might be genuinely good for one of our most fragile cohorts, older adults living with dementia?
The flip
The population most at risk from AI emotional attachment - young people in their formative years - is the population loudest in this conversation.
The population for whom AI companion tools are now being specifically designed are older adults: people who are lonely, people who are living with dementia, people who are losing memory, losing independence, losing the ability to stay safely in their own home.
Companion AI is already a mass behaviour.
The people most at risk use it the most.
The people it may help barely use it at all
Sources: Age breakdown & revenue: Sensor Tower / Similarweb, 2024, via Consumer Startups newsletter, April 2026. 52M users / 8 of top 50 apps: Similarweb, January 2026, via Consumer Startups PolyBuzz. Daily usage: Consumer Startups newsletter, April 2026. Teen usage / serious topics preference: Common Sense Media, 2024. Replika romantic relationships: Replika company data, reported via Consumer Startups, April 2026.
Statistically, people over 65 represent less than one per cent of AI companion app users.
Yet, a wave of startups, researchers, and clinicians is quietly building something for them. These tools are not Replika1 for the elderly. They are tools that are clinician-supervised, research-backed, and purpose-built for cognitive and emotional support.
Tools that have made lonely people feel less lonely, that have been cited to reduce anxiety in dementia patients by over fifty per cent,2 and have helped a woman with Alzheimer's correctly identify the objects in her own kitchen.
Are we okay with that? And if we are, why is it different?
The divide is not about AI technology.
It’s about context.
Vulnerability is not the same thing in everybody.
A teenager forming an attachment to an AI companion risks developing distorted expectations of human relationships at exactly the moment those expectations are being formed. They are choosing AI over something available to them.
An older adult with dementia is not choosing AI over human connection.
They are living with diminishing access to it.
Through isolation, through the sheer relentlessness of the condition, through care systems that are financially inaccessible or stretched to the point of fracture. If an AI companion can provide stimulation, reduce loneliness, and give a carer an hour of rest, the balance of considerations is simply not the same. That is not a comfortable conclusion. But it may be an honest one.
Companion AI vs Conversational AI
A distinction appears to be emerging, but it is not yet clean.
Conversational AI responds. It answers questions, completes tasks, and remains, in principle, neutral. The interaction is functional.
Companion AI relates. It remembers, adapts, and builds continuity over time. The interaction is designed to feel ongoing, not transactional.
This tension sits underneath the entire category. When a system is designed to feel like a companion, users do not treat it as a tool, they treat it as something closer to a person. That design choice has consequences. Products like Replika show what happens when systems are explicitly built for relationships. Users return more often. They disclose more and in some cases, they become dependent.
For younger users, that raises concerns about attachment and distorted expectations of human relationships.
For older adults with dementia, the same mechanism may be the point.
THIS IS A CATEGORY BEING DEFINED IN REAL TIME
The question is not whether these systems can form relationships. The question is whether they should, and under what conditions. No clear boundary exists yet between helpful companionship and engineered dependence.
What Companion AI actually is
Companion AI for dementia care is not a chatbot. It is not Siri. It is purpose-built software accessed through a tablet, screen, or device designed specifically for people with mild cognitive impairment or early-to-mid stage dementia.
The more robust versions are built with clinician oversight. They are trained to prompt reminiscence, encourage gentle cognitive exercises, track mood over time, and report back to families or care teams.
Frank Poulsen is 72.
He was diagnosed with mild cognitive impairment in 2019. He now talks daily to Sunny, an AI companion developed by the US startup NewDays. When he first started, Sunny was "kind of formal," he said. Over time, their conversations have become more familiar.
"It really does feel like you're having a conversation with someone who's interested."One advantage, says his wife and carer, Cheryl: Sunny has no judgement. He can tell the same story to her three times in a day, and she listens every time.3
Jan Worrell is 85.
She lives alone on a remote stretch of the Washington coast. Her days used to pass largely in silence, sometimes without seeing another person. Then she was given ElliQ, an AI companion designed to talk, prompt and engage. At first, she resisted.
“This thing is a robot, right?”, but over time ElliQ began to learn her routines, her memories, her moods, and to speak into the quiet at just the right moments. It asked about her past, suggested activities, and listened as she told stories from across her life. Remembering details and bringing them back with care.
“Oh, Jan, this sounds just like you! Always looking for the next mountain to climb,” it told her as she shared an old story. Jan began to talk more, to play games, to reflect. After a year, her doctor noted improvements in her memory and a lower resting heart rate. 4
These are individual stories, but the data behind them is beginning to accumulate.
Much of that data comes from provider-led programmes rather than independent clinical trials. Research from Harvard's 5 has shown that talking to an empathetic AI is roughly as effective at reducing loneliness as speaking with a human stranger, at least in healthy adults.
In a New York State programme of over 800 older adults using ElliQ, 94% of users said they felt less lonely. In memory care settings, engagement can be high. In one study, residents spent an average of 47 minutes a day interacting with an AI companion over 65 days.6
In a separate memory care study using CloundMind’s Bright Path, now Kathy, an AI companion, one resident’s anxiety episodes fell by more than fifty per cent.2
The cognitive case is more cautious.
NewDays cites an evidence base of over 300 clinical trials and fifty meta-analyses supporting the underlying techniques, such as reminiscence and cognitive stimulation therapy. Trials were conducted with human practitioners, not AI systems.
NewDays' own Chief Innovation Officer has stated on record that the company cannot make the same claims as the Harvard study, but they follow the methodology and intend to run their own trials. The translation from human-led intervention to AI-led delivery has not yet been independently validated.7