AI in senior care: where it helps, where it's oversold, and what stays human

When Dr. Jeff Armstrong asked me on The Aging Well Podcast where AI genuinely helps in healthcare and where it is oversold, I had to admit where I stand. I am a software engineer, and I am not neutral about this technology.

“I'm more on the side of an AI maximalist because I no longer write my own code.”

Patrick Donahue, on The Aging Well Podcast

That might sound like a strange thing for the founder of a senior-care product to say, because I built Porchlight to use as little AI as it can get away with. Those two positions are not in conflict. Knowing what AI is good at is exactly what tells you where to keep it out. This article is for caregivers, families and care teams trying to sort the real from the hype.

Where AI is genuinely useful in care right now

Turning speech into text. Speech recognition has become cheap and accurate in the last few years. My own primary care doctor now records the visit, gets a transcript, and has a model draft the note. That means the doctor looks at me instead of a keyboard. The same leap is why voice-first tools for older adults are suddenly practical: a resident can talk, and the words get written down reliably.

Tidying and summarizing. Raw transcripts are hard to read. People repeat themselves, trail off, and circle back.

“These raw transcripts that you get aren't really very nice to read until they've been smoothed over by a large language model.”

Patrick Donahue, on The Aging Well Podcast

A model is also good at pulling a few useful lines out of a long conversation: loves baseball, worked as a machinist, would rather not talk about the war. That kind of summary is what a busy aide can actually read before walking into a room.

Helping people find human help. I shared on the show that during a hard stretch after my baby was born, a chatbot pointed me to a free state warm line, and I talked with a real person there for a long time. The best version of AI in a moment like that is a doorway to people.

Where AI is oversold

Anything with consequences. Medical decisions, medication changes, legal questions. Large language models write fluent, confident answers whether or not they are right.

“The problem with AI is that it's... kind of a liar. It's very good at telling you something that sounds plausible but... doesn't hold up.”

Patrick Donahue, on The Aging Well Podcast

“A human being should... be the ultimate authority. It's called human in the loop.”

Patrick Donahue, on The Aging Well Podcast

Replacing relationships. Consumer voice assistants have become astonishingly good to talk to. I practice my German with one around the house. In a sense it is a competitor to what I build. But it is a closed loop between one person and a company.

“What it won't do is it won't share the information easily with your care team, and it won't do that whole relay thing.”

Patrick Donahue, on The Aging Well Podcast

For an older adult in a care community, the point is not to have a nicer chatbot to talk to. The point is for the aide, the activity director, and the daughter in another state to know the person better.

Privacy, by default. Consumer AI apps are built for individuals, and depending on the app and its settings, your conversations may be stored and used to improve future models. That is a reasonable trade for some people asking about recipes. It is a very different question for an older adult describing their health, their marriage, or their grief. Before anyone in your family pours their life into a general chatbot, read the data settings.

AI should be the plumbing that carries a story to the people who care, not the one doing the caring.

What should never be delegated to software

Jeff asked me directly what should never be handed to technology when caring for another person. My short answer is judgment and presence.

• Medical and care decisions. A person reviews and decides. Always.
• Noticing that someone is not okay. I said on the show that in severe mental health situations, a system "needs to immediately escalate." In care settings, that responsibility belongs to trained people who know the resident. Software is not a safety net, and nobody should treat it as one.
• The relationship itself. Showing up, holding a hand, laughing at the same story for the tenth time.

“My grandiose goal for my product would be AI will facilitate human relationship, facilitate human health, and facilitate human interaction. It's not meant to replace it in any respect.”

Patrick Donahue, on The Aging Well Podcast

How Porchlight uses AI, and where it deliberately does not

Some older adults find AI intrusive or creepy, and so do some of their families. I take that seriously.

“Some people find AI invasive, intrusive, or creepy. So... at this stage of development, we are deliberately low tech.”

Patrick Donahue, on The Aging Well Podcast

Here is the honest list.

• The questions are not AI. A resident hears questions from a curated life-story sequence, read aloud in a fixed order that starts gentle. An early version experimented with AI-generated follow-up questions. We took that out. Nobody is being interviewed by a chatbot.
• Speech-to-text writes down what they said. A transcription model turns the recording into text, and a screen throws out the made-up sentences speech recognition sometimes produces over silence. The original audio is kept.
• A language model tidies the transcript so it reads cleanly, the way a careful editor would fix stumbles without changing what the person meant.
• A language model drafts the staff briefing from the resident's own stories (and any stories family or staff add about them), so a new caregiver can read the essentials in a few minutes. A weekly digest of story highlights goes to staff every Monday.
• People do the rest. Stories go to the resident's care team and to the family members the resident and community approve. Our who sees what page spells it out.

And here is what Porchlight does not do. It does not detect distress, diagnose anything, or escalate on its own. It is a conversation tool. Staff and family, the humans who know the resident, are the ones watching how a session is going, and they should stop it whenever it is not going well.

Why I won't make clinical claims

Jeff noticed that I am careful not to say Porchlight improves health. That is on purpose.

“You know, it's a legal liability. Like if... my business advances to the point that I am able to launch some kind of a study, you know, I'd love to report on the results of that study.”

Patrick Donahue, on The Aging Well Podcast

Reminiscence and life review have real research behind them, with modest benefits that depend on the setting.¹ That does not mean any particular app has been shown to deliver them. If a product in this space tells you it is clinically proven to improve memory or mood, ask for the study. I would rather tell you what Porchlight actually does and let you judge.

Questions to ask any AI caregiving tool

1. Where exactly is AI used, and where is a person making the call?
2. Who can see what my family member says, and can I see that list?
3. Is the original recording kept, or only an AI summary?
4. What happens if a resident says something worrying? Who notices?
5. What does the company claim about health outcomes, and what is the evidence?

Good tools will have plain answers. The best ones will make the people in the room more present, not less.