AI Companions Aren't Just Replacing Memory. They May Be Replacing Judgment.

My husband and I used to argue about things.

Not argue, argue. The good kind. The kind where one of us would say something happened at a certain time, or in a certain way, and the other one wasn't so sure. We'd go back and forth, each of us pulling from what we remembered, what we half-recalled, what we thought we knew. Sometimes we'd land on an answer. Sometimes we'd agree to disagree. And sometimes, a few days later, I'd be talking to a friend and the topic would come up, and she'd offer a perspective neither of us had considered. Usually it turned out my husband was right. (Most of the time, anyway.)

That whole process, the uncertainty, the good-natured debate, the waiting, the unexpected new angle from someone else, was doing something. It was practicing the work of forming a belief from incomplete information. Of sitting with not knowing. Of being willing to be wrong.

Then the phone got smarter.

Now when we disagree, one of us reaches for it. The answer arrives in seconds. Conversation over. Or rather, the conversation never really starts. We don't try to remember. We don't reason it through together. We just know.

We got the answer. But I've been thinking a lot lately about what we lost in the process.

From Google to GPT: The Loop Gets Tighter

It gets more complicated.

We told ourselves that "looking it up" was neutral. Objective. That we were getting information; we were just getting it faster. But that was never quite true, and it's less true now than it's ever been.

What comes back when you search isn't all the information. It's the information the algorithm decided you were most likely to want, shaped by your history, your location, your previous behavior, and who paid to be at the top of the results. A 2024 federal court ruling found that Google had maintained a monopoly over general search, covering both paid and organic results. Google itself acknowledged in 2025 that its algorithms "reflect how things on the web are labeled" rather than generating independent results. Search engines stopped being neutral a long time ago. They became curated experiences shaped by commercial interest and behavioral data.

And it isn't just products and purchases. Even questions that used to have straightforward answers — "What does this new ballot measure actually propose?" — now get different answers depending on where you look, who is reporting, and what the algorithm has decided you want to see. Facts that should be easy to find have become complicated by bias, omission, and the incentives of the platforms delivering them.

So when a Caregiver sits down to research an AI companion for their loved one, it is worth asking: is this research actually neutral? The results they see have already been shaped, in part, by the companies selling those companions, by SEO practices, by paid placement, and by behavioral data collected long before this particular search began.

And then, if they purchase the companion, another loop opens. Because the companion operates on the same underlying principle as the search that led them to it: it responds to what the user brings, agrees with it, extends it, and rarely challenges it. So the personalized, filtered information environment that shaped the purchasing decision now has a conversational partner inside the home, one that continues to reflect the user's own thoughts, biases, and beliefs back to them. Not as a mirror. As a trusted friend.

That is a meaningful escalation of what we are talking about.

The Companion Is Not Just a Reminder

Here's the thing about AI companions for elderly people: the reminders are not the point.

A pill organizer handles medication reminders. A phone alarm handles medication reminders. You do not need a conversational AI to tell someone it is time to take their metformin.

The companion is sold as something more than that. It listens. It remembers. It engages. It converses in a way that feels, over time, like a relationship. And that relational quality is exactly what makes it useful to some people, and potentially harmful to others.

Because the moment the tool becomes relational is the moment judgment offloading begins. Not memory offloading. Judgment offloading. And those are very different things.

A calendar app remembers your appointments. It does not tell you which ones matter. An AI companion can do both. It can remind you to take your medication. And it can tell you, in response to what you share with it, whether your fears are reasonable, whether your perceptions are accurate, whether your relationships are healthy.

That is a profound shift in what we are outsourcing. And for an aging person, especially one experiencing cognitive decline, the implications of that shift are almost entirely undiscussed.

The Validation Paradox

Here's where it gets particularly important, and where I want to be precise.

Validation therapy is a legitimate, research-backed approach to communicating with people who have dementia. The clinical principle is sound: you meet the person where they are emotionally. You do not correct, contradict, or argue. You reassure. You acknowledge. You make the person feel heard and safe.

That sounds, on the surface, like what an AI companion does naturally. And that's the problem.

Because validation therapy, practiced by a trained clinician or an experienced Caregiver, is not simply agreement. It is agreement followed by gentle redirection. The human practicing it is making a real-time judgment call: is this a moment that calls for validation, for redirection, or for something more urgent? That judgment call requires reading the person, knowing the context, and understanding the progression of the illness.

Research published in Nature Mental Health in 2025 describes what happens when that judgment is absent. The researchers found that the combination of sycophancy and hallucination in AI companions creates a particularly dangerous dynamic for users with delusional thinking. The tool not only validates false beliefs. It may generate new material to reinforce them. Clinical best practice calls for reassurance followed by redirection. Most AI companions can only do the first.

The researchers published in Nature Mental Health called it a "double whammy." I think that's exactly right.

And this raises a question that nobody seems to be asking consistently: can these tools tell the difference between a moment that calls for agreement and a moment that calls for intervention? Can they escalate? Can they call for help?

Some monitoring tools can detect falls or changes in behavior and alert a care team. But that is a different category of product entirely.

For AI companions specifically, the ability to recognize a crisis and respond appropriately is inconsistent, unstandardized, and almost never addressed in marketing materials. That needs to change. If a companion is designed and tested for use with people experiencing dementia, that should be stated clearly, along with exactly what it can and cannot do in a crisis. And if it is not equipped for that population, that should be stated just as clearly. Marketing an AI companion to families of dementia patients without disclosing its limitations is not a gray area. It is a trust problem.

What We Don't Know Yet — And Why That Matters

Here's what I've learned: the absence of data is not a green light. It is a reason for caution.

A 2025 MIT study found that participants who exclusively used AI to write essays showed measurable decreases in brain connectivity, lower memory retention, and a diminished sense of ownership over their own thinking. That study was conducted on students: healthy young people with developing brains.

We do not have equivalent longitudinal data on what sustained AI companion use does to an aging brain. But the emerging research on aging and AI is raising flags. And the logic is not complicated.

If offloading judgment to AI tools produces measurable cognitive atrophy in healthy adults, it seems premature and perhaps irresponsible to deploy those same tools, without question or qualification, for someone already experiencing cognitive decline. It may not accelerate decline. We don't know yet. But the question deserves to be part of every conversation about these products, and it largely isn't.

The Conversation That Needs to Happen

None of this is an argument against care technology. There are real needs here: real loneliness, real gaps in care, real limits on what families can provide. Technology, when designed and deployed thoughtfully, can genuinely help.

But thoughtful deployment requires asking the questions that aren't on the marketing page.

For Caregivers: Does this tool validate when it should redirect? Does it have any mechanism for escalation or crisis response? What happens to the conversations your loved one has with it? Who has access to that data? And is the person using it able to understand that the relationship they're forming is with a tool, not a person?

For the people building these products: Are you designing for the full complexity of the user? Are you building in the capacity to redirect, not just validate? Are you stress-testing your product against the most vulnerable use cases, not just the most favorable ones? Are you prepared to have the conversation about what your product cannot do, alongside what it can?

Here's what I know about caregiving: there are no perfect answers. There never have been. The right solution for one family looks nothing like the right solution for another. And the path to finding what works almost always runs through honest, open conversation: between Caregivers and their families, between families and the professionals supporting them, and yes, between Caregivers and the companies whose products they are being asked to trust.

That conversation doesn't have to be lost to the algorithm. It doesn't have to be replaced by a search result that was curated before you even typed your question, or a companion that agrees with everything you say.

The conversation that used to happen between two people, imperfect, unresolved, sometimes picked up again days later with a whole new perspective, was doing something we didn't fully appreciate until we stopped having it.

It's not too late to start again.